Assume your robot is a point mass

As always in our battle against complexity, I want to find a setup that is as simple as possible (but no simpler)! Here is my proposal for the box-flipping example. First, I will restrict all motion to a 2D plane; this allows me to chop off two sides of the bin (for you to easily see inside), but also drops the number of degrees of freedom we have to consider. In particular, we can avoid the quaternion floating base coordinates, by adding a PlanarJoint to the box. Instead of using a complete gripper, let's start even simpler and just use a "point finger". I've visualized it as a small sphere, and modeled two control inputs as providing force directly on the $x$ and $z$ coordinates of the point mass.

You will get efficient and thoughtful service from Fuxin Intelligent.

Even in this simple model, and throughout the discussions in this chapter, we will have two dynamic models that we care about. The first is the full dynamic model used for simulation: it contains the finger, the box, and the bin, and has a total of 5 degrees of freedom ($x, z, \theta$ for the box and $x, z$ for the finger). The second is the robot model used for designing the controllers: this model has just the two degrees of freedom of the finger, and experiences unmodelled contact forces. By design, the equations of this second model are particularly simple: $$\begin{bmatrix}m & 0 \\ 0 & m \end{bmatrix} \dot{v} = m \begin{bmatrix}\ddot{x} \\ \ddot{z} \end{bmatrix} = \tau_g + \begin{bmatrix} u_x \\ u_z \end{bmatrix} + f^{F_c},$$ where $m$ is the mass, $\tau_g$ is the gravitational "torque" which here is just $\tau_g = [0, -mg]^T$, $u$ is the control input vector, and $f^{F_c}$ is the Cartesian contact force applied to the finger, $F$, at $c$.

This chapter will make heavy use of our spatial force notation. To make sure it's clear, I've used $f^{F_c}$ above to denote force applied to the finger, $F$, at point $c$. If we want to denote the same force applied to the box, $B$, at point $c$, then we'll denote it as $f^{B_c}$. Naturally, we have $f^{F_c} = - f^{B_c}$, because the forces must be equal and opposite. If we're not careful about this notation, the signs will definitely get confusing below.

Trajectory tracking

Before we even make contact with the box, let's make sure we know how to move our robot finger around through the air. Assume that you've done some motion planning, perhaps using optimization, and have developed a beautiful desired trajectory, $q_d(t)$, that you want your finger to follow. Let's assume that $q_d(t)$ is twice differentiable. What generalized-force commands should you send to the robot to execute that trajectory?

One of the first and most common ways to execute the trajectory is with proportional-integral-derivative (PID) control, which we've discussed when we talked about position-controlled robots. The PidController uses the command $$u(t) = K_p (q_d(t) - q(t)) + K_d (\dot{q}_d(t) - \dot{q}(t)) + K_i \int_0^t (q_d(t) - q(t)) dt,$$ with $K_p$, $K_d$, and $K_i$ being the position, velocity, and integral gains. Often these gains are diagonal matrices, making the command for each joint independent of the rest. Note that in manipulation, we tend to avoid the integral term, setting $K_i = 0$. You can imagine that if the robot is pushing up against the wall and is unable to achieve $q(t) == q_d(t)$, then the integral term will "wind-up", sending larger and larger commands to the robot until some fault is reached.

PD control can be incredibly effective; it assumes almost nothing about the dynamics of the robot. But this assumption is also a limitation; if we do know something about the dynamics of the robot then we can achieve better tracking performance. To make that clear, let's write the closed-loop dynamics of the $z$ position of our robot finger under PD control:$$m \ddot{z} = -mg + k_p (z_d - z) + k_d (\dot{z}_d - \dot{z}).$$ Here is a sample roll-out when we command a sinusoidal trajectory, $z_d(t)$, and start the finger slightly off the trajectory in $z$ and $\dot{z}$: The actual trajectory does track the desired trajectory, but it will have some persistent errors. Even if the desired trajectory was a constant, $\dot{z}_d = 0$, we would expect to have a steady-state error, $\tilde{z} = z_d - z$, obtained informally by setting $\dot{z} = \ddot{z} = 0$: $$\tilde{z} = z_d - z = \frac{1}{k_p}mg.$$ In this configuration, the restoring force from the PD controller is exactly balancing the perturbation force from gravity, and is not doing anything to drive the error further towards zero.

Looking at the equations, it seems clear that we can do better if we inform the controller about the gravity term. This is commonly referred to as "gravity compensation". For reasons that will become clear soon, the gravity compensation + PD controller is available in Drake as the JointStiffnessController, $$u = -\tau_g + K_p (q_d - q) + K_d (\dot{q}_d - \dot{q}).$$ For the $z$ axis of our point finger, this results in the closed-loop dynamics: $$m \ddot{z} = k_p (z_d - z) + k_d (\dot{z}_d - \dot{z}).$$ This controller has no steady-state error, and tracks better in practice:

Although the steady-state error is gone, we can still see errors when tracking a more complicated desired trajectory. There is one more relevant piece of information, however, which we have access to, but which we have not yet given to our controller: $\ddot{q}_d(t)$. Let's write our controller access in a subtly different form: $$u = -\tau_g + m\left[\ddot{q}_d + K_p (q_d - q) + K_d (\dot{q}_d - \dot{q}) \right].$$ The general form of this controller is available in Drake as the InverseDynamicsController. Two things changed here; in addition to adding the feed-forward acceleration term into the controller, we also multiplied the PD terms by the mass. Combined, this has the result that the closed-loop error dynamics simplifies to $$\ddot{\tilde{z}} + k_d \dot{\tilde{z}} + k_p \tilde{z} = 0,$$ which may be familiar to you as a simple mass-spring-damper model. Given reasonable choices for $k_p$ and $k_d$, the error dynamics converge nicely to zero, giving us our best tracking performance yet:

Trajectory tracking for the point mass

I've put the code to generate these simple plots into a notebook so that you can play around with them yourself:

(Direct) force control

Things get more interesting when we start making contact (in our example, it means our finger starts contacting the box). In this case, controlling the position / velocity / acceleration of the finger might not be the only goal. We might also want to control the interaction forces that we are applying to the box.

In the simplest case, let's not try to regulate the finger position at all, but instead implement a low-level controller that accepts the desired contact forces and produces generalized force inputs to the robot that try to make the actual contact forces match the desired.

What information do we need to regulate the contact forces? Certainly we need the desired contact force, $f^{F_c}_{desired}$. In general, we will need the robot's state (though in the immediate example, the dynamics of our point mass are not state dependent). But we also need to either (1) measure the robot accelerations (which we've try to avoid, since they are often noisy measurements), (2) assume the robot accelerations are zero, or (3) provide a measurement of the contact force so that we can regulate it with feedback.

Let's consider the case where the robot is already in contact with the box. Let's also assume for a moment that the accelerations are (nearly) zero. This is actually not a horrible assumption for most manipulation tasks, where the movements are relatively slow. In this case, our equations of motion reduce to $$f^{F_c} = - mg - u.$$ Our force controller implementation can be as simple as $u = -mg - f^{F_c}_{desired}.$ Note that we only need to know the mass of the robot (not the box) to implement this controller.

What happens when the robot is not in contact? In this case, we cannot reasonably ignore the accelerations, and applying the same control results in $m\dot{v} = - f^{F_c}_{desired}.$ That's not all bad. In fact, it's one of the defining features of force control that makes it very appealing. When you specify a desired force and don't get it, the result is accelerating the contact point in the (opposite) direction of the desired force. In practice, this (locally) tends to bring you into contact when you are not in contact.

Commanding a constant force

Let's see what happens when we run a full simulation which includes not only the non-contact case and the contact case, but also the transition between the two (which involves collision dynamics). I'll start the point finger next to the box, and apply a constant force command requesting to get a horizontal contact force from the box. I've drawn the $x$ trajectory of the finger for different (constant) contact force commands.

For all strictly positive desired force commands, the finger will accelerate at a constant rate until it collides with the box (at $x=0.089$). For small $f^{F_c}_{desired}$, the box barely moves. For an intermediate range of $f^{F_c}_{desired}$, the collision with the box is enough to start it moving, but friction eventually brings the box and therefore also the finger to rest. For large values, the finger will keep pushing the box until it runs into the far wall.

Consider the consequences of this behavior. By commanding force, we can write a controller that will come into a nice contact with the box with essentially no information about the geometry of the box (we just need enough perception to start our finger in a location for which a straight-line approach will reach the box).

This is one of the reasons that researchers working on legged robots also like force control. On a force-capable walking robot, we might mimic position control during the "swing phase", to get our foot approximately where we are hoping to step. But then we switch to a force control mode to actually set the foot down on the ground. This can significantly reduce the requirements for accurate sensing of the terrain.

A force-based flip-up strategy

Update the code to match this updated derivation. In particular, \theta_d(t) should get more negative.

Here is my proposal for a simple strategy for flipping up the box. Once in contact, we will use the contact force from the finger to apply a moment around the bottom left corner of the box to generate the rotation. But we'll add constraints to the forces that we apply so that the bottom corner does not slip.

These conditions are very natural to express in terms of forces. And once again, we can implement this strategy with very little information about the box. The position of the finger will evolve naturally as we apply the contact forces. It's harder for me to imagine how to write an equally robust strategy using a (high-gain) position-controller finger; it would likely require many more assumptions about the geometry (and possibly the mass) of the box.

Let's encode the textual description above, describing the forces that are applied to the box. I'll use $C$ for the contact frame between the finger and the box, with its normal pointing into the box normal to the surface, and $A$ for the contact frame for the lower left corner of the box contacting the bin, with the normal pointing straight up in positive world $z$. $${}^BR^C = R_y(-\frac{\pi}{2}), \qquad {}^WR^A = I,$$ where $R_y(\theta)$ is a rotation by $\theta$ around the y axis. Of course we also have the force of gravity, which is applied at the body center of mass (com): $$f^{B_{com}}_{gravity,W} = mg.$$ As you can see, we'll make heavy use of the spatial force notation / spatial algebra described here.

The friction cone provides (linear inequality) constraints on the forces we want to apply. \begin{gather*} f^{B_C}_{\text{finger}, C_z} \ge 0, \qquad |f^{B_C}_{\text{finger}, C_x}| \le \hat\mu_C f^{B_C}_{\text{finger}, C_z}, \\ f^{B_A}_{\text{ground}, A_z} \ge 0, \qquad |f^{B_A}_{\text{ground}, A_x}| \le \hat\mu_A f^{B_A}_{\text{ground}, A_z}.\end{gather*} Please make sure you understand the notation. Within those constraints, we would like to rotate up the box.

To rotate the box about $A$, let's reason about the total torque being applied about $A$: $$\tau^{B_A}_{total,W} = \tau^{B_A}_{gravity, W} + \tau^{B_A}_{ground, W} + \tau^{B_A}_{finger, W},$$ but we know that $\tau^{B_A}_{ground, W} = 0$ since the moment arm is zero. I have a goal here of not making too many assumptions about the mass and geometry of the box in our controller, so rather than try to regulate this torque perfectly, let's write $$\tau^{B_A}_{total,W_y} = \tau^{B_A}_{finger,W_y} + \text{unknown terms}.$$ A reasonable control strategy in the face of these unmodeled terms is to use feedback control on the angle of the box (call it $\theta$) which is the $y$ component of ${}^WR^B$: $$ \tau^{B_A}_{finger_d,W_y} = \text{PID}(\theta_d, \theta), \qquad {}^WR^B(\theta) = R_y(\theta),$$ where I've used $\theta_{d}$ as shorthand for the desired box angle and $\text{PID}$ as shorthand for a simple proportional-integral-derivative term. Note that $\tau^{B_A}_{finger,W_y} \propto f^{B_C}_{finger, C_x},$ where the (constant) coefficient only depends on ${}^B\hat{p}^C;$ it does not actually depend on $\hat\theta.$

To execute the desired PID control subject to the friction-cone constraints, we can use constrained least-squares: \begin{align*}\min_{f^{B_C}_{finger,C}, \, f^{B_A}_{ground,A}} \quad& \left| \tau^{B_A}_{finger,W_y} - \text{PID}(\theta_d, \hat\theta) \right|^2,\\ \subjto \quad & f^{B_C}_{\text{finger}, C_z} \ge 0, \qquad |f^{B_C}_{\text{finger}, C_x}| \le \hat\mu_C f^{B_C}_{\text{finger}, C_z}, \\ & f^{B_A}_{\text{ground}, A_z} \ge 0, \qquad |f^{B_A}_{\text{ground}, A_x}| \le \hat\mu_A f^{B_A}_{\text{ground}, A_z}, \\ & f^{B_A}_{\text{ground}, A} + \hat{f}^{B_A}_{gravity, A} + f^{B_A}_{finger, A} = 0.\end{align*} Note that the last line is still a linear constraint once $\hat\theta$ is given, despite requiring some spatial algebra operations. Implementing this strategy assumes:

• We have some approximation, $\hat\theta$, for the orientation of the box. We could obtain this from a point cloud normal estimation, or even from tracking the path of the fingers.
• We have conservative estimates of the coefficients of static friction between the wall and the box, $\hat\mu_A$, and between the finger and the box, $\hat\mu_C$, as well as the approximate location of the finger relative to the box corner, and the box mass, $\hat{m}.$

You should experiment and decide for yourself whether these estimates can be loose. I think that you'll find it's quite robust to knowing the mass and geometry of the box. Generate random boxes.

We have multiple controllers in this example. The first is the low-level force controller that takes a desired contact force and sends commands to the robot to attempt to regulate this force. The second is the higher-level controller that is looking at the orientation of the box and deciding which forces to request from the low-level controller.

Please also understand that this is not some unique or optimal strategy for box flipping. I'm simply trying to demonstrate that sometimes controllers which might be difficult to express otherwise can easily be expressed in terms of forces!

Indirect force control

There is a nice philosophical alternative to controlling the contact interactions by specifying the forces directly. Instead, we can program our robot to act like a (simple) mechanical system that reacts to contact forces in a desired way. This philosophy was described nicely in an important series of papers by Ken Salisbury introducing stiffness control Salisbury80 and then Neville Hogan introducing impedance control Hogan85a+Hogan85b+Hogan85c.

This approach is conceptually very nice. Imagine we were to walk up and push on the end-effector of the iiwa. With only knowledge of the parameters of the robot itself (not the environment), we can write a controller so that when we push on the end-effector, the end-effector pushes back (using the entire robot) as if you were pushing on, for instance, a simple spring-mass-damper system. Rather than attempting to achieve manipulation by moving the end-effector rigidly through a series of position commands, we can move the set points (and perhaps stiffness) of a soft virtual spring, and allow this virtual spring to generate our desired contact forces.

This approach can also have nice advantages for guaranteeing that your robot won't go unstable even in the face of unmodeled contact interactions. If the robot acts like a dissipative system and the environment is a dissipative system, then the entire system will be stable. Arguments of this form can ensure stability for even very complex system, building on the rich literature on passivity theory or more generally Port-Hamiltonian systemsDuindam09.

Our simple model with a point finger is ideal for understanding the essence of indirect force control. The original equations of motion of our system are $$m\begin{bmatrix}\ddot{x} \\ \ddot{z} \end{bmatrix} = mg + u + f^{F_c}.$$ We can write a simple controller to make this system act, instead, like (passive) mass-spring-damper system: $$m \begin{bmatrix}\ddot{x} \\ \ddot{z} \end{bmatrix} + b \begin{bmatrix} \dot{x} \\ \dot{z} \end{bmatrix} + k \begin{bmatrix} x - x_d \\ z - z_d \end{bmatrix} = f^{F_c},$$ with the rest position of the spring at $(x_d, z_d).$ The controller that implements this follows easily; in the point finger case this has the familiar form of a proportional-derivative (PD) controller, but with an additional "feed-forward" term to cancel out gravity.

Technically, if we are just programming the stiffness and damping, as I've written here, then a controller of this form would commonly be referred to as "stiffness control", which is a subset of impedance control. We could also change the effective mass of the system; this would be impedance control in its full glory. My impression, though, is that the consensus in robot control experts is that changing the effective mass is most often considered not worth the complexity that comes from the extra sensing and bandwidth requirements.

The literature on indirect force control has a lot of terminology and implementation details that are important to get right in practice. Your exact implementation will depend on, for instance, whether you have access to a force sensor and whether you can command forces/torque directly. The performance can vary significantly based on the bandwidth of your controller and the quality of your sensors. See e.g. Villani08 for a more thorough survey (and also some fun older videos), or Whitney87 for a nice earlier perspective.

Teleop with stiffness control

I didn't give you a teleop interface with direct force control; it would have been very difficult to use! Moving the robot by positioning the set points on virtual springs, however, is quite natural. Take a minute to try moving the box around, or even flipping it up.

To help your intuition, I've made the bin and the box slightly transparent, and added a visualization (in orange) of the virtual finger or gripper that you are moving with the sliders.

Let's embrace indirect force control to come up with another approach to flipping up the box. Flipping the box up in the middle of the bin required very explicit reasoning about forces in order to stay inside the friction cone in the bottom left corner of the box. But there is another strategy that doesn't require as precise control of the forces. Let's push the box into the corner, and then flip it up.

To make this one happen, I'd like to imagine creating a virtual spring -- you can think of it like a taut rubber band -- that we attach from the finger to a point near the wall just a little above the top left corner of the box. The box will act like a pendulum, rotating around the top corner, with the rubber band creating the moment. At some point the top corner will slip, but the very same rubber band will have the finger pushing the box down from the top corner to complete the maneuver.

Consider the alternative of writing an estimator and controller that needed to detect the moment of slip and make a new plan. That is not a road to happiness. By only using our model of the robot to make the robot act like a different dynamical system at the point we can accomplish all of that!

A stiffness-control-based flip-up strategy

This controller almost couldn't be simpler. I will just command a trajectory the move the virtual finger to just in front of the wall. This will push the box into contact and establish our bracing contact force. Then I'll move the virtual finger (the other end of our rubber band) up the wall a bit, and we can let mechanics take care of the rest!

So far we've made our finger act like two independent mass-spring-damper systems, one in $x$ and the other in $z$. We even used the same stiffness, $k$, and damping, $b$, parameters for each. More generally, we can write stiffness and damping matrices, which we typically would call $K_p$ and $K_d$, respectively: $$m \begin{bmatrix}\ddot{x} \\ \ddot{z} \end{bmatrix} + K_d \begin{bmatrix} \dot{x} \\ \dot{z} \end{bmatrix} + K_p \begin{bmatrix} x - x_d \\ z - z_d \end{bmatrix} = f^{F_c}.$$ In order to emphasize the coordinate frame of these stiffness matrices, let's write exactly the same equation, realizing that with our point finger we have ${}^Wp^F = \begin{bmatrix} x, z\end{bmatrix}^T,$ so: $$m\,{}^W\dot{v}^F + K_d \, {}^Wv^F + K_p\left({}^Wp^F - {}^Wp^{F_d}\right) = f^F.$$ Sometimes it can be convenient to express the stiffness (and/or damping) matrices in a different frame, $A$: \begin{align*} m\,{}^W\dot{v}^F + {}^W R^A K_d \, {}^Wv^F_A + {}^WR^W K_p\left({}^Wp^F_A - {}^Wp^{F_d}_A\right) =& \\ m\,{}^W\dot{v}^F + {}^W R^A K_d {}^A R^W\, {}^Wv^F + {}^WR^A K_p {}^A R^W\left(\,{}^Wp^F - {}^Wp^{F_d}\right) =& \, f^F.\end{align*}

Hybrid position/force control

There are a number of applications where we would like to explicitly command force in one direction, but command positions in another. One classic examples if you are trying to wipe or polish a surface -- you might care about the amount of force you are applying normal to the surface, but use position feedback to follow a trajectory in the directions tangent to the surface. In the simplest case, imagine controlling force in $z$ and position in $x$ for our point finger: $$u = -mg + \begin{bmatrix} k_p (x_d - x) + k_d (\dot{x}_d - \dot{x}) \\ -f^F_{desired, W_z} \end{bmatrix}.$$ If want the forces/positions in a different frame, for instance the contact frame, $C$, then we can use \begin{equation} u = -mg + {}^WR^C \begin{bmatrix} k_p (p^{F_d}_{C_x} - p^{F}_{C_x}) + k_d (v^{F_d}_{C_x} - v^F_{C_x}) \\ -f^{F}_{desired, C_z}\end{bmatrix}. \label{eq:position-or-force}\end{equation} By commanding the normal force, you not only have the benefit of controlling how hard the robot is pushing on the surface, but also gain some robustness to errors in estimating the surface normal. If a position-controlled robot estimated the normal of a wall badly, then it might leave the wall entirely in one direction, and push extremely hard in the other. Having a commanded force in the normal direction would allow the position of the robot in that direction to become whatever is necessary to apply the force, and it will follow the wall.

Check yourself: can the finger move the box? (only if the coefficient of friction is bigger).

The choice of position control or force control need not be a binary decision. We can simply apply both the position command (as in stiffness/impedance control) and a "feed-forward" force command: $$u = -mg + K_p (p^{F_d} - p^{F}) + K_d (v^{F_d} - v^{F}) - f^{F}_{\text{feedforward}}.$$ Certainly this is only more general that the explicit position-or-force mode in Eq. (\ref{eq:position-or-force}); we could achieve the explicit position-or-force in this interface by the appropriate choices of $K_p$, $K_d$, and $f^F_{feedforward}.$ As we'll see, this is quite similar to the interface provided by the iiwa (and many other torque-controlled robots).

The general case (using the manipulator equations)

Using the floating finger/gripper is a good way to understand the main concepts of force control without the details. But now it's time to actually implement those strategies using joint commands that we can send to the arm.

Our starting point is understanding that the equations of motion for a fully-actuated robotic manipulator have a very structured form: \begin{equation}M(q)\ddot{q} + C(q,\dot{q})\dot{q} = \tau_g(q) + u + \sum_i J^T_i(q)f^{c_i}.\label{eq:manipulator} \end{equation} The left side of the equation is just a generalization of "mass times acceleration", with the mass matrix, $M$, and the Coriolis terms $C$. The right hand side is the sum of the (generalized) forces, with $\tau_g(q)$ capturing the forces due to gravity, $u$ the joint torques supplied by the motors, and $f^{c_i}$ is the Cartesian force due to the $i$th contact, where $J_i(q)$ is the $i$th "contact Jacobian". I introduced a version of these briefly when we described multibody dynamics for dropping random objects into the bin, and have more notes available here. For the purposes of the remainder of this chapter, we can assume that the robot is bolted to the table, so does not have any floating joints; I've therefore used $\dot{q}$ and $\ddot{q}$ instead of $v$ and $\dot{v}$.

Trajectory tracking

Let us ignore the contact terms for a moment: $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} = \tau_g(q) + u.$$ The forward dynamics problem, which we use during simulation, is to compute the accelerations, $\ddot{q}$, given the joint torques, $u$, (and $q, \dot{q}$). The inverse dynamics problem, which we can use in control, is to compute $u$ given $\ddot{q}$ (and $q, \dot{q}$). As a system, it looks like

This system implements the controller: $$u = M(q)\ddot{q}_d + C(q,\dot{q})\dot{q} - \tau_g(q),$$ so that the resulting dynamics are $M(q)\ddot{q} = M(q)\ddot{q}_d.$ Since we know that $M(q)$ is invertible, this implies that $\ddot{q} = \ddot{q}_d$.

When we put the contact forces back in, In practice, we only have estimates of the dynamics (mass matrix, etc) and even the states $q$ and $\dot{q}$. We will need to account for these errors when we determine the acceleration commands to send. For example, if we wanted to follow a desired joint trajectory, $q_0(t)$, we do not just differentiate the trajectory twice and command $\ddot{q}_0(t)$, but we could instead command e.g. $$\ddot{q}_d = \ddot{q}_0(t) + K_p(q_0(t) - q) + K_d(\dot{q}_0(t) - \dot{q}).$$ One could also include an integral gain term.

The resulting System has one additional input port for the desired state:

This "inverse dynamics control" also appears in the literature with a few other names, such as "computed-torque control" and even the "feedforward-plus-feedback-linearizing control" Lynch17.

Check yourself: What is the difference between the joint stiffness control and the inverse dynamics control?

The closed-loop dynamics with the joint stiffness control (taking $\tau_{ff} = 0$) are given by $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + K_p(q - q_d) + K_d(\dot{q} - \dot{q}_d) = \tau_{ext},$$ where $\tau_{ext}$ summarized the results of (unmodeled) forces like contacts. The closed-loop dynamics with the inverse dynamics controllers are given by $$\ddot{q} + K_p(q - q_d) + K_d(\dot{q} - \dot{q}_d) = M^{-1}(q) \tau_{ext}.$$ So the inverse dynamics controller writes the feedback law in the units of accelerations rather than forces; this would be similar to shaping the internal matrix into the diagonal matrix.

The ability to realize this cancellation in hardware, however, will be limited by the accuracy with which we estimate the model parameters, $\hat{M}(q)$ and $\hat{C}(q,\dot{q})$. The stiffness controller, on the other hand, achieves much of the desired performance but only attempts to cancel the gravity (and perhaps friction) terms; it does not shape the inertia.

Joint stiffness control

In practice, the way that we most often interface with the iiwa is through the its "joint-impedance control" mode, which is written up nicely in Ott08+Albu-Schaffer07. For our purposes, we can view this as a stiffness control in joint space: $$u = -\tau_g(q) + K_p(q_d - q) + K_d(\dot{q}_d - \dot{q}) + \tau_{ff},$$ where $K_p, K_d$ are positive diagonal matrices, and $\tau_{ff}$ is a "feed-forward" torque. In practice the controller also includes some joint friction compensationAlbu-Schaffer01, but I've left those friction terms out here for the sake of brevity. The controller does also do some shaping of the inertias (earning it the label "impedance control" instead of only "stiffness control"), but only the rotor inertias, and the user does not set these effective inertias. From the user perspective it should be viewed as stiffness control.

Check yourself: What is the difference between traditional position control with a PD controller and joint-stiffness control?

The difference in the algebra is quite small. A PD control would typically have the form $$u=K_p(q_d-q) + K_d(\dot{q}_d - \dot{q}),$$ whereas stiffness control is $$u = -\tau_g(q) + K_p(q_d-q) + K_d(\dot{q}_d - \dot{q}).$$ In other words, stiffness control tries to cancel out the gravity and any other estimated terms, like friction, in the model. As written, this obviously requires an estimated model (which we have for iiwa, but don't believe we have for most arms with large gear-ratio transmissions) and torque control. But this small difference in the algebra can make a profound difference in practice. The controller is no longer depending on error to achieve the joint position in steady state. As such we can turn the gains way down, and in practice have a much more compliant response while still achieving good tracking when there are no external torques.

Gravity comp is a classic demo of how well you model/control your robot. Show kuka videos. C++ implementation. Add code example(s) here.

Cartesian stiffness and operational space control

A better analogy for the control we were doing with the point finger example is to write a controller so that the robot acts like a simple dynamical system in the world frame. To do that, we have to identify a frame, $E$ for "end-effector", on the robot where we want to impose these simple dynamics -- ideally the origin of this frame will be the expected point of contact with the environment. Following our treatment of kinematics and differential kinematics, we'll define the forward kinematics of this frame as: \begin{equation}p^E = f_{kin}(q), \qquad v^E = \dot{p}^E = J(q) \dot{q}, \qquad a^E = \ddot{p}^E = J(q)\ddot{q} + \dot{J}(q) \dot{q}.\label{eq:kinematics}\end{equation} We haven't actually written the second derivative before, but it follows naturally from the chain rule. Also, I've restricted this to the Cartesian positions for simplicity; one can think about the orientation of the end-effector, but this requires some care in defining the 3D stiffness in orientation (e.g. Suh22).

One of the big ideas from manipulator control is that we can actually write the dynamics of the robot in the frame $E$, by parameterizing the joint torques as a translational force applied to body $B$ at $E$, $f^{B_E}_u$: $u = J^T(q) f^{B_E}_u$. This comes from the principle of virtual work. By substituting this and the manipulator equations (\ref{eq:manipulator}) into (\ref{eq:kinematics}) and assuming that the only external contacts happen at $E$, we can write the dynamics: \begin{equation} M_E(q) \ddot{p}^E + C_E(q,\dot{q})\dot{q} = f^{B_E}_g(q) + f^{B_E}_u + f^{B_E}_{ext} \label{eq:cartesian_dynamics},\end{equation} where $$M_E(q) = (J M^{-1} J^T)^{-1}, \qquad C_E(q,\dot{q}) = M_E \left(J M^{-1} C - \dot{J} \right), \qquad f^{B_E}_g(q) = M_E J M^{-1} \tau_g.$$ Now if we simply apply a controller analogous to what we used in joint space, e.g.: $$f^{B_E}_u = -f^{B_E}_g(q) + K_p(p^{E_d} - p^E) + K_d(\dot{p}^{E_d} - \dot{p}^E),$$ then we can achieve the desired closed-loop dynamics at the end-effector: $$M_E(q) \ddot{p}^E + C_E(q,\dot{q})\dot{q} + K_p(p^{E} - p^{E_d}) + K_d(\dot{p}^{E} - \dot{p}^{E_d}) = f^{B_E}_{ext}.$$ So if I push on the robot at the end-effector, the entire robot will move in a way that will make it feel like I'm pushing on a spring-damper system. Beautiful!

The company is the world’s best electronic manipulators supplier. We are your one-stop shop for all needs. Our staff are highly-specialized and will help you find the product you need.

When $J(q)$ can be rank-deficient (for instance, if you have more degrees of freedom than you are trying to control at the end-effector), you'll also want to add some terms to similar to stabilize the null space of your Jacobian. Since we are operating here with torques instead of velocity commands, the natural analogue is to write the a joint-centering PD controller that operates in the null space of the end-effector control: $$u = J^Tf^{B_E}_u + P[K_{p,joint} (q_0 - q) - K_{d,joint}\dot{q}].$$ Here $P$ is the so-called "dynamically-consistent null-space projection" Khatib87. In differential IK operate in the null space of the velocity controller, but here we have to operate in the null space for the accelerations.

Add a derivation, or at least the equations for, the dynamically consistent null-space projection.

This approach of programming the task-space dynamics with secondary joint-space objectives operating in the null space was developed in Khatib87 and is known as operational space control. It can be generalized to a rich composition of prioritized task-space and joint-space control objectives; the prioritization is accomplished by putting lower-priority tasks in the null space of the primary task. Like the joint-centering differential inverse kinematics, these can also be understood through the lens of least-squares optimization.

Some implementation details on the iiwa

The implementation of the low-level impedance controllers has many details that are explained nicely in Ott08+Albu-Schaffer07. In particular, the authors go to quite some length to implement the impedance law in the actuator coordinates rather than the joint coordinates (remember that they have an elastic transmission in between the two). I suspect there are many subtle reasons to prefer this, but they go to lengths to demonstrate that they can make a true passivity argument with this controller.

The iiwa interface offers a Cartesian impedance control mode. If we want high performance stiffness control in end-effector coordinates, then we should definitely use it! The iiwa controller runs at a much higher bandwidth (faster update rate) than the interface we have over the provided network API, and many implementation details that they have gone to great lengths to get right. But in practice we do not use it, because we cannot convert between joint impedance control and Cartesian impedance control without shutting down the robot. Sigh. In fact we cannot even change the stiffness gains nor the frame $C$ (aka the "end-effector location") without stopping the robot. So we stay in joint impedance mode and command some Cartesian forces through $\tau_{ff}$ if we desire them. (If you are interested in the driver details, then I would recommend the documentation for the Franka Control Interface which is much easier to find and read, and is very similar to functionality provided by the iiwa driver.)

You might also notice that the interface we provide to the IiwaDriver in the HardwareStation takes a desired joint position for the robot, but not a desired joint velocity. That is because we cannot actually send a desired joint velocity to the iiwa controller. In practice, we believe that they are numerically differentiating our position commands to obtain the desired velocity, which adds a delay of a few control timesteps (and sometimes non-monotonic behavior). I believe that the reason why we aren't allowed to send our own joint velocity commands is to to make the passivity argument in their paper go through, and perhaps to make a simpler/safer API -- they don't want users to be able to send a series of inconsistent position and velocity commands (where the velocities are not actually the time derivatives of the positions).

Summarize the exchange with the kuka folks describing these details.

Since the controller attempts to cancel gravitational terms, one can also tell the iiwa firmware about the lumped inertia of the gripper. This is set just once (in the pendant), and cannot be changed while the robot is moving; if you pick something up (or even move the fingers), those changes in the gravitational terms will not be compensated in the controller.

Although the JointStiffnessController in Drake is the best model for the iiwa control stack, we commonly use the InverseDynamicsController in our simulations instead. This is for a fairly subtle reason -- the numerics are better. The effective stiffness of the differential equations to achieve comparable stiffness in the physics is smaller, and one can take bigger integration timesteps. On the iiwa hardware, we commonly use [800., 600, 600, 600, 400, 200, 200] Nm/rad as the stiffness values for the JointStiffnessController; but simulating with those requires small integration steps.

Admittance control Adaptive control. Slotine WAM throwing. Estimating contact location. contact particle filter. haddadin videos (e.g. http://www.diag.uniroma1.it/~deluca/IIT_Seminar_Jan23__ADL.pdf) De Luca pHRI https://www.youtube.com/playlist?list=PLvAUmIzqq6oaRtwX9l9sjDhcNMXNCGSN0

Abstract

Since , when the Facebook/Cambridge Analytica scandal began to emerge, public concern has grown around the threat of “online manipulation”. While these worries are familiar to privacy researchers, this paper aims to make them more salient to policymakers—first, by defining “online manipulation”, thus enabling identification of manipulative practices; and second, by drawing attention to the specific harms online manipulation threatens. We argue that online manipulation is the use of information technology to covertly influence another person’s decision-making, by targeting and exploiting their decision-making vulnerabilities. Engaging in such practices can harm individuals by diminishing their economic interests, but its deeper, more insidious harm is its challenge to individual autonomy. We explore this autonomy harm, emphasising its implications for both individuals and society, and we briefly outline some strategies for combating online manipulation and strengthening autonomy in an increasingly digital world.

Public concern is growing around an issue previously discussed predominantly amongst privacy and surveillance scholars—namely, the ability of data collectors to use information about individuals to manipulate them (e.g., Abramowitz, ; Doubek, ; Vayena, ). Knowing (or inferring) a person’s preferences, interests, and habits, their friends and acquaintances, education and employment, bodily health and financial standing, puts the knower in a position to exercise considerable influence over the known (Richards, ). It enables them to better understand what motivates their targets, what their weaknesses and vulnerabilities are, when they are most susceptible to influence and how most effectively to frame pitches and appeals. Because information technology makes generating, collecting, analysing, and leveraging such data about us cheap and easy, and at a scarcely comprehendible scale, the worry is that such technologies render us deeply vulnerable to the whims of those who build, control, and deploy these systems.

Initially, for academics studying this problem, that meant the whims of advertisers, as these technologies were largely developed by firms like Google and Facebook, who identified advertising as a means of monetising the troves of personal information they collect about internet users (Zuboff, ). Accordingly, for some time, scholarly worries centred (rightly) on commercial advertising practices, and policy solutions focused on modernising privacy and consumer protection regulations to account for the new capabilities of data-driven advertising technologies (e.g., Calo, ; Nadler & McGuigan, ; Turow, ). As Ryan Calo put it, “the digitization of commerce dramatically alters the capacity of firms to influence consumers at a personal level. A specific set of emerging technologies and techniques will empower corporations to discover and exploit the limits of each individual consumer’s ability to pursue his or her own self-interest” (, p. 999).

More recently, however, the scope of these worries has expanded. After concerns were raised in and about the use of information technology to influence elections around the world, many began to reckon with the fact that the threat of targeted advertising is not limited to the commercial sphere. By harnessing ad targeting platforms, like those offered by Facebook, YouTube, and other social media services, political campaigns can exert meaningful influence over the decision-making and behaviour of voters (Vaidhyanathan, ; Yeung, ; Zuiderveen Borgesius et al., ). Global outrage over the Cambridge Analytica scandal—in which the data analytics firm was accused of profiling voters in the United States, United Kingdom, France, Germany, and elsewhere, and targeting them with advertisements designed to exploit their “inner demons”—brought such worries to the forefront of public consciousness (“Cambridge Analytica and Facebook: The Scandal so Far”, ; see also, Abramowitz, ; Doubek, ; Vayena, ).

Indeed, there is evidence that the pendulum is swinging well to the other side. Rather than condemning the particular harms wrought in particular contexts by strategies of online influence, scholars are beginning to turn their attention to the big picture. In their recent book Re-Engineering Humanity, Brett Frischmann and Evan Selinger describe a vast array of related phenomena, which they collectively term “techno-social engineering”—i.e., “processes where technologies and social forces align and impact how we think, perceive, and act” (, p. 4). Operating at a grand scale reminiscent of mid-20th century technology critique (like that of Lewis Mumford or Jacques Ellul), Frischmann and Selinger point to cases of technologies transforming the way we carry out and understand our lives—from “micro-level” to the “meso-level” and “macro-level”— capturing everything from fitness tracking to self-driving cars to viral media (, p. 270). Similarly, in her book The Age of Surveillance Capitalism (), Shoshana Zuboff raises the alarm about the use of information technology to effectuate what she calls “behavior modification”, arguing that it has become so pervasive, so central to the functioning of the modern information economy, that we have entered a new epoch in the history of political economy.

These efforts help to highlight the fact that there is something much deeper at stake here than unfair commerce. When information about us is used to influence our decision-making, it does more than diminish our interests—it threatens our autonomy. At the same time, there is value in limiting the scope of the analysis. The notions of “techno-social engineering” and “surveillance capitalism” are too big to wield surgically—the former is intended to reveal a basic truth about the nature of our human relationship with technology, and the latter identifies a broad set of economic imperatives currently structuring technology development and the technology industry. Complementing this work, our intervention aims smaller. For the last several years, public outcry has coalesced against a particular set of abuses effectuated through information technology—what many refer to as “online manipulation” (e.g., Abramowitz, ; Doubek, ; Vayena, ). In what follows, we theorise and vindicate this grievance.

In the first section, we define manipulation, distinguishing it from neighbouring concepts like persuasion, coercion, deception, and nudging, and we explain why information technology is so well-suited to facilitating manipulation. In the second section, we describe the harms of online manipulation—the use of information technology to manipulate—focusing primarily on its threat to individual autonomy. Finally, we suggest directions for future policy efforts aimed at curbing online manipulation and strengthening autonomy in human-technology relations.

1. What is online manipulation?

The term “manipulation” is used, colloquially, to designate a wide variety of activities, so before jumping in it is worth narrowing the scope of our intervention further. In the broadest sense, manipulating something simply means steering or controlling it. We talk about doctors manipulating fine instruments during surgery and pilots manipulating cockpit controls during flight. “Manipulation” is also used to describe attempts at steering or controlling institutions and systems. For example, much has been written of late about allegations made (and evidence presented) that internet trolls under the authority of the Russian government attempted to manipulate the US media during the presidential election. Further, many suspect that the goal of those efforts was, in turn, to manipulate the election itself (by influencing voters). However, at the centre of this story, and at the centre of stories like it, is the worry that people are being manipulated, that individual decision-making is being steered or controlled, and that the capacity of individuals to make independent choices is therefore being compromised. It is manipulation in this sense—the attempt to influence individual decision-making and behaviour—that we focus on in what follows.

Philosophers and political theorists have long struggled to define manipulation. According to Robert Noggle, there are three main proposals (Noggle, b). Some argue that manipulation is non-rational influence (Wood, ). On that account, manipulating someone means influencing them by circumventing their rational, deliberative decision-making faculties. A classic example of manipulation understood in this way is subliminal messaging, and depending on one’s conception of rationality we might also imagine certain kinds of emotional appeals, such as guilt trips, as fitting into this picture. The second approach defines manipulation as a form of pressure, as in cases of blackmail (Kligman & Culver, , qtd. in Noggle, b). Here the idea is that manipulation involves some amount of force—a cost is extracted for non-compliance—but not so much force as to rise to the level of coercion. Finally, a third proposal defines manipulation as trickery. Although a variety of subtly distinct accounts fall under this umbrella, the main idea is that manipulation, at bottom, means leading someone along, inducing them to behave as the manipulator wants, like Iago in Shakespeare’s Othello, by tempting them, insinuating, stoking jealousy, and so on.

Each of these theories of manipulation has strengths and weaknesses, and our account shares certain features in common with all of them. It hews especially close to the trickery view, but operationalises the notion of trickery more concretely, thus offering more specific tools for diagnosing cases of manipulation. In our view, manipulation is hidden influence. Or more fully, manipulating someone means intentionally and covertly influencing their decision-making, by targeting and exploiting their decision-making vulnerabilities. Covertly influencing someone—imposing a hidden influence—means influencing them in a way they aren’t consciously aware of, and in a way they couldn’t easily become aware of were they to try and understand what was impacting their decision-making process.

Understanding manipulation as hidden influence helps to distinguish it from other forms of influence. In what follows, we distinguish it first from persuasion and coercion, and then from deception and nudging. Persuasion—in the sense of rational persuasion—means attempting to influence someone by offering reasons they can think about and evaluate. Coercion means influencing someone by constraining their options, such that their only rational course of action is the one the coercer intends (Wood, ). Persuasion and coercion carry very different, indeed nearly opposite, normative connotations: persuading someone to do something is almost always acceptable, while coercing them almost always isn’t. Yet persuasion and coercion are alike in that they are both forthright forms of influence. When someone is trying to persuade us or trying to coerce us we usually know it. Manipulation, by contrast, is hidden—we only learn that someone was trying to steer our decision-making after the fact, if we ever find out at all.

What makes manipulation distinctive, then, is the fact that when we learn we have been manipulated we feel played. Reflecting back on why we behaved the way we did, we realise that at the time of decision we didn’t understand our own motivations. We were like puppets, strung along by a puppet master. Manipulation thus disrupts our capacity for self-authorship—it presumes to decide for us how and why we ought to live. As we discuss in what follows, this gives rise to a specific set of harms. For now, what is important to see is the kind of influence at issue here. Unlike persuasion and coercion, which address their targets openly, manipulation is covert. When we are coerced we are usually rightly upset about it, but the object of our indignation is the set of constraints placed upon us. When we are manipulated, by contrast, we are not constrained. Rather, we are directed, outside our conscious awareness, to act for reasons we can’t recognise, and toward ends we may wish to avoid.

Given this picture, one can detect a hint of deception. On our view, deception is a special case of manipulation—one way to covertly influence someone is to plant false beliefs. If, for example, a manipulator wanted their partner to clean the house, they could lie and tell them that their mother was coming for a visit, thereby tricking them into doing what they wanted by prompting them to make a rational decision premised on false beliefs. But deception is not the only species of manipulation; there are other ways to exert hidden influence. First, manipulators need not focus on beliefs at all. Instead, they can covertly influence by subtly tempting, guilting, seducing, or otherwise playing upon desires and emotions. As long as the target of manipulation is not conscious of the manipulator’s strategy while they are deploying it, it is “hidden” in the relevant sense.

Some argue that even overt temptation, guilting, and so on are manipulative (these arguments are often made by proponents of the “non-rational influence” view of manipulation, described above), though they almost always concede that such strategies are more effective when concealed. We suspect that what is usually happening in such cases is a manipulator attempting to covertly tempt, guilt, etc., but failing to successfully hide their strategy. On our account, it is the attempted covertness that is central to manipulation, rather than the particular strategy, because once one learns that they are the target of another person’s influence that knowledge becomes a regular part of their decision-making process. We are all constantly subject to myriad influences; the reason we do not feel constantly manipulated is that we can usually reflect on, understand, and account for those influences in the process of reaching our own decisions about how to act (Raz, , p. 204). The influences become part of how we explain to ourselves why we make the decisions we do. When the influence is hidden, however, that process is undermined. Thus, while we might naturally call a person who frequently engages in overt temptation or seduction manipulative—meaning, they frequently attempt to manipulate—strictly speaking we would only say that they have succeeded in manipulating when their target is unaware of their machinations.

Second, behavioural economists have catalogued a long list of “cognitive biases”—unreliable mental shortcuts we use in everyday decision-making—which can be leveraged by would-be manipulators to influence the trajectory of our decision-making by shaping our beliefs, without the need for outright deception. Manipulators can frame information in a way that disposes us to a certain interpretation of the facts; they can strategically “anchor” our frame of reference when evaluating the costs or benefits of some decision; they can indicate to us that others have decided a certain way, in order to cue our intrinsic disposition to social conformity (the so-called “bandwagon effect”); and so on. Indeed, though deception and playing on people’s desires and emotions have likely been the most common forms of manipulation in the past—which is to say, the most common strategies for covertly influencing people—as we explain in what follows, there is reason to believe that exploiting cognitive biases and vulnerabilities is the most alarming problem confronting us today.

Talk of exploiting cognitive vulnerabilities inevitably gives rise to questions about nudging, thus finally, we briefly distinguish between nudging and manipulation. The idea of “nudging”, as is well known, comes from the work of Richard Thaler and Cass Sunstein, and points to any intentional alteration of another person’s decision-making context (their “choice architecture”) made in order to influence their decision-making outcome (Thaler & Sunstein, , p. 6). For Thaler and Sunstein, the fact that we suffer from so many decision-making vulnerabilities, that our decision-making processes are inalterably and unavoidably susceptible to even the subtlest cues from the contexts in which they are situated, suggests that when we design other people’s choice-making environments—from the apps they use to find a restaurant to the menus they order from after they arrive—we can’t help but influence their decisions. As such, on their account, we might as well use that power for good, by steering people’s decisions in ways that benefit them individually and all of us collectively. For these reasons, Thaler and Sunstein recommend a variety of nudges, from setting defaults that encourage people to save for retirement to arranging options in a cafeteria in way that encourages people to eat healthier foods.

Given our definition of manipulation as intentionally hidden influence, and our suggestion that influences are frequently hidden precisely by leveraging decision-making vulnerabilities like the cognitive biases nudge advocates reference, the question naturally arises as to whether or not nudges are manipulative. Much has been written on this topic and no consensus has been reached (see, e.g., Bovens, ; Hausman & Welch, ; Noggle, a; Nys & Engelen, ; Reach, ; Selinger & Whyte, ; Sunstein, ). In part, this likely has to do with the fact that a wide and disparate variety of changes to choice architectures are described as nudges. In our view, some are manipulative and some are not—the distinction hinging on whether or not the nudge is hidden, and whether it exploits vulnerabilities or attempts to rectify them. Many of the nudges Thaler and Sunstein, and others, recommend are not hidden and work to correct cognitive bias. For example, purely informational nudges, such as nutrition labels, do not seem to us to be manipulative. They encourage individuals to slow down, reflect on, and make more informed decisions. By contrast, Thaler and Sunstein’s famous cafeteria nudge—placing healthier foods at eye-level and less healthy foods below or above—seems plausibly manipulative, since it attempts to operate outside the individual’s conscious awareness, and to leverage a decision-making bias. Of course, just because it’s manipulative does not mean it isn’t justified. To say that a strategy is manipulative is to draw attention to the fact that it carries a harm, which we discuss in detail below. It is possible, however, that the harm is justified by some greater benefit it brings with it.

Having defined manipulation as hidden or covert influence, and having distinguished manipulation from persuasion, coercion, deception, and nudging, it is possible to define “online manipulation” as the use of information technology to covertly influence another person’s decision-making, by targeting and exploiting decision-making vulnerabilities. Importantly, we have adopted the term “online manipulation” from public discourse and interpret the word “online” expansively, recognising that there is no longer any hard boundary between online and offline life (if there ever was). “Online manipulation”, as we understand it, designates manipulation facilitated by information technology, and could just as easily be termed “digital manipulation” or “automated manipulation”. Since traditionally “offline” spaces are increasingly digitally mediated (because the people occupying them carry smartphones, the spaces themselves are embedded with internet-connected sensors, and so on), we should expect to encounter online manipulation beyond our computer screens.

Given this definition, it is not difficult to see why information technology is uniquely suited to facilitating manipulative influences. First, pervasive digital surveillance puts our decision-making vulnerabilities on permanent display. As privacy scholars have long pointed out, nearly everything we do today leaves a digital trace, and data collectors compile those traces into enormously detailed profiles (Solove, ). Such profiles comprise information about our demographics, finances, employment, purchasing behaviour, engagement with public services and institutions, and so on—in total, they often involve thousands of data points about each individual. By analysing patterns latent in this data, advertisers and others engaging in behavioural targeting are able to detect when and how to intervene in order to most effectively influence us (Kaptein & Eckles, ).

Moreover, digital surveillance enables detection of increasingly individual- or person-specific vulnerabilities. Beyond the well-known cognitive biases discussed above (e.g., anchoring and framing effects), which condition most people’s decision-making to some degree, we are also each subject to particular circumstances that can impact how we choose. We are each prone to specific fears, anxieties, hopes, and desires, as well as physical, material, and economic realities, which—if known—can be used to steer our decision-making. In , the voter micro-targeting firm Cambridge Analytica claimed to construct advertisements appealing to particular voter “psychometric” traits (such as openness, extraversion, etc.) by combining information about social media use with personality profiles culled from online quizzes. And in , an Australian newspaper exposed internal Facebook strategy documents detailing the company’s alleged ability to detect when teenage users are feeling insecure. According to the report, “By monitoring posts, pictures, interactions and internet activity in real-time, Facebook can work out when young people feel ‘stressed’, ‘defeated’, ‘overwhelmed’, ‘anxious’, ‘nervous’, ‘stupid’, ‘silly’, ‘useless’, and a ‘failure’” (Davidson, ). Though Facebook claims it never used that information to target advertisements at teenagers, it did not deny that it could. Extrapolating from this example it is easy to imagine others, such as banks targeting advertisements for high-interest loans at the financially desperate or pharmaceutical companies targeting advertisements for drugs at those suspected to be in health crisis.

Second, digital platforms, such as websites and smartphone applications, are the ideal medium for leveraging these insights into our decision-making vulnerabilities. They are dynamic, interactive, intrusive, and adaptive choice architectures (Lanzing, ; Susser, b; Yeung, ). Which is to say, the digital interfaces we interact with are configured in real time using the information about us described above, and they continue to learn about us as we interact with them. Unlike advertisements of old, they do not wait, passively, for viewers to drive past them on roads or browse over them in magazines; rather, they send text messages and push notifications, demanding our attention, and appear in our social media feeds at the precise moment they are most likely to tempt us. And because all of this is automated, digital platforms are able to adapt to each individual user, creating what Karen Yeung calls “highly personalised choice environment[s]”—decision-making contexts in which the vulnerabilities catalogued through pervasive digital surveillance are put to work in an effort to influence our choices (, p. 122).

Third, if manipulation is hidden influence, then digital technologies are ideal vehicles for manipulation because they are already in a real sense hidden. We often think of technologies as objects we attend to and use with focus and attention. The language of technology design reflects this: we talk about “users” and “end users,” “user interfaces,” and “human-computer interaction”. In fact, as philosophers (especially phenomenologists) and science and technology studies (STS) scholars have long shown, once we become habituated to a particular technology, the device or interface itself recedes from conscious attention, allowing us to focus on the tasks we are using it to accomplish. Think of a smartphone or computer: we pay little attention to the devices themselves, or even to the way familiar websites or app interfaces are arranged. Instead, after becoming acclimated to them, we attend to the information, entertainment, or conveniences they offer (Rosenberger, ). Philosophers refer to this as “technological transparency”—the fact that we see, hear, or otherwise perceive through technologies—as though they were clear, transparent—onto the perceptual objects they convey to us (Ihde, ; Van Den Eede, ; Verbeek, ). Because this language of transparency can be confused with the concept of transparency familiar from technology policy discussions, we might more helpfully describe it as “invisibility” (Susser, b). In addition to pervasive digital surveillance making our decision-making vulnerabilities easy to detect, and digital platforms making them easy to exploit, the ease with which our technologies become invisible to us—simply through frequent use and habituation—means the influences they facilitate are often hidden, and thus potentially manipulative.

Finally, although we focus primarily on the example of behavioural advertising to illustrate these dynamics, it is worth emphasising that advertisers are not the only ones engaging in manipulative practices. In the realm of user interface/experience (UI/UX) design, increasing attention is being paid to so-called “dark patterns”—design strategies that exploit users’ decision-making vulnerabilities to nudge them into acting against their interests (or, at least, acting in the interests of the website or app), such as requiring automatically-renewing paid subscriptions that begin after an initial free trial period (Brignull, ; Gray, Kou, Battles, Hoggatt, & Toombs, ; Murgia, ; Singer, ). Though many of these strategies are as old as the internet and not all rise to the level of manipulation—sometimes overtly inconveniencing users, rather than hiding their intentions—their growing prevalence has led some to call for legislation banning them (Bartz, ).

Worries about online manipulation have also been raised in the context of gig economy services, such as Uber and Lyft (Veen, Goods, Josserand, & Kaine, ). While these platforms market themselves as freer, more flexible alternatives to traditional jobs, providing reliable and consistent service to customers requires maintaining some amount of control over workers. However, without access to the traditional managerial controls of the office or factory floor, gig economy firms turn to “algorithmic management” strategies, such as notifications, customer satisfaction ratings, and other forms of soft control enabled through their apps (Rosenblat & Stark, ). Uber, for example, rather than requesting (or demanding) that workers put in longer hours, prompts drivers trying to exit the app with a reminder about their progress toward some earnings goal, exploiting the desire to continue making progress toward that goal; Lyft issues game-like “challenges” to drivers and stars and badges for accomplishing them (Mason, ; Scheiber, ).

In their current form, not all such practices necessarily manipulate—people are savvy, and many likely understand what they are facing. These examples are important, however, because they illustrate our present trajectory. Growing reliance on digital tools in all parts of our lives—tools that constantly record, aggregate, and analyse information about us—means we are revealing more and more about our individual and shared vulnerabilities. The digital platforms we interact with are increasingly capable of exploiting those insights to nudge and shape our choices, at home, in the workplace, and in the public sphere. And the more we become habituated to these systems, the less attention we pay to them.

2. The harm(s) of online manipulation

With this picture in hand, the question becomes: what exactly is the harm that results from influencing people in this way? Why should we be worried about technological mediation rendering us so susceptible to manipulative influence? In our view, there are several harms, but each flows from the same place—manipulation violates its target’s autonomy.

The notion of autonomy points to an individual’s capacity to make meaningfully independent decisions. As Joseph Raz puts it: “(t)he ruling idea behind the ideal of personal autonomy is that people should make their own lives” (Raz, , p. 369). Making one’s own life means freely facing both existential choices, like whom to spend one’s life with or whether to have children, and pedestrian, everyday ones. And facing them freely means having the opportunity to think about and deliberate over one’s options, considering them against the backdrop of one’s beliefs, desires, and commitments, and ultimately deciding for reasons one recognises and endorses as one’s own, absent unwelcome influence (J. P. Christman, ; Oshana, ; Veltman & Piper, ). Autonomy is in many ways the guiding normative principle of liberal democratic societies. It is because we think individuals can and should govern themselves that we value our capacity to collectively and democratically self-govern.

Philosophers sometimes operationalise the notion of autonomy by distinguishing between its competency and authenticity conditions (J. P. Christman, , p. 155f). In the first place, being autonomous means having the cognitive, psychological, social, and emotional competencies to think through one’s choices, form intentions about them, and act on the basis of those intentions. Second, it means that upon critical reflection one identifies with one’s values, desires, and goals, and endorses them authentically as one’s own. Of course, many have criticised such conceptions of autonomy as overly rationalistic and implausibly demanding, arguing that we rarely decide in this way. We are emotional actors and creatures of habit, they argue, socialised and enculturated into specific ways of choosing that we almost never reflect upon or endorse. But we understand autonomy broadly—our conception of deliberation includes not only beliefs and desires, but also emotions, convictions, and experiences, and critical reflection can be counterfactual (we must in principle be able to critically reflect on and endorse our motivations for acting, but we need not actually reflect on each and every move we make).

In addition to rejecting overly demanding and rationalistic conceptions of autonomy, we also reject overly atomistic ones. In our view, autonomous persons are socially, culturally, historically, and politically situated. Which is to say, we acknowledge the “intersubjective and social dimensions of selfhood and identity for individual autonomy and moral and political agency” (Mackenzie & Stoljar, , p. 4). Though social contexts can constrain our choices, by conditioning us to believe and behave in stereotypical ways (as, for example, in the case of gendered social expectations), it is also our social contexts that bestow value on autonomy, teaching us what it means to make independent decisions, and providing us with rich sets of options from which to choose. Moreover, it is crucial for present purposes that we emphasise our understanding of autonomy as more than an individual good—it is an essential social and political good too. Individuals express their autonomy across a variety of social contexts, from the home to the marketplace to the political sphere. Democratic institutions are meant to register and reflect the autonomous political decisions individuals make. Disrupting individual autonomy is thus more than an ethical concern; it has social and political import.

Against this picture of autonomy and its value, we can more carefully explain why online manipulation poses such a grave threat. To manipulate someone is, again, to covertly influence them, to intentionally alter their decision-making process without their conscious awareness. Doing so undermines the target’s autonomy in two ways: first, it can lead them to act toward ends they haven’t chosen, and second, it can lead them to act for reasons not authentically their own.

To see the first problem, consider examples of targeted advertising in the commercial sphere. Here, the aim of manipulators is fairly straightforward: they want people to buy things. Rather than simply put products on display, however, advertisers can construct decision-making environments—choice architectures—that subtly tempt or seduce shoppers to purchase their wares, and at the highest possible price (Calo, ). A variety of strategies might be deployed, from pointing out that one’s friends have purchased the item to countdown clocks that pressure one to act before some offer expires, the goal being to hurry, evade, or undermine deliberation, and thus to encourage decisions that may or may not align with an individual’s deeper, reflective, self-chosen ends and values.

Of course, these strategies are familiar from non-digital contexts; all commercial advertising (digital or otherwise) functions in part to induce consumers to buy things, and worries about manipulative ads emerged long before advertising moved online. Equally, not all advertising—perhaps not even all targeted advertising—involves manipulation. Purely informational ads displayed to audiences actively seeking out related products and services (e.g., online banner ads displaying a doctor’s contact information shown to visitors to a health-related website) are unlikely to covertly influence their targets. Worries about manipulation arise in cases where advertisements are sneaky—which is to say, where their effects are achieved covertly. If, for example, the doctor was a psychiatrist, his advertisements were shown to people suspected of suffering from depression, and only at the specific times of day they were thought to be most afflicted, our account would offer grounds for condemning such tactics as manipulative.

It might also be the case that manipulation is not a binary phenomenon. We are the objects of countless influence campaigns and we understand some of them more than others; perhaps we ought to say that they are more or less manipulative in equal measure. On such a view, online targeted (or “behavioural”) advertising could be understood as exacerbating manipulative dynamics common to other forms of advertising, by making the tweaks to individual choice architectures more subtle, and the seductions and temptations that result from them more difficult to resist (Yeung, ). Worse still, the fluidity and porousness of online environments makes it easy for marketers to conflate other distinct contexts with shopping, further blurring a person’s reasoning about whether they truly want to make some purchase. For example, while chatting with friends over social media or searching for some place to eat, an ad may appear, thus requiring the target to juggle several tasks—in this case, communication and information retrieval—along with deliberation over whether or not to respond to the marketing ploy, thus diminishing the target’s ability to sustain focus on any of the them. This problem is especially clearly illustrated by so-called “native advertising” (advertisements designed to look like user-generated, non-commercial content). Such advertisements are a kind of Trojan horse, intentionally conflating commercial and non-commercial activities in an attempt to undermine our capacity for focused, careful deliberation.

In the philosophical language introduced above, these strategies challenge both autonomy’s competency and authenticity conditions. By deliberately and covertly engineering our choice environments to steer our decision-making, online manipulation threatens our competency to deliberate about our options, form intentions about them, and act on the basis of those intentions. And since, as we’ve seen, manipulative practices often work by targeting and exploiting our decision-making vulnerabilities—concealing their effects, leaving us unaware of the influence on our decision-making process—they also challenge our capacity to reflect on and endorse our reasons for acting as authentically on our own. Online manipulation thus harms us both by inducing us to act toward ends not of our choosing and for reasons we haven’t endorsed.

Importantly, undermining personal autonomy in the ways just described can lead to further harms. First, since autonomous individuals are wont to protect (or at least to try and protect) their own interests, we can reasonably expect that undermining people’s autonomy will lead, in many cases, to a diminishment of those interests. Losing the ability to look out for ourselves is unlikely to leave us better off in the long run. This harm—e.g., being tricked into buying things we don’t need or paying more for them than we otherwise would—is well described by those who have analysed the problem of online manipulation in the commercial sphere (Calo, ; Nadler & McGuigan, ; Zarsky, ; Zarsky, ). And it is a serious harm, which we would do well to take seriously, especially given the fact that law and policy around information and internet practices (at least in the US) assume that individuals are for the most part capable of safeguarding their interests (Solove, ). However, it is equally important to see that this harm to welfare is derivative of the deeper harm to autonomy. Attempting to “protect consumers” from threats to their economic or other interests, without addressing the more fundamental threat to their autonomy, is thus to treat the symptoms without addressing the cause.

To bring this into sharper relief, it is worth pointing out that even purely beneficent manipulation is harmful. Indeed, it is harmful to manipulate someone even in an effort to lead them more effectively toward their own self-chosen ends. That is because the fundamental harm of manipulation is to the process of decision-making, not its outcome. A well-meaning, paternalistic manipulator, who subtly induces his target to eat better food, exercise, and work hard, makes his target better off in one sense—he is healthier and perhaps more materially well-off—but it harms him as well by rendering him opaque to himself. Imagine if some bad habit, which someone had spent their whole life attempting to overcome, one day, all of a sudden, disappeared. They would be happy, of course, to be rid of the habit, but they might also be deeply confused and suspicious about the source of the change. As T.M. Scanlon writes, “I want to choose the furniture for my own apartment, pick out the pictures for the walls, and even write my own lectures despite the fact that these things might be done better by a decorator, art expert, or talented graduate student. For better or worse, I want these things to be produced by and reflect my own taste, imagination, and powers of discrimination and analysis. I feel the same way, even more strongly, about important decisions affecting my life in larger terms: what career to follow, where to work, how to live” (Scanlon, ).

Having said that, we have not demonstrated that manipulation is necessarily wrong in every case—only that it always carries a harm. One can imagine cases where the harm to autonomy is outweighed by the benefit to welfare. (For example, a case where someone’s life is in immediate danger, and the only way to save them is by manipulating them.) But such cases are likely few and far between. What is so worrying about online manipulation is precisely its banality—the fact that it threatens to become a regular part of the fabric of everyday experience. As Jeremy Waldron argues, if we allow that to happen, our lives will be drained of something deeply important: “What becomes of the self-respect we invest in our own willed actions, flawed and misguided though they often are, when so many of our choices are manipulated to promote what someone else sees (perhaps rightly) as our best interest?” (Waldron, ) That we also lack reason to believe online manipulators really do have our best interests at heart is only more reason to resist them.

Finally, beyond the harm to individuals, manipulation promises a collective harm. By threatening our autonomy it threatens democracy as well. For autonomy is writ small what democracy is writ large—the capacity to self-govern. It is only because we believe individuals can make meaningfully independent decisions that we value institutions designed to register and reflect them. As the Cambridge Analytica case—and the public outcry in response to it—demonstrates, online manipulation in the political sphere threatens to undermine these core collective values. The problem of online manipulation is, therefore, not simply an ethical problem; it is a social and political one too.

3. Technology and autonomy

If one accepts the arguments advanced thus far, an obvious response is that we need to devise law and policy capable of preventing and mitigating manipulative online practices. We agree that we do. But that response is not sufficient—the question for policymakers is not simply how to mitigate online manipulation, but how to strengthen autonomy in the digital age. In making this claim, we join our voices with a growing chorus of scholars and activists—like Frischmann, Selinger, and Zuboff—working to highlight the corrosive effects of digital technologies on autonomy. Meeting these challenges requires more than consumer protection—it requires creating the positive conditions necessary for supporting individual and collective self-determination.

We don’t pretend to have a comprehensive solution to these deep and complex problems, but some suggestions follow from our brief discussion. It should be noted that these suggestions—like the discussion, above, that prompted them—are situated firmly in the terrain of contemporary liberal political discourse, and those convinced that online manipulation poses a significant threat (especially some European readers) may be struck by how moderate our responses are. While we are not opposed to more radical interventions, we formulate our analysis using the conceptual and normative frameworks familiar to existing policy discussions in hopes of having an impact on them.

Curtail digital surveillance

Data, as Tal Zarsky writes, is the “fuel” powering online manipulation (, p. 186). Without the detailed profiles cataloguing our preferences, interests, habits, and so on, the ability of would-be manipulators to identify our weaknesses and vulnerabilities would be vastly diminished, and so too their capacity to leverage them to their ends. Of course, the call to curtail digital surveillance is nothing new. Privacy scholars and advocates have been raising alarms about the ills of surveillance for half a century or more. Yet, as Zarsky argues, manipulation arguments could add to the “analytic and doctrinal arsenal of measures which enable legal intervention in the new digital environment” (, p. 185). Furthermore, outcry over apparent online manipulation in both the commercial and political spheres appears to be generating momentum behind new policy interventions to combat such strategies. In the US, a number of states have recently passed or are considering passing new privacy legislation, and the U.S. Congress appears to be weighing new federal privacy legislation as well. (“Congress Is Trying to Create a Federal Privacy Law”, ; Merken, ). And, of course, all of that takes place on the heels of the new General Data Protection Regulation (GDPR) taking effect in Europe, which places new limits on when and what kinds of data can be collected about European citizens and by firms operating on European soil. To curb manipulation and strengthen autonomy online, efforts to curtail digital surveillance ought to be redoubled.

Problematise personalisation

When asked to justify collecting so much data about us, data collectors routinely argue that the information is needed in order to personalise their services to the needs and interests of individual users. Mark Zuckerberg, for example, attempted recently to explain Facebook’s business model in the pages of the Wall Street Journal: “People consistently tell us that if they're going to see ads, they want them to be relevant,” he wrote. “That means we need to understand their interests” (). Personalisation seems, on the face of it, like an unalloyed good. Who wouldn’t prefer a personalised experience to a generic one? Yet research into different forms of personalisation suggests that individualising—personalising—our experiences can carry with it significant risks.

These worries came to popular attention with Eli Pariser’s book Filter Bubble (), which argued forcefully (though not without challenge) that the construction of increasingly singular, individualised experiences, means at the same time the loss of common, shared ones, and describes the detriments of that transformation to both individual and collective decision-making. In addition to personalised information environments—Pariser’s focus—technological advances enable things like personalised pricing - sometimes called “dynamic pricing” or “price discrimination” (Calo, ) and personalised work scheduling - or “just-in-time” scheduling (De Stefano, ). For the reasons discussed above, many such strategies may well be manipulative. The targeting and exploiting of individual decision-making vulnerabilities enabled by digital technologies—the potential for online manipulation they create—gives us reason to question whether the benefits of personalisation really outweigh the costs. At the very least, we ought not to uncritically accept personalisation as a rationale for increased data collection, and we ought to approach with care (if not skepticism) the promise of an increasingly personalised digital environment.

Promote awareness and understanding

If the central problem of online manipulation is its hiddenness, then any response must involve a drive toward increased awareness. The question is what form such awareness should take. Yeung argues that the predominant vehicle for notifying individuals about information flows and data practices—the privacy notice, or what is often called “notice-and-consent”—is insufficient (). Indeed, merely notifying someone that they are the target of manipulation is not enough to neutralise its effects. Doing so would require understanding not only that one is the target of manipulation, but also who the manipulator is, what strategies they are deploying, and why. Given the well-known “transparency paradox”, according to which we are bound to either deprive users of relevant information (in an attempt to be succinct) or overwhelm them with it (in an attempt to be thorough), there is little reason to believe standard forms of notice alone can equip users to face the challenges of online manipulation.

Furthermore, the problem of online manipulation runs deeper than any particular manipulative practice. What worries many people is the fact that manipulative strategies, like targeted advertising, are becoming basic features of the digital world—so commonplace as to escape notice or mention. In the same way that machine learning and artificial intelligence tools have quickly and quietly been delegated vast decision-making authorities in a variety of contemporary contexts and institutions, and in response, scholars and activists have mounted calls to make their decision-making processes more explainable, transparent, and accountable, so too must we give people tools to understand and manage a digital environment designed to shape and influence them.

Attend to context

Finally, it is important to recognise that moral intuitions about manipulation are indexed to social context. Which is to say, we are willing to tolerate different levels of outside influence on our decision-making in different decision-making spheres. As relatively lax commercial advertising regulations indicate, we are—at least in the US—willing to accept a fair amount of interference in the commercial sphere. By contrast, somewhat more stringent regulations around elections and campaign advertising suggest that we are less willing to accept such interference in the realm of politics. Responding to the threats of online manipulation therefore requires sensitivity to where—in which spheres of life—we encounter them.

Conclusion

The idea that technological advancements bring with them new arrangements of power is, of course, nothing new. That online manipulation threatens to subordinate the interests of individuals to those of data collectors and their clients is thus, in one respect, a familiar (if nonetheless troubling) problem. What we hope to have shown, however, is that the threat of online manipulation is deeper, more insidious, than that. Being steered or controlled, outside our conscious awareness, violates our autonomy, our capacity to understand and author our own lives. If the tools that facilitate such control are left unchecked, it will be to our individual and collective detriment. As we’ve seen, information technology is in many ways an ideal vehicle for these forms of control, but that does not mean that they are inevitable. Combating online manipulation requires both depriving it of personal data—the oxygen enabling it—and empowering its targets with awareness, understanding, and savvy about the forces attempting to influence them.

Contact us to discuss your requirements of electronic stamping robots. Our experienced sales team can help you identify the options that best suit your needs.