From end-effector pose to joint angles

With its obvious importance in robotics, you probably won't be surprised to hear that there is an extensive literature on inverse kinematics. But you may be surprised at how extensive and complete the solutions can get. The forward kinematics, $f_{kin}$, is a nonlinear function in general, but it is a very structured one. In fact, with rare exceptions (like if your robot has a helical joint, aka screw joint), the equations governing the valid Cartesian positions of our robots are actually polynomial. "But wait! What about all of those sines and cosines in my kinematic equations?" you say. The trigonometric terms come when we want to relate joint angles with Cartesian coordinates. In $\Re^3$, for two points, $A$ and $B$, on the same rigid body, the (squared) distance between them, $\|p^A - p^B\|^2,$ is a constant. And a joint is just a polynomial constraint between positions on adjoining bodies, e.g. that they occupy the same point in Cartesian space. See Wampler11 for an excellent overview.

Understanding the solutions to polynomial equations is the subject of algebraic geometry. There is a deep literature on kinematics theory, on symbolic algorithms, and on numerical algorithms. For even very complex kinematic topologies, such as four-bar linkages and Stewart-Gough platforms, we can count the number of solutions, and/or understand the continuous manifold of solutions. For instance, Wampler11 describes a substantial toolbox for numerical algebraic geometry (based on homotopy methods) with impressive results on difficult kinematics problems.

The kinematics algorithms based on algebraic geometry have traditionally been targeted for offline global analysis, and are often not designed for fast real-time inverse kinematics solutions needed in a control loop. The most popular tool these days for real-time inverse kinematics for six- or seven-DOF manipulators is a tool called "IKFast", described in Section 4.1 of Diankov10, that gained popularity because of its effective open-source implementation. Rather than focus on completeness, IKFast uses a number of approximations to provide fast and numerically robust solutions to the "easy" kinematics problems. It leverages the fact that a six-DOF pose constraint on a six-DOF manipulator has a "closed-form" solution (for most serial-chain robot arms) with a finite number of joint space configurations that produce the same end-effector pose, and for seven-DOF manipulators it adds a layer of sampling in the last degree of freedom on top of the six-DOF solver.

These explicit solutions are important to understand because they provide deep insight into the equations, and because they can be fast enough to use inside a more sophisticated solution approach. But the solutions don't immediately provide the rich specification I advocated for above; in particular, they break down once we have inequality constraints instead of equality constraints. For those richer specifications, we will turn to optimization.

IK as constrained optimization

Rather than formulate inverse kinematics as $$q = f^{-1}_{kin}(X^G),$$ let's consider solving the same problem as an optimization: \begin{align} \min_q & \quad |q - q_0|^2, \\ \subjto &\quad X^G = f_{kin}(q), \end{align} where $q_0$ is some comfortable nominal position. While writing the inverse directly is a bit problematic, especially because we sometimes have multiple (even infinite) solutions or no solutions. This optimization formulation is slightly more precise -- if we have multiple joint angles which achieve the same end-effector position, then we prefer the one that is closest to the nominal joint positions. But the real value of switching to the optimization perspective of the problem is that it allows us to connect to a rich library of additional costs and constraints.

We have already discussed the idea of solving differential inverse kinematics as an optimization problem. In that workflow, we started by using the pseudo-inverse of the kinematic Jacobian, but then graduated to thinking about the least-squares formulation of the inverse problem. The more general least-squares setting, we could add additional costs and constraints that would protect us from (nearly) singular Jacobians and could take into account additional constraints from joint limits, joint velocity limits, etc. We could even add collision avoidance constraints. Some of these constraints are quite nonlinear / nonconvex functions of the configuration $q$, but in the differential kinematics setting we were only seeking to find a small change $\Delta q$ around the nominal configuration, so it was quite reasonable to make linear/convex approximations of these nonlinear/nonconvex constraints.

Now we will consider the full formulation, where we try to solve the nonlinear / nonconvex optimization directly, without any constraints on only making a small change to an initial $q$. This is a much harder problem computationally. Using powerful nonlinear optimization solvers like SNOPT, we are often able to solve the problems, even at interactive rates (the example below is quite fun). But there are no guarantees. It could be that a solution exists even if the solver returns "infeasible".

Of course, the differential IK problem and the full IK problem are closely related. In fact, you can think about the differential IK algorithm as doing one step of (projected) gradient descent or one-step of Sequential Quadratic Programming, for the full nonlinear problem.

Drake provides a nice InverseKinematics class that makes it easy to assemble many of the standard kinematic/multibody constraints into a MathematicalProgram. Take a minute to look at the constraints that are offered. You can add constraints on the relative position and/or orientation on two bodies, or that two bodies are more than some minimal distance apart (e.g. for non-penetration) or closer than some distance, and more. This is the way that I want you to think about the IK problem; it is an inverse problem, but one with a potentially very rich set of costs and constraints.

Interactive IK

Despite the nonconvexity of the problem and nontrivial computational cost of evaluating the constraints, we can often solve it at interactive rates. I've assembled a few examples of this in the chapter notebook:

In the first version, I've added sliders to let you control the desired pose of the end-effector. This is the simple version of the IK problem, amenable to more explicit solutions, but we nevertheless solve it with our full nonlinear optimization IK engine (and it does include joint limit constraints). This demo won't look too different from the very first example in the notes, where you used teleop to command the robot to pick up the red brick. In fact, differential IK offers a fine solution to this problem, too.

In the second example, I've tried to highlight the differences between the nonlinear IK problem and the differential IK problem by adding an obstacle directly in front of the robot. Because both our differential IK and IK formulations are able to consume the collision-avoidance constraints, both solutions will try to prevent you from crashing the arm into the post. But if you move the target end-effector position from one side of the post to the other, the full IK solver can switch over to a new solution with the arm on the other side of the post, but the differential IK will never be able to make that leap (it will stay on the first side of the post, refusing to allow a collision).

With great power comes great responsibility. The inverse kinematics toolbox allows you to formulate complex optimizations, but your success with solving them will depend partially on how thoughtful you are about choosing your costs and constraints. My basic advice is this:

1. Try to keep the objective (costs) simple; I typically only use the "joint-centering" quadratic cost on $q$. Putting terms that should be constraints into the cost as penalties leads to lots of cost-function tuning, which can be a nasty business.
2. Write minimal constraints. You want the set of feasible configurations to be as big as possible. For instance, if you don't need to fully constrain the orientation of the gripper, then don't do it.

Grasp the cylinder

Let's use IK to grasp a cylinder. You can think of it as a hand rail, if you prefer. Suppose it doesn't matter where along the cylinder we grasp, nor the orientation at which we grasp it. Then we should write the IK problem using only the minimal version of those constraints.

In the notebook, I've coded up one version of this. I've put the cylinder's pose on the sliders now, so you can move it around the workspace, and watch how the IK solver decides to position the robot. In particular, if you move the cylinder in $\pm y$, you'll see that the robot doesn't try to follow... until the hand gets to the end of the cylinder. Very nice!

One could imagine multiple ways to implement that constraint. Here's how I did it:

ik.AddPositionConstraint(
    frameB=gripper_frame, p_BQ=[0, 0.1, -0.02],
    frameA=cylinder_frame, p_AQ_lower=[0, 0, -0.5], p_AQ_upper=[0, 0, 0.5])
ik.AddPositionConstraint(
    frameB=gripper_frame, p_BQ=[0, 0.1, 0.02], 
    frameA=cylinder_frame, p_AQ_lower=[0, 0, -0.5], p_AQ_upper=[0, 0, 0.5])

In words, I've defined two points in the gripper frame; in the notation of the AddPositionConstraint method they are ${}^Bp^{Q}$. Recall the gripper frame is such that $[0, .1, 0]$ is right between the two gripper pads; you should take a moment to make sure you understand where $[0,.1,-0.02]$ and $[0,.1,0.02]$ are. Our constraints require that both of those points should lie exactly on the center line segment of the cylinder. This was a compact way for me to leave the orientation around the cylinder axis as unconstrained, and capture the cylinder position constraints all quite nicely.

We've provided a rich language of constraints for specifying IK problems, including many which involve the kinematics of the robot and the geometry of the robot and the world (e.g., the minimum-distance constraints). Let's take a moment to appreciate the geometric puzzle that we are asking the optimizer to solve.

Visualizing the configuration space

Let's return to the example of the iiwa reaching into the shelf. This IK problem has two major constraints: 1) we want the center of the target sphere to be in the center of the gripper, and 2) we want the arm to avoid collisions with the shelves. In order to plot these constraints, I've frozen three of the joints on the iiwa, leaving only the three corresponding motion in the $x-z$ plane.

To visualize the constraints, I've sampled a dense grid in the three joint angles of the planarized iiwa, assigning each grid element to 1 if the constraint is satisfied or 0 otherwise, then run a marching cubes algorithm to extract an approximation of the true 3D geometry of this constraint in the configuration space. The "grasp the sphere" constraint produces the nice green geometry I've pictured above on the right; it is clipped by the joint limits. The collision-avoidance constraint, on the other hand, is quite a bit more complicated. To see that, you'd better open up this 3D visualization so you can navigate around it yourself. Scary!

Global inverse kinematics

For unconstrained inverse kinematics with exactly six degrees of freedom, we have closed-form solutions. For the generalized inverse kinematics problem with rich costs and constraints, we've got a nonlinear optimization problem that works well in practice but is subject to local minima (and therefore can fail to find a feasible solution if it exists). If we give up on solving the optimization problem at interactive rates, is there any hope of solving the richer IK formulation robustly? Ideally to global optimality?

This is actually an extremely relevant question. There are many applications of inverse kinematics that work offline and don't have any strict timing requirements. Imagine if you wanted to generate training data to train a neural network to learn your inverse kinematics; this would be a perfect application for global IK. Or if you want to do workspace analysis to see if the robot can reach all of the places it needs to reach in the workspace that you're designing for it, then you'd like to use global IK. Some of the motion planning strategies that we'll study below will also separate their computation into an offline "building" phase to make the online "query" phase much faster.

In my experience, general-purpose global nonlinear solvers -- for instance, based on mixed-integer nonlinear programming (MINLP) approaches or the interval arithmetic used in satisfiability-modulo-theories (SMT) solvers -- typically don't scale the complexity of a full manipulator. But if we only slightly restrict the class of costs and constraints that we support, then we can begin to make progress.

Drake provides an implementation of the GlobalInverseKinematics approach described in Dai17 using mixed-integer convex optimization. The solution times are on the order of a few seconds; it can solve a full constrained bimanual problem in well under a minute. Trutman22 solves the narrow version of the problem (just end-effector poses and joint limits) using convex optimization via a hierarchy of semi-definite programming relaxations; it would be very interesting to understand how well this approach works with the larger family of costs and constraints.

Inverse kinematics vs differential inverse kinematics

When should we use IK vs Differential IK? IK solves a more global problem, but is not guaranteed to succeed. It is also not guaranteed to make small changes to $q$ as you make small changes in the cost/constraints; so you might end up sending large $\Delta q$ commands to your robot. Use IK when you need to solve the more global problem, and the trajectory optimization algorithms we produce in the next section are the natural extension to producing actual $q$ trajectories. Differential IK works extremely well for incremental motions -- for instance if you are able to design smooth commands in end-effector space and simply track them in joint space.

Grasp planning using inverse kinematics

In our first version of grasp selection using sampling, we put an objective that rewarded grasps that were oriented with the hand grasping from above the object. This was a (sometimes poor) surrogate for the problem that we really wanted to solve: we want the grasp to be achievable given a "comfortable" position of the robot. So a simple and natural extension of our grasp scoring metric would be to solve an inverse kinematics problem for the grasp candidate, and instead of putting costs on the end-effector orientation, we can use the joint-centering cost directly as our objective. Furthermore, if the IK problem returns infeasible, we should reject the sample.

There is a quite nice extension of this idea that becomes natural once we take the optimization view, and it is a nice transition to the trajectory planning we'll do in the next section. Imagine if the task requires us not only to pick up the object in clutter, but also to place the object carefully in clutter as well. In this case, a good grasp involves a pose for the hand relative to the object that simultaneously optimizes both the pick configuration and the place configuration. One can formulate an optimization problem with decision variables for both $q_{pick}$ and $q_{place}$, with constraints enforcing that ${}^OX^{G_{pick}} = {}^OX^{G_{place}}$. Of course we can still add in all of our other rich costs and constraints, In the dish-loading project at TRI, this approach proved to be very important. Both picking up a mug in the sink and placing it in the dishwasher rack are highly constrained, so we needed the simultaneous optimization to find successful grasps.

Trajectory parameterizations

The interesting question, then, becomes how do we actually parameterize the trajectory $q(t)$ with a parameter vector $\alpha$? These days, you might think that $q_\alpha(t)$ could be a neural network that takes $t$ as an input, offers $q$ as an output, and uses $\alpha$ to represent the weights and biases of the network. Of course you could, but for inputs with a scalar input like this, we often take much simpler and sparser parameterizations, often based on polynomials.

There are many ways one can parameterize a trajectory with polynomials. For example in dynamic motion planning, direct collocation methods use piecewise-cubic polynomials to represent the state trajectory, and the pseudo-spectral methods use Lagrange polynomials. In each case, the choice of basis functions is made so that algorithm can leverage a particular property of the basis. In dynamic motion planning, a great deal of focus is on the integration accuracy of the dynamic equations to ensure that we obtain feasible solutions to the dynamic constraints.

When we are planning the motions of our fully-actuated robot arms, we typically worry less about dynamic feasibility, and focus instead on the kinematics. For kinematic trajectory optimization, the so-called B-spline trajectory parameterization has a few particularly nice properties that we can leverage here:

• The derivative of a B-spline is still a B-spline (the degree is reduced by one), with coefficients that are linear in the original coefficients.
• The bases themselves are non-negative and sparse. This gives the coefficients of the B-spline polynomial, which are referred to as the control points, a strong geometric interpretation.
• In particular, the entire trajectory is guaranteed to lie inside the convex hull of the active control points (the control points who's bases are not zero).

Note that B-splines are closely related to Bézier curves. But the "B" in "B-spline" actually just stands for "basis" (no, I'm not kidding) and "Bézier splines" are slightly different.

Optimization algorithms

The default KinematicTrajectoryOptimization class in Drake optimizes a trajectory defined using a B-spline to represent a path, $r(s)$ over the interval $s \in [0,1]$, plus an additional scalar decision variable corresponding to the trajectory duration, $T$. The final trajectory combines the path with the time-rescaling: $q(t) = r(t/T).$ This is a particularly nice way to represent a trajectory of unknown duration, and has the excellent feature that the convex hull property can still be used. Velocity constraints are still linear; constraints on acceleration and higher derivatives do become nonlinear, but if satisfied they still imply strict bounds $\forall t \in [0, T].$

Since the KinematicTrajectoryOptimization is written using Drake's MathematicalProgram, by default it will automatically select what we think is the best solver given the available solvers. If the optimization has only convex costs and constraints, it will be dispatched to a convex optimization solver. But most often we add in nonconvex costs and constraints from kinematics. Therefore in most cases, the default solver would again be the SQP-solver, SNOPT. You are free to experiment with others!

One of the most interesting set of constraints that we can add to our kinematic trajectory optimization problem is the MinimumDistanceLowerBoundConstraint; when the minimum distance between all potential collision pairs is greater than zero then we have avoided collisions. Please note, though, that these collision constraints can only be enforced at discrete samples, $s_i \in [0,1]$, along the path. They do not guarantee that the trajectory is collision free $\forall t\in[0,T].$ It required very special properties of the derivative constraints to leverage the convex hull property; we do not have them for more general nonlinear constraints. A common strategy is to add constraints at some modest number of samples along the interval during optimization, then to check for collisions more densely on the optimized trajectory before executing it on the robot.

There are a number of related approaches to kinematic trajectory optimization in the motion planning literature, which differ either by their parameterization or by the solution technique (or both). Some of the more well-known include CHOMP Ratliff09a, STOMP Kalakrishnan11a, and KOMOToussaint17.

KOMO, for instance, is one of a handful of trajectory optimization techniques that use the Augmented Lagrangian method of transcribing a constrained optimization problem into an unconstrained problem, then using a simple but fast gradient-based solverToussaint14. Augmented-Lagrangian-based approaches appear to be the most popular and successful these days; I hope to provide a nice implementation in Drake soon! But one does have to be careful in comparing these different solvers -- SNOPT may declare failure if it cannot optimize the cost and satisfy constraints to the default tolerances (around 1e-6), while an Augmented-Lagrangian approach will never report failure and will rarely aim for this level of accuracy.

When kinematic trajectory optimizations succeed, they are an incredibly satisfying solution to the motion planning problem. They give a very natural and expressive language for specifying the desired motion (with needing to sample nonlinear constraints as perhaps the one exception), and they can be solved fast enough for online planning. The only problem is: they don't always succeed. Because they are based on nonconvex optimization, they are susceptible to local minima, and can fail to find a feasible path even when one exists.

Local minima in collision-free trajectory optimization

Consider the extremely simple example of finding the shortest path for a point robot moving from the start (blue circle) to the goal (green circle) in the image below, avoiding collisions with the red box. Avoiding even the complexity of B-splines, we can write an optimization of the form: \begin{aligned} \min_{q_0, ..., q_N} \quad & \sum_{n=0}^{N-1} | q_{n+1} - q_n|_2^2 & \\ \text{subject to} \quad & q_0 = q_{start} \\ & q_N = q_{goal} \\ & |q_n|_1 \ge 1 & \forall n, \end{aligned} where the last line is the collision-avoidance constraint saying that each sample point has to be outside of the $\ell_1$-ball (my convenient choice for the geometry of the red box). Alternatively, we can write a slightly more advanced constraint to maintain that each line segment is outside of the obstacle (rather than just the vertices). Here are some possible solutions to this optimization problem:

Fix meshcat line width.

The solution on the left is the (global) minimum for the problem; the solution on the right is clearly a local minima. Once a nonlinear solver is considering paths that go right around the obstacle, it is very unlikely to find a solution that goes left around the obstacle, because the solution would have to get worse (violate the collision constraint) before it gets better.

Make a version of this with B-splines

To deal with this limitation, the field of collision-free motion planning has trended heavily towards sampling-based methods.

Rapidly-exploring random trees (RRT)

The Probabilistic Roadmap (PRM)

Post-processing

Short-cutting, anytime b-spline smoother

Sampling-based planning in practice

There are many optimizations and heuristics which can dramatically improve the effectiveness of these methods... (optimized collision checking, weighted Euclidean distances, ...)

Graphs of Convex Sets

The GCS framework was originally introduced in Marcucci21, and we have a mature implementation in Drake. It is a general optimization framework for combining combinatorial optimizations, like graph search, with continuous optimizations, like the kinematic trajectory optimization we studied above. GCS can provide a continuous extension to any "network flow" optimization (c.f. Ahuja93), but for motion planning, the first problem we are interested in is shortest paths on a graph.

Consider the very classical problem of finding the (weighted) shortest path from a source, $s$, to a target, $t,$ on a graph, pictured on the left. The problem is described by a set of vertices, a set of (potentially directed) edges, and the cost of traversing each edge.

The shortest path problem on a graph of convex sets can encode problems that are NP-Hard, but these can be formulated as a mixed-integer convex optimization (MICP). What makes the framework so powerful is that this MICP has a very strong and efficient convex relaxation; meaning that if you relax the binary variables to continuous variables you get (nearly) the solution to the original MICP. This means that you can solve GCS problems to global optimality orders of magnitude faster than previous transcriptions. But in practice, we find that solving only the convex relaxation (plus a little rounding) is almost always sufficient to recover the optimal solution. Nowadays, we almost never solve the full MICP in our robotics applications of GCS.

You can find more details and examples about GCS in my underactuated notes, and a thorough and very readable treatment of GCS as a general tool for mixed-discrete and continous optimization in Marcucci24a.

GCS (Kinematic) Trajectory Optimization

GCS provides a very general machinery for optimizing over the mixed discrete/continuous optimization problems that are naturally represented on a graph. But how do we transcribe the motion planning problem into a GCS? We first presented the transcription I'll describe here in the paper titled "Motion planning around obstacles with convex optimization" Marcucci23a. Since we've already spent time in this chapter talking about the inherent nonconvexity of the motion planning problem, you can see why we like this title! It should be slightly surprising that we can use convex optimization to effectively solve these problems to global optimality.

Let us assume for a moment that we are given a convex decomposition of the collision-free space. This is a strong assumption, which we will justify below.

Observe that if we have two points in the same convex collision-free region, then we know that the straight line connecting them is also guaranteed to be collision free. When points lie in the intersection of two C-space regions, then they can connect continuous paths through multiple regions. This is the essence of our transcription. What this implies is that for each visit to a vertex in the GCS, then we want to pick two points in a C-space region, so that the path lies in the region. So the convex sets in the GCS are not the C-space regions themselves, they are the set with $2n$ variables for an $n$-dimensional configuration space where the first $n$ and the last $n$ are both in the C-space region. In set notation, we'd say it's the Cartesian product of the set with itself. We form undirected edges between these sets iff the C-space regions intersect, and we put a constraint on each edge saying that the second point in the first set must be equal to the first point in the second set.

GcsTrajOpt in a simple 2D configuration space

Let's run GCS trajectory optimization on the simple configuration space depicted above. I've put the start in the bottom left corner and the goal in the top right. I'll run the algorithm twice, once with a minimum-distance objective, and once with a minimum-time objective (and velocity constraints):

We can generalize this significantly. Rather than just putting a line segment in each set, we can use the convex hull property of Bezier curves to guarantee that an entire continuous path of fixed degree lies inside the set. Using the derivative properties of the Bezier curves, we can add convex constraints on the velocities (e.g. velocity limits), and guarantee that the curves are smooth even at the intersections between regions. To define these velocities, though, we need to know something about the rate at which we are traversing the path, so we also introduce the time-scaling parameterization, much like we did in the kinematic trajectory optimization above, with time scaling parameters in each convex set.

The result is a fairly rich trajectory optimization framework that encodes many (but admittedly not all) of the costs and constraints that we want for trajectory optimization in the GCS convex optimization framework. For instance, we can write objectives that include:

• time (trajectory duration),
• path length (approximated with an upper bound),
• an upper bound on path "energy" (e.g. $\int |\dot{q}(t)|_2^2$),

• path derivative continuity (to arbitrary degree),
• velocity constraints (strictly enforced for all $t$, not just at sample points),
• additional convex position and/or path derivative constraints, such as initial and final positions and velocities.

GcsTrajOpt in a simple 2D configuration space (cont.)

This is the result of running nearly the same minimum-time optimization as in the example above, but this time we parameterize each segment with a fifth order Bezier curve and a time scaling, and ask for continuity up to the fourth derivative ("snap").

Inspection of the trajectory derivatives reveals that the velocities are smooth and that they stay within the velocity bounds $[-1, 1]$ for the entire trajectory; not just at sample points.

Nonconvex trajectory optimization can, of course, consume richer costs and constraints, but does not have the global optimization and completeness elements that GCS can provide. We are constantly adding more costs and constraints into the GCS Trajectory Optimization framework -- some seemingly nonconvex constraints actually have an exact convex reformulation, others have nice convex approximations. This leads to a natural merger of the two techniques: we can use convex approximations of more complex costs and constraints in the GCS convex relaxation, but then use nonlinear optimization during in the rounding step to handle the nonconvexity exactly once GCS selects the discrete path through the graph to give a strong initial guess Wrangel24. This can be accomplished in Drake via the use_in_transcription option of the AddCost and AddConstraint methods of GCS, and adding nonconvex constraints only to the Transcription::kRestriction.

Convex decomposition of (collision-free) configuration space

Ok, so how do we obtain a convex decomposition of collision-free configuration space? Let's start by taking the analogy with the PRM seriously. If we sample a collision-free point in the configuration space, then what is the right way to inflate that point into a convex region?

The answer is simpler when we are dealing with convex geometries. In my view, it is reasonable to approximate the geometry of a robot with convex shapes in the Euclidean space, $\Re^3.$ Even though many robots (like the iiwa) have nonconvex mesh geometries, we can often approximate them nicely as the union of simpler convex geometries; sometimes these are primitives like cylinders and spheres, alternatively we can perform a convex decomposition directly on the mesh. But even when the geometries are convex in the Euclidean space, it is generally unreasonable to treat them as convex in the configuration space.

There are a few exceptions. If your configuration space only involves translations, or if your robot geometry is invariant to rotations (e.g. a point robot or a spherical robot), then the configuration space will still be convex; this was one of the first key observations that cemented the notion of configuration space as a core mathematical object in robotics Lozano-Perez90. This can work for a mobile robot (even a quadrotor approximated by a sphere). In this case, one can easily compute the minimum Euclidean distance between any pairs of convex bodies at the sample point, and for instance, inflate the point into a sphere of the appropriate radius Shkolnik11a. In order to find bigger regions, we have made heavy use of the Iterative Region Inflation by Semidefinite Program (IRIS) algorithm introduced in Deits14 and implemented in Drake. This algorithm alternates between finding separating hyperplanes and finding the maximum volume contained ellipse to "inflate" the region to locally maximize (an efficient approximation of) the volume of the region.

It is possible to extend the IRIS algorithm to find large convex regions in the more general (nonconvex) configuration space. We've done it in two ways:

• Using nonlinear programming (NP). In the IRIS-NP algorithm Petersen23a (implemented in Drake), we replace the convex optimization for finding the closest collision with a nonlinear optimization solver like SNOPT. We randomly sample potential collisions in order to make the algorithm probabilistically sound; in practice this algorithm is relatively fast but does not guarantee that the region is completely collision free. We've recently dramatically improve the performance of this pipeline with IRIS-NP2, introduced in Werner24, with creates bigger regions with less faces orders of magnitude faster than the original IRIS-NP.
• Using zero-order optimization (ZO). By replacing the nonlinear optimization in IRIS-NP with a zeroth-order optimization (no gradients required; it uses only samples), IRIS-ZO produces slightly smaller regions, but can be trivially parallelized Werner24.
• Using algebraic kinematics + sums-of-squares optimization. Leveraging the idea that the kinematics of most of our robots can be expressed using (rational) polynomials, we can use tools from polynomial optimization to search for rigorous certificates of non-collision Dai23 (implemented in Drake). This method tends to be slower, but the results are sound. The one sacrifice that we have to make here is that we find convex regions in the stereographic projection coordinates of the original space. In practice, this represents a slight warping of the coordinate system (and consequently a trajectory length) over the region.

IRIS in Configuration Space

Let's run the IrisInConfigurationSpace (aka IRIS-NP) algorithm on the example of iiwa reaching into a shelf that we played with before. We will use the full 7 degrees of freedom of the arm, and the original collision geometry. I'd like for this example to demonstrate two things: (1) the mechanics of using IrisInConfigurationSpace and (2) the fact that these IRIS regions can cover a lot of volume, even when they are up against complicated configuration-space geometries.

IrisInConfigurationSpace can be configured using a single seed point; iterations of the algorithm will start with a small region around that point and "inflate" it to maximize a local approximation of configuration space volume. We can find a seed configuration, $q,$ that is reaching inside the top shelf using InverseKinematics. But if we only run IRIS from there, then the region it will find will not fill space inside the shelf, it will instead creep out of the shelf in order to cover the much larger configuration space outside of the shelf. To avoid this, we'll run IRIS twice; once with a seed that is the "home configuration" of the robot, which will grow a region outside the shelf. Then we'll run IRIS again with the IK seed inside the shelf, but with an additional constraint using the first region as an obstacle (asking that they do not overlap). This will force the second region to use the volume inside the shelf to maximize volume.

A different approach for obtaining this volume, which I now favor, is to collect many configuration space samples in the region of interest using inverse kinematics, then compute the MinimumVolumeCircumscribedEllipsoid to be used as the starting_ellipse for IRIS. Then we run only one iteration of IRIS to get a nice region with a notion of volume corresponding to covering this area of configuration space. This approach avoids the false constraint of using previous regions as obstacles (so leads to bigger regions), runs faster (due to using only one iteration of IRIS), and is more parallelizable.

If we were to continue to follow the analogy with the PRM, then we could imagine performing a convex decomposition of the space by sampling at random, rejecting samples that are either in collision or in an existing IRIS region, and then inflating any remaining samples. But in fact, we can do much better than this -- we can cover a larger percentage of the configuration free space with the same number of regions, or alternatively cover a similar volume with less regions. Our current best algorithm for performing this convex decomposition proceeds by computing a "minimum clique cover" of the "visibility graph" Werner23 (implemented in Drake).

GcsTrajOpt in a simple 2D configuration space (cont.)

If we run the "minimum clique cover" algorithm with a coverage threshold of 96% of the configuration space on the simple 2D example, we obtain automatic decompositions like this:

In practice, it seems reasonably tractable to try to compute a quite dense covering of the configuration space for 7-DOF manipulators, and perhaps even up to say 10 degrees of freedom. But for higher dimensions, I think that trying to cover every nook and cranny of the configuration space is a false goal. I strongly prefer the idea of intializing the clique-cover algorithm with sample points that represent the "important" regions of your configuration space. Sometimes we do this by manual seeding (via calls to inverse kinematics), or alternatively, we can do it by examining trajectories that come from teleop demonstrations, from another planner (that can generate plans, but without the same guarantees), or even from a reinforcement learning policy.

GcsTrajOpt Examples

Variations and Extensions

Fast path planning (FPP) Marcucci23. Many motion planning instances don't need the full power of GCS; heuristic approximations can often find very good solutions. Rather than solving the full GCS problem, considering the discrete and continuous variables simultaneously, FPP alternates between solving the discrete and continuous problems to find a local minima. FPP is (typically) faster than GcsTrajectoryOptimization, and by virtue of using alternations, it can handle constraints on the derivatives (e.g. acceleration constraints) which GcsTrajectoryOptimization cannot handle. However, FPP is more limited on the class of convex sets/constraints that it can support. (Drake implementation coming soon!)

GCS on a manifold including for mobile manipulation Cohn23, and for bimanual manipulation Cohn23a.

GCS with dynamic constraints. For discrete-time systems described by piecewise-affine dynamics and convex objectives and constraints, GCS provides the strongest formulations to dateMarcucci21. It is also possible to reason more directly about continuous dynamic constraints over continuous curves (more coming soon!).

Planning through contact. Amazingly, we now have tight convex relaxations for quasi-static manipulation with contact Graesdal24. The initial examples are quite simple (so that we could study the problem carefully, and defer the details of scaling the solver), but you can expect much more from us in this direction very soon.

Planning under uncertainty (coming soon!).

Task and motion planning (coming soon!). Kurtz23 showed a natural transcription from temporal logic to GCS.

GCS as a (feedback) policy (coming soon!).