Underactuated Robotics

Algorithms for Walking, Running, Swimming, Flying, and Manipulation

Russ Tedrake

© Russ Tedrake, 2020
How to cite these notes   |   Send me your feedback

Note: These are working notes used for a course being taught at MIT. They will be updated throughout the Spring 2020 semester. Lecture videos are available on YouTube.

Previous Chapter Table of contents Next Chapter

Motion Planning as Search

The term "motion planning" is a hopelessly general term which almost certainly encompasses the dynamic programming, feedback design, and trajectory optimization algorithms that we have already discussed. However, there are a number of algorithms and ideas that we have not yet discussed which have grown from the idea of formulating motion planning as a search problem -- for instance searching for a path from a start to a goal in a graph which is too large solve completely with dynamic programming. Some, but certainly not all, of these algorithms sacrifice optimality in order to find any path if it exists, and the notion of a planner being "complete" -- guaranteed to find a path if one exists -- is highly valued. This is precisely our goal for this chapter, to add some additional tools that will be able to provide some form of solutions for our most geometrically complex, highly non-convex, robot control problems.

LaValle06 is a very nice book on planning algorithms in general and on motion planning algorithms in particular. Compared to other planning problems, motion planning typically refers to problems where the planning domain is continuous (e.g. continuous state space, continuous action space), but many motion planning algorithms trace their origins back to ideas in discrete domains (e.g., graph search).

For this chapter, we will consider the following problem formulation: given a system defined by the nonlinear dynamics (in continuous- or discrete-time) $$\dot{\bx} = f(\bx,\bu) \quad \text{or} \quad \bx[n+1] = f(\bx[n],\bu[n]),$$ and given a start state $\bx(0) = \bx_0$ and a goal region ${\cal G}$, find any finite-time trajectory from $\bx_0$ to to $\bx \in {\cal G}$ if such a trajectory exists.

Artificial Intelligence as Search

A long history... some people feel that the route to creating intelligent machines is to collect large ontologies of knowledge, and then perform very efficient search. (The more modern view of AI is focused instead on machine learning, the right answer probably involves pieces of both.) Samuels checker players, Deep Blue playing chess, theorem proving, Cyc, IBM Watson,...

One of the key ideas is the use of "heuristics" to guide the search. "Is it possible to find and optimal path from the start to a goal without visiting every node?". $A^*$. Admissible heuristics. Example: google maps.

Online planning. $D^*$, $D^*$-Lite.

Randomized motion planning

If you remember how we introduced dynamic programming initially as a graph search, you'll remember that there were some challenges in discretizing the state space. Let's assume that we have discretized the continuous space into some finite set of discrete nodes in our graph. Even if we are willing to discretize the action space for the robot (this might be even be acceptable in practice), we had a problem where discrete actions from one node in the graph, integrated over some finite interval $h$, are extremely unlikely to land exactly on top of another node in the graph. To combat this, we had to start working on methods for interpolating the value function estimate between nodes.

add figure illustrating the interpolation here

Interpolation can work well if you are trying to solve for the cost-to-go function over the entire state space, but it's less compatible with search methods which are trying to find just a single path through the space. If I start in node 1, and land between node 2 and node 3, then which node to I continue to expand from?

One approach to avoiding this problem is to build a search tree as the search executes, instead of relying on a predefined mesh discretization. This tree will contains nodes rooted in the continuous space at exactly the points where system can reach.

Another other problem with any fixed mesh discretization of a continuous space, or even a fixed discretization of the action space, is that unless we have specific geometric / dynamic insights into our continuous system, it very difficult to provide a complete planning algorithm. Even if we can show that no path to the goal exists on the tree/graph, how can we be certain that there is no path for the continuous system? Perhaps a solution would have emerged if we had discretized the system differently, or more finely?

One approach to addressing this second challenge is to toss out the notion of fixed discretizations, and replace them with random sampling (another approach would be to adaptively add resolution to the discretization as the algorithm runs). Random sampling, e.g. of the action space, can yield algorithms that are probabilistically complete for the continuous space -- if a solution to the problem exists, then a probabilistically complete algorithm will find that solution with probability 1 as the number of samples goes to infinity.

With these motivations in mind, we can build what is perhaps the simplest probabilistically complete algorithm for finding a path from the a starting state to some goal region with in a continuous state and action space:

Planning with a Random Tree

Let us denote the data structure which contains the tree as ${\cal T}$. The algorithm is very simple:

  • Initialize the tree with the start state: ${\cal T} \leftarrow \bx_0$.
  • On each iteration:
    • Select a random node, $\bx_{rand}$, from the tree, ${\cal T}$
    • Select a random action, $\bu_{rand}$, from a distribution over feasible actions.
    • Compute the dynamics: $\bx_{new} = f(\bx_{rand},\bu_{rand})$
    • If $\bx_{new} \in {\cal G}$, then terminate. Solution found!
    • Otherwise add the new node to the tree, ${\cal T} \leftarrow \bx_{new}$.
It can be shown that this algorithm is, in fact, probabilistically complete. However, without strong heuristics to guide the selection of the nodes scheduled for expansion, it can be extremely inefficient. For a simple example, consider the system $\bx[n] = \bu[n]$ with $\bx \in \Re^2$ and $\bu_i \in [-1,1]$. We'll start at the origin and put the goal region as $\forall i, 15 \le x_i \le 20$. Try it yourself:

T = struct('parent',zeros(1,1000),'node',zeros(2,1000));  % pre-allocate memory for the "tree"
for i=2:size(T.parent,2)
T.parent(i) = randi(i-1);
x_rand = T.node(:,T.parent(i));
u_rand = 2*rand(2,1)-1;
x_new = x_rand+u_rand;
if (15<=x_new(1) && x_new(1)<=20 && 15<=x_new(2) && x_new(2)<=20)
  disp('Success!'); break;
T.node(:,i) = x_new;
Again, this algorithm is probabilistically complete. But after expanding 1000 nodes, the tree is basically a mess of node points all right on top of each other:

We're nowhere close to the goal yet, and it's not exactly a hard problem.

While the idea of generating a tree of feasible points has clear advantages, we have lost the ability to cross off a node (and therefore a region of space) once it has been explored. It seems that, to make randomized algorithms effective, we are going to at the very least need some form of heuristic for encouraging the nodes to spread out and explore the space.

Rapidly-Exploring Random Trees (RRTs)

(Click here to watch the animation)
(Click here to watch the animation)

RRTs for robots with dynamics

Variations and extensions

Multi-query planning with PRMs, ...

RRT*, RRT-sharp, RRTx, ...

Kinodynamic-RRT*, LQR-RRT(*)

Complexity bounds and dispersion limits


Not sure yet whether randomness is fundamental here, or whether is a temporary "crutch" until we understand geometric and dynamic planning better.

Decomposition methods

Cell decomposition...

Mixed-integer planning.

Approximate decompositions for complex environments (e.g. IRIS)


RRT Planning

In this notebook we will write code for the Rapidly-Exploring Random Tree (RRT). Building on this implementation we will also implement RRT*, a variant of RRT that converges towards an optimal solution.

  1. Implement RRT
  2. Implement RRT*
Previous Chapter Table of contents Next Chapter