Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Motion Planning Networks

Ahmed H. Qureshi, Anthony Simeonov, Mayur J. Bency, Michael C. Yip

Seminal June 2018

TL;DR MPNet learns motion planning from RRT* demonstrations. One network encodes the obstacle point cloud and another predicts the next configuration. In 2D, 3D and rigid-body tests it finds paths in 0.11-0.48 s on average, 18-36x faster than BIT*, including in workspaces it never saw.

What they do

Sampling-based planners such as RRT* slow down sharply as the planning problem’s dimension grows. MPNet replaces most of the search with two networks trained by imitation on RRT* paths. One encodes the obstacle point cloud and the other predicts the robot’s next configuration. The test robots are 2D and 3D point masses, a rigid body and a 7-DoF Baxter arm.

How it works

  • Encoder (Enet). Embeds the obstacle point cloud into a latent vector. It is a contractive autoencoder in the 2D/3D cases and is trained end-to-end with the planner for Baxter.
  • Planning network (Pnet). A feed-forward net maps the obstacle embedding, current state and goal state to the next state, trained with an MSE loss against expert paths. Dropout (p = 0.5) stays on during planning, so repeated calls produce different paths.
  • Bidirectional generation. Paths grow from the start and the goal in alternation until a straight collision-free segment joins them. Lazy states contraction then drops waypoints that can be skipped.
  • Replanning. Colliding segments are repaired by calling the neural planner recursively for a fixed number of steps (NR). The hybrid version (HR) passes any remaining gaps to RRT, so the method inherits RRT’s probabilistic completeness.

Results

  • Across simple 2D, complex 2D, complex 3D and rigid-body tests, MPNet (NR) averages 0.11-0.48 s, and HR averages 0.19-1.98 s. BIT* takes 2.63-11.91 s and Informed-RRT* 5.36-30.38 s.
  • Times stay about the same in the 10 unseen test workspaces as in the training workspaces (e.g. rigid body: 0.33 s vs 0.32 s).
  • Mean accuracy is 97% with neural replanning alone and 100% with the hybrid.
  • On Baxter, MPNet takes about 1 s with 85% success. BIT* takes about 9 s and succeeds 56% of the time at finding paths within 40% of MPNet’s path length.

Limitations

The networks alone carry no guarantees; completeness comes only from the RRT* fallback. The method targets feasible paths, not optimal ones. Each environment type needs 4000 RRT* demonstrations in each of 100 training workspaces. Outside the Baxter experiments, the baselines were Python implementations. Kinodynamic planning and changing environments are left for future work.

Figures

From Figure 1: Baxter's 7-DoF arm executes an MPNet path, frames 1-4, with a stopwatch showing execution time. Here M...
From Figure 1: Baxter's 7-DoF arm executes an MPNet path, frames 1-4, with a stopwatch showing execution time. Here MPNet took under 1 second; BIT* took 3.1 minutes on average to find a path of comparable cost.
From Figure 2: The online neural planner. The planning network takes the obstacle encoding, the current state and the...
From Figure 2: The online neural planner. The planning network takes the obstacle encoding, the current state and the goal, and its predicted next state is fed back as input. Blue blocks are frozen, already-trained modules.
From Figure 3: Paths from MPNet (red) and RRT* (blue) in a complex 3D environment. RRT* took 6.9 s and MPNet 0.50 s.
From Figure 3: Paths from MPNet (red) and RRT* (blue) in a complex 3D environment. RRT* took 6.9 s and MPNet 0.50 s.
Abstract

Fast and efficient motion planning algorithms are crucial for many state-of-the-art robotics applications such as self-driving cars. Existing motion planning methods become ineffective as their computational complexity increases exponentially with the dimensionality of the motion planning problem. To address this issue, we present Motion Planning Networks (MPNet), a neural network-based novel planning algorithm. The proposed method encodes the given workspaces directly from a point cloud measurement and generates the end-to-end collision-free paths for the given start and goal configurations. We evaluate MPNet on various 2D and 3D environments including the planning of a 7 DOF Baxter robot manipulator. The results show that MPNet is not only consistently computationally efficient in all environments but also generalizes to completely unseen environments. The results also show that the computation time of MPNet consistently remains less than 1 second in all presented experiments, which is significantly lower than existing state-of-the-art motion planning algorithms.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.