Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Motion Policy Networks

Adam Fishman, Adithyavairan Murali, Clemens Eppner, Bryan Peele, Byron Boots, Dieter Fox

Seminal October 2022

TL;DR MπNets is a neural policy that drives a 7-DoF arm to a task-space target while avoiding obstacles, using only a segmented point cloud. It is trained on 3.27 million expert trajectories and succeeds on 95.33% of test problems, against 65.28% for the earlier neural planner MPNets; each solution takes 0.33 s, versus 16.46 s for a global planner.

What they do

MπNets replaces a motion planning pipeline with a single learned policy for a Franka Panda arm. From the current joint configuration and a point cloud of the scene, it predicts the next joint configuration toward an end-effector target. It needs no collision checker or scene model at test time. It is trained by behavior cloning on 3.27 million planner demonstrations across 576,532 procedurally generated tabletop, cubby and dresser scenes.

How it works

  • Inputs. The target is a gripper mesh at the goal pose, sampled into the point cloud. The cloud is segmented into robot, obstacle and target points. A PointNet++ encoder and an MLP configuration encoder feed a decoder that outputs a joint-space displacement (19M parameters).
  • Loss. L1 + L2 distance between 1,024 points on the predicted and expert robot surfaces, plus a hinge collision loss computed from the obstacles’ signed distance fields.
  • Expert. A hybrid planner: task-space AIT* gives waypoints that Geometric Fabrics follow. Hindsight Goal Revision relabels each problem’s goal as wherever the expert actually ended.
  • Robustness. Gaussian noise on input joints, and a fresh point cloud sampled for every example.

Results

  • On hybrid-solvable problems: 95.33% success vs 65.28% for MPNets. On problems both experts solve: 95.06%, vs 76.00% for STORM and 60.06% for Geometric Fabrics (both given an oracle collision checker).
  • With a single-view partial point cloud, success drops only from 95.06% to 93.22%, with 3.06% collisions; AIT* on a voxel map collides on 16.41%.
  • Ablations: without Hindsight Goal Revision, success is 58.11%; without robot points in the cloud, 65.06%.
  • Against a moving block, success is 88.1%, 57.4% and 28.3% at slow, medium and fast speeds. Real-robot demos with an L515 camera are qualitative only.

Limitations

Performance is bounded by the expert. On harder global-planner-solvable problems, success falls to 75.78% and collisions reach 11%. The policy collides on 64.25% of unsolvable problems, keeps no memory of occluded obstacles, and sometimes ends 180° off about the gripper axis.

Videos

Full video (YouTube)

Figures

From Figure 1: MπNets are trained on a large dataset of synthetic demonstrations (left) and solve motion planning pro...
From Figure 1: MπNets are trained on a large dataset of synthetic demonstrations (left) and solve motion planning problems from raw point cloud observations (right).
From Figure 2: The input is a normalized robot configuration plus a point cloud segmented into robot, obstacles and t...
From Figure 2: The input is a normalized robot configuration plus a point cloud segmented into robot, obstacles and target. The policy outputs a displacement in normalized joint space, which gives the next configuration.
From Figure 4: MπNets success rate keeps rising with more training data, while MPNets stays roughly constant.
From Figure 4: MπNets success rate keeps rising with more training data, while MPNets stays roughly constant.
Abstract

Collision-free motion generation in unknown environments is a core building block for robot manipulation. Generating such motions is challenging due to multiple objectives; not only should the solutions be optimal, the motion generator itself must be fast enough for real-time performance and reliable enough for practical deployment. A wide variety of methods have been proposed ranging from local controllers to global planners, often being combined to offset their shortcomings. We present an end-to-end neural model called Motion Policy Networks (M$π$Nets) to generate collision-free, smooth motion from just a single depth camera observation. M$π$Nets are trained on over 3 million motion planning problems in over 500,000 environments. Our experiments show that M$π$Nets are significantly faster than global planners while exhibiting the reactivity needed to deal with dynamic scenes. They are 46% better than prior neural planners and more robust than local control policies. Despite being only trained in simulation, M$π$Nets transfer well to the real robot with noisy partial point clouds. Code and data are publicly available at https://mpinets.github.io.

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.