Prabin Kumar Rath, Nakul Gopalan

Arizona State University

Official Webpage: https://prabinrath.github.io/xmop/      |     Code: https://github.com/prabinrath/xmop

Abstract

Classical manipulator motion planners work across different robot embodiments. However they plan on a pre-specified static environment representation, and are not scalable to unseen dynamic environments. Neural Motion Planners (NMPs) are an appealing alternative to conventional planners as they incorporate different environmental constraints to learn motion policies directly from raw sensor observations. Contemporary state-of-the-art NMPs can successfully plan across different environments.  However none of the existing NMPs generalize across robot embodiments. In this paper we propose Cross-Embodiment Motion Policy (XMoP), a neural policy for learning to plan over a distribution of manipulators. XMoP implicitly learns to satisfy kinematic constraints for a distribution of robots and zero-shot transfers the planning behavior to unseen robotic manipulators within this distribution. We achieve this generalization by formulating a whole-body control policy that is trained on planning demonstrations from over three million procedurally sampled robotic manipulators in different simulated environments. Despite being completely trained on synthetic embodiments and environments, our policy exhibits strong sim-to-real generalization across manipulators with different kinematic variations and degrees of freedom with the same set of frozen policy parameters. We evaluate XMoP on 7 commercially available manipulators and show successful cross-embodiment motion planning, achieving an average 70% success rate on baseline benchmarks. Furthermore, we show sim-to-real demonstrations on two unseen manipulators solving novel planning problems across eight unstructured real-world environments even with dynamic obstacles.

Cross Embodiment Motion Policy (XMoP) is a Behavior Cloning (BC) policy trained on synthetic planning demonstrations that zero-shot transfers to unseen robotic manipulators and can plan in unstructured real-world environments. 

All rollouts shown in the videos (both simulated and real) use XMoP with a fixed set of frozen policy parameters.

XMoP uses Model Predictive Control (MPC) which allows for locally reactive planning in the presence of dynamic obstacles.

XMoP generates smooth and optimal trajectories for reaching SE(3) targets in obstacle-free environments. Zero-shot control of 7 commercial robot embodiments.

Real-World Experiments on 3 Domains

Unstructured Obstacle Domain: We test XMoP’s ability to plan in unstructured and cluttered environments which are hard for conventional planning algorithms that use primitive-based obstacle representations.

Wall Hopping Domain: We test XMoP's ability to plan in a structured environment with large obstacles that significantly occupy the space in front of the robot.

Bin-to-Bin Domain: We test XMoP’s ability to plan for a common industrial bin-to-bin motion task, where the manipulator needs to move its end-effector from inside one bin to another.

Failure Modes of XMoP