RMPs for Safe Impedance Control in Contact-Rich Manipulation

Seiji Shaw, Ben Abbatematteo, George Konidaris

Abstract

Contact-rich manipulation remains a significant challenge in robotics, particularly when balancing task performance with safety constraints like joint-limit avoidance and collision prevention. While Variable Impedance Control in End-Effector Space (VICES) allows for compliant interaction, it often neglects the safety of the robot's full body. This paper presents RMP-VICES, a framework that integrates Riemannian Motion Policies (RMPs) with VICES. By embedding safety constraints as geometric priors within an RMP-Tree, the system ensures that reinforcement learning agents operate within a safe action space. Experimental results across three manipulation domains demonstrate a reduction in joint-limit violations by over 90% and improved generalization to novel obstacles.

Introduction

In robotics, "contact-rich manipulation"—tasks where the robot must touch and move objects—is notoriously difficult. A popular method called VICES (Variable Impedance Control in End-Effector Space) helps by allowing a reinforcement learning (RL) agent to control the compliance of the robot's hand. This is effective for interacting with objects, but VICES has a major blind spot: it only considers the end-effector. Because the RL agent only "sees" the hand's position, it often ignores the configuration of the rest of the arm, leading to frequent joint-limit violations or elbow collisions during the learning process. Usually, researchers attempt to mitigate this by penalizing the robot with negative rewards, but this approach is often ineffective for ensuring consistent safety or adapting to new environments.

Methodology: The RMP-VICES Framework

The authors introduce RMP-VICES, which integrates safety constraints directly into the robot's control architecture before learning begins. This is achieved through two primary components:

  1. Riemannian Motion Policies (RMPs): A mathematical framework that treats different robot behaviors (such as reaching for a goal or avoiding a wall) as individual forces on a curved geometric surface.
  2. The RMP-Tree: These behaviors are organized in a tree structure where the "root" represents the robot's joints and the "leaves" represent specific tasks.
The RMP-Tree structure hierarchically organizes joint-limit avoidance, collision policies, and goal attractors.

By synthesizing these policies, the RMP framework mathematically "blends" competing objectives. If an RL agent attempts to move the hand in a way that would cause a collision, the RMP layer automatically nudges the arm into a safer configuration without the RL agent needing explicit awareness of the obstacle.

Experimental Evaluation

The researchers evaluated RMP-VICES using the Robosuite simulation framework across three distinct contact-rich domains: Trajectory-Following, Door-Opening, and Block-Pushing. To test generalization, they utilized versions of these environments with randomly generated obstacles (represented as red spheres).

The three evaluation domains: door-opening, trajectory-following, and block-pushing with obstacles.

Key Results and Insights

The quantitative improvements over standard VICES were substantial:

  • Safety Improvements: Joint-limit violations were reduced by up to 90.7% in door-opening and 93.2% in block-pushing. Collisions were reduced by approximately 50% across all domains.
  • Accelerated Learning: Treating safety as a geometric prior actually helped the robot learn faster in some tasks. Because the robot violated joint limits less often, training episodes were not terminated prematurely, providing the agent with more high-quality data.
  • Robustness: When placed in environments with unseen obstacles, RMP-VICES agents successfully navigated around them, whereas standard VICES agents moved blindly and crashed.
  • Collision Severity: Even when collisions occurred, the force of impact was significantly lower due to the damping policies inherent in the RMP framework.

Conclusion and Future Work

RMP-VICES demonstrates that embedding safety as a geometric necessity rather than a learned penalty significantly enhances the reliability of contact-rich manipulation. While the system requires intensive tuning to balance policy weights and currently lacks formal stability proofs, it represents a major step toward "body-aware" robotic control. Future research will focus on learning the Riemannian metrics automatically to allow robots to dynamically prioritize behaviors based on their workspace context.

Summary updated

BibTeX
@inproceedings{shaw2022rmps,
  author = {Seiji Shaw and Ben Abbatematteo and George Konidaris},
  title = {{RMPs for Safe Impedance Control in Contact-Rich Manipulation}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2022},
  url = {https://openreview.net/forum?id=Mq7ZOxS4_X},
}
Figure