Record STMT-W611DD1Aspirational

RMP2: A Structured Composable Policy Class for Robot Learning

White paper State lawGlobalin force

This paper introduces RMP2, a more efficient and flexible alternative algorithm to the RMPflow message passing routine for computing Riemannian Motion Policies using modern automatic differentiation tools. The authors demonstrate that RMP2 retains the analytical guarantees of RMPflow while offering an easier programming interface, support for general directed acyclic graphs, and end-to-end differentiability for robot learning. Experimental results in reinforcement learning tasks with a three-link robot and a Franka robot arm show that RMP2-based structured policies improve performance and safety.

Summary and content fields are drawn from the document text and are pending editorial review.

Who issued it

University of WashingtonAcademic

Lead issuer

Microsoft ResearchIndustry

Co-issuer

Georgia Institute of TechnologyAcademic · US

Co-issuer

NvidiaIndustry · US

Co-issuer

Legal forceAspirational

Under whose authorityState law

Judged fromfull text

Who it governs

Researchers and academicsAI developers and providersAI deployers and users

Robotics researchers and AI developers working on motion planning, reinforcement learning, and structured policy optimization for robotic systems.

Sectors

Research and science · robot learning and motion controlManufacturing and industry · robotics systems

What the document contains

What's covered

The paper covers the algorithmic design, complexity analysis, and learning integration of RMP2, a structured policy class implementation for acceleration-based robot motion control. It evaluates the approach through simulated reinforcement learning experiments involving obstacle avoidance and goal reaching for multi-link robotic arms.

What's missing

The document does not provide legal compliance frameworks, regulatory enforcement mechanisms, or policy guidelines for commercial deployment outside of simulated research robotics environments.

Key takeaways

RMP2 replaces the complex message passing algorithm in RMPflow with standard automatic differentiation operators.
RMP2 supports general directed acyclic graphs (DAGs) rather than being limited to tree-structured task maps.
RMP2 achieves the same time complexity as RMPflow while maintaining a smaller memory footprint.
The method enables fully differentiable structured policies, facilitating end-to-end learning via reinforcement learning and supervised learning.
Experiments demonstrate that incorporating RMP2 policies as priors or residuals improves both task performance and safety during robotic reaching tasks.

Outline

· Introduction
· Background
  - Acceleration-based Motion Control
  - RMPflow: A Framework for Multi-Task Problems
· RMP2 based on Automatic Differentiation
  - Key Idea
  - RMP2 based on Reverse Accumulation
  - Complexity of RMP2
  - Discussion: A Naïve Alternative Algorithm
  - Key Benefits of RMP2
· RMP2 for Learning
  - Parameterizing RMP2 Policies
  - Learning Setups
  - Learning with Residuals
· Experiments
  - Three-Link Robot Reaching
  - Franka Robot Reaching

Issued10 March 2021

Length11,043 words · 14 pages

FormatNot recorded

3/5

Openness about how AI systems work and are used

RMP2 allows for a simper user interface. The user only needs to specify the task map using an automatic differentiation library

Section I

screening score · 2026-10-24

Enforcement

0/5

Consequences and compliance mechanisms with real force

Abstract

screening score · 2026-10-24

Topics

robotics reinforcement_learning motion_planning automatic_differentiation policy_optimization