RMPflow: A Computational Graph for Automatic Motion Policy Generation

Ching-An Cheng, Mustafa Mukadam, Jan Issac, Stan Birchfield, Dieter Fox, Byron Boots, Nathan D. Ratliff

Introduction

Robot motion generation typically faces a trade-off between computationally intensive motion planning and fast but often limited reactive control. While planners can navigate complex, non-Euclidean spaces, they often struggle with real-time adaptation to dynamic environments. Conversely, reactive controllers like Operational Space Control (OSC) are fast but typically assume simple Euclidean geometry, relying on extrinsic potential fields that can lead to jittery movements or local minima. RMPflow bridges this gap by providing a globally stable, highly reactive framework that intrinsically understands non-Euclidean task geometries.

The RMPflow Framework

At the core of RMPflow is the RMP-tree, a computational data structure where the root represents the robot's configuration space (e.g., joint angles) and the leaves represent specific task spaces, such as end-effector positions or distances to obstacles.

Evolution of RMP structures from simple mappings (a) to multi-task mappings (b) and complex RMP-trees (c).

To manage these tasks efficiently, the authors define an RMP-algebra consisting of three primary operators:

  1. Pushforward: Propagates the robot's state (position and velocity) from the root down to the leaves.
  2. Pullback: Propagates desired forces and inertias from local task policies back up to the root.
  3. Resolve: Combines these pulled-back policies at the root into a single global acceleration command for the robot's joints.

Geometric Dynamical Systems (GDSs)

RMPflow extends classical mechanics through Geometric Dynamical Systems (GDSs), which allow for velocity-dependent metrics. Unlike standard physics where inertia depends only on position or mass, GDSs allow the "importance" or weight of a task to change based on the robot's direction of motion. For instance, an obstacle's repulsive policy can be weighted heavily when the robot is moving toward it and ignored when moving away, preventing unnecessary deceleration.

Furthermore, these non-constant metrics naturally generate "curvature" terms (similar to Coriolis forces). Including these terms ensures smooth navigation around obstacles, effectively coaxing the robot to orbit or curve around them rather than just bouncing off.

Key Insights and Stability

RMPflow is intrinsically coordinate-free, meaning task policies designed for one robot (like a 7-DOF arm) can be seamlessly transferred to a different platform while maintaining the same task-space behavior. Crucially, the authors provide mathematical proof that the resulting global policy remains Lyapunov-stable despite the highly nonlinear transformations occurring within the RMP-tree.

Experimental Results

The framework was evaluated in both simulation and real-world scenarios:

  • Simulation: In a 7-DOF YuMi robot navigating cluttered environments, RMPflow achieved a 0% collision failure rate, significantly outperforming standard potential field baselines which were often too conservative or failed to reach targets.
  • Real-world Integration: Deployed on Baxter and YuMi platforms with real-time vision tracking, the system handled dynamic tasks like reacting to human perturbations and opening drawers at interactive speeds.

Conclusion

RMPflow offers a robust solution for generating smooth, reactive motion policies by combining simple local tasks into a global controller. While it currently requires manual design of subtask RMPs, future work aims to integrate machine learning to automatically learn or fine-tune these policies, further reducing the engineering burden.

Summary updated

Abstract

We develop a novel policy synthesis algorithm, RMPflow, based on geometrically consistent transformations of Riemannian Motion Policies (RMPs). RMPs are a class of reactive motion policies designed to parameterize non-Euclidean behaviors as dynamical systems in intrinsically nonlinear task spaces. Given a set of RMPs designed for individual tasks, RMPflow can consistently combine these local policies to generate an expressive global policy, while simultaneously exploiting sparse structure for computational efficiency. We study the geometric properties of RMPflow and provide sufficient conditions for stability. Finally, we experimentally demonstrate that accounting for the geometry of task policies can simplify classically difficult problems, such as planning through clutter on high-DOF manipulation systems.
BibTeX
@inproceedings{cheng2018rmpflow,
  author = {Ching-An Cheng and Mustafa Mukadam and Jan Issac and Stan Birchfield and Dieter Fox and Byron Boots and Nathan D. Ratliff},
  title = {{RMPflow: A Computational Graph for Automatic Motion Policy Generation}},
  booktitle = {WAFR},
  year = {2018},
  url = {https://openreview.net/forum?id=pb1v4oOiQzO},
}
Figure