On the Design of Region-Avoiding Metrics for Collision-Safe Motion Generation on Riemannian Manifolds

Holger Klein, Noémie Jaquier, André Meixner, Tamim Asfour

TL;DR

This paper introduces a novel way to generate robot motions by reshaping the "geometry" of a robot's configuration space. By incorporating inverse barrier functions into Riemannian metrics, the authors ensure that the shortest, most energy-efficient paths (geodesics) naturally and strictly avoid joint limits, self-collisions, and obstacles.

What Problem Does This Paper Address?

In robotics, generating motion that is both energy-efficient and safe (avoiding collisions and joint limits) is a classic challenge. One elegant approach is to treat the robot's configuration space as a Riemannian manifold, where the "distance" between points is defined by the robot's dynamics (the kinetic energy metric). In this framework, the ideal motion is a geodesic—the shortest path on this curved surface.

However, standard kinetic energy metrics don't "know" about obstacles or joint limits. A geodesic might be energy-optimal but could easily smash the robot into a table or exceed its joint range. Previous attempts to fix this often resulted in "soft" constraints (the robot might still collide if moving fast enough) or produced weird, jerky accelerations that didn't point directly away from danger.

Methods

The authors propose a class of region-avoiding metrics that reshape the manifold around "no-go" zones.

  1. Barrier Functions: They use barrier functions (mathematical functions that go to infinity at a boundary) to modify the metric. They demonstrate that inverse barriers ($1/dist$) are superior to exponential barriers because they provide "strict" avoidance—the energy required to cross the boundary becomes infinite, making it impossible for a geodesic to pass through.
  2. Direction-Dependent Basis Change: A key innovation is how they handle multi-dimensional space. Simply adding a barrier can cause irregular accelerations. The authors implement a change of basis (using the Gram-Schmidt process) so that the repelling "force" generated by the metric always points directly away from the obstacle, regardless of the robot's current velocity direction.
  3. Directional Pullbacks: While joint limits are easy to define in joint space, obstacles are defined in 3D task space. The authors propose pulling back the direction to the obstacle from task space to joint space using the Jacobian pseudo-inverse. This ensures the resulting joint-space metric produces the correct repelling acceleration in the physical world.
  4. Optimization-based Geodesics: To find the path between two points, they represent the trajectory as a cubic spline and optimize the control points to minimize the total Riemannian energy.

Datasets

The researchers evaluated their approach using:

  • KIT Whole-Body Human Motion Database: They used five human movement sequences (reaching, throwing, pointing, and two types of waving) and "retargeted" them to the ARMAR-6 humanoid robot.
  • ARMAR-6 Humanoid Model: A complex 8-degree-of-freedom arm was used for testing joint limits, self-collision (arms crossing), and environment obstacles (moving a hand from under a table to above it).

Key Insights and Ideas

  • Hard vs. Soft Constraints: Exponential barriers (common in previous work) only provide soft avoidance. If a robot has enough momentum, it can "tunnel" through the barrier. Inverse barriers create a "hard" geometric wall that geodesics cannot cross.
  • Geometry as Control: By changing the metric, the robot's "natural" behavior becomes safe. You don't need a separate collision-avoidance controller; the physics of the manifold itself pushes the robot away from danger.
  • The Importance of Gradient Direction: The paper highlights that the geodesic acceleration depends not just on the metric's value, but on its relationship with its gradient. Their basis-change method ensures this relationship is physically intuitive (pushing straight back from the obstacle).

Key Results

The authors compared their inverse-barrier metric against a standard kinetic energy metric and an exponential-barrier metric:

  • Safety: In "arms crossing" and "waving" tasks, the standard metric and exponential barrier resulted in significant joint limit violations (up to 100% in some cases) and self-collisions. The inverse-barrier metric achieved 0% violations across all tests.
  • Efficiency: The generated paths remained close to the energy-optimal geodesics of the original kinetic energy metric, only deviating when necessary to stay safe.
  • Obstacle Avoidance: In the "table" scenario, the inverse-barrier metric successfully guided the hand around the table, whereas other methods either collided or got stuck in local minima.

Limitations and Future Work

  • Velocity Independence: Currently, the metric only depends on the robot's position. It doesn't distinguish between a robot moving toward an obstacle and one moving away from it. The authors suggest that incorporating velocity into the metric (creating a Finsler metric) is a promising next step.
  • Computational Complexity: While they achieved computation times under 0.5s for an 8-DoF arm, scaling this to full-body humanoid motion (20+ DoF) in real-time remains a challenge.
  • Local Minima: Although the spline optimization helps, complex environments with many "traps" might still pose challenges for finding the global energy-minimum geodesic.

Summary updated

Abstract

The generation of energy-efficient and dynamic-aware robot motions that satisfy constraints such as joint limits, self-collisions, and collisions with the environment remains a challenge. In this context, Riemannian geometry offers promising solutions by identifying robot motions with geodesics on the so-called configuration space manifold. While this manifold naturally considers the intrinsic robot dynamics, constraints such as joint limits, self-collisions, and collisions with the environment remain overlooked. In this paper, we propose a modification of the Riemannian metric of the configuration space manifold allowing for the generation of robot motions as geodesics that efficiently avoid given regions. We introduce a class of Riemannian metrics based on barrier functions that guarantee strict region avoidance by systematically generating accelerations away from no-go regions in joint and task space. We evaluate the proposed Riemannian metric to generate energy-efficient, dynamic-aware, and collision-free motions of a humanoid robot as geodesics and sequences thereof.
BibTeX
@inproceedings{klein2023design,
  author = {Holger Klein and Noémie Jaquier and André Meixner and Tamim Asfour},
  title = {{On the Design of Region-Avoiding Metrics for Collision-Safe Motion Generation on Riemannian Manifolds}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=WXu3XFwRKw},
}

Loading related content...

Figure