CMP: Robust Whole-Body Tracking for Loco-Manipulation via Competence Manifold Projection

2026-04-08

AI Summary

This work addresses the vulnerability of whole-body control in legged mobile manipulators to out-of-distribution (OOD) inputs caused by sensor noise or infeasible commands. To mitigate this, the authors propose Capability Manifold Projection (CMP), a method that reformulates infinite-horizon safety constraints into a single-step manifold inclusion problem via a frame-level safety mechanism. CMP constructs an isomorphic latent space that aligns manifold geometry with safety probability, enabling real-time OOD detection and correction with O(1) computational complexity. The approach ensures safety while preserving task continuity, exhibiting a “best-effort” generalization behavior. Experimental results demonstrate that, under representative OOD conditions, CMP improves system survival rates by up to tenfold compared to baseline methods, with less than a 10% degradation in task-tracking performance.

Abstract

While decoupled control schemes for legged mobile manipulators have shown robustness, learning holistic whole-body control policies for tracking global end-effector poses remains fragile against Out-of-Distribution (OOD) inputs induced by sensor noise or infeasible user commands. To improve robustness against these perturbations without sacrificing task performance and continuity, we propose Competence Manifold Projection (CMP). Specifically, we utilize a Frame-Wise Safety Scheme that transforms the infinite-horizon safety constraint into a computationally efficient single-step manifold inclusion. To instantiate this competence manifold, we employ a Lower-Bounded Safety Estimator that distinguishes unmastered intentions from the training distribution. We then introduce an Isomorphic Latent Space (ILS) that aligns manifold geometry with safety probability, enabling efficient O(1) seamless defense against arbitrary OOD intents. Experiments demonstrate that CMP achieves up to a 10-fold survival rate improvement in typical OOD scenarios where baselines suffer catastrophic failure, incurring under 10% tracking degradation. Notably, the system exhibits emergent ``best-effort'' generalization behaviors to progressively accomplish OOD goals by adhering to the competence boundaries. Result videos are available at: https://shepherd1226.github.io/CMP.

Problem

• whole-body tracking
• loco-manipulation
• Out-of-Distribution
• robustness
• mobile manipulators

Innovation

• Competence Manifold Projection
• Out-of-Distribution Robustness
• Whole-Body Control
• Isomorphic Latent Space
• Safety-Constrained Learning