UMC: A Unified Approach for Resilient Control of Legged Robots Across Masked Malfunction Training

Yu Qiu, Xin Lin, Runfa Li, Dizhe Zhang, Jingbo Wang, Xiangtai Li, Bo Du, Truong Nguyen, Lu Qi

Keywords: Embodied AI

Abstract: Adaptation to unpredictable damages is crucial for autonomous legged robots, yet existing methods based on multi-policy or meta-learning frameworks face challenges like limited generalization and maintenance. In this work, we first provide a systematic categorization of eight representative malfunction scenarios, covering both detectable and undetectable cases. Then, we propose a novel model-free framework, \textbf{U}nified \textbf{M}alfunction \textbf{C}ontroller (UMC). UMC employs a two-stage training pipeline: the first stage learns strong baseline locomotion in undamaged environments, while the second stage fine-tunes the controller with mixed malfunction scenarios to encourage adaptive and robust behavior. For detectable settings, we introduce a masking strategy that explicitly filters corrupted signals, preventing error propagation and enabling policies to rely on functional joints. UMC is compatible with both transformer and MLP backbones. Experiments across multiple humanoid and quadruped tasks show that UMC consistently reduces failure rates and improves task completion under diverse damage conditions. The source code and trained models will be made available to the public.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 9655