UMC: Unified Resilient Controller for Legged Robots with Joint Malfunctions

TL;DR

This work tackles the challenge of making legged robots robust to unpredictable damages by introducing the Unified Malfunction Controller (UMC), a model-free framework that uses a masking mechanism to ignore malfunctioning joints. UMC employs a two-stage training pipeline—pretraining on normal conditions and fine-tuning with damage scenarios—allowing rapid adaptation without sacrificing normal performance. The framework supports both transformer- and MLP-based actors through a damage-detection module, a tokenizer, a mask encoder, and a detokenizer, with a masking strategy applied in self-attention. Empirical results across A1-Walk, Unitree H1/G1, and Solo8 show substantial reductions in failure rates and improvements in task completion, outperforming state-of-the-art methods while preserving normal-condition performance, signaling strong practical impact for resilient legged robotics.

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 complex maintenance. To address this issue, we first analyze and summarize eight types of damage scenarios, including sensor failures and joint malfunctions. Then, we propose a novel, model-free, two-stage training framework, Unified Malfunction Controller (UMC), incorporating a masking mechanism to enhance damage resilience. Specifically, the model is initially trained with normal environments to ensure robust performance under standard conditions. In the second stage, we use masks to prevent the legged robot from relying on malfunctioning limbs, enabling adaptive gait and movement adjustments upon malfunction. Experimental results demonstrate that our approach improves the task completion capability by an average of 36% for the transformer and 39% for the MLP across three locomotion tasks. The source code and trained models will be made available to the public.

Figures (12)

• Figure 1: Qualitative and Quantitative Comparison of Our UMC Framework with Baselines and a SOTA Method. 'Trf' is the transformer, 'NM' represents the baseline normal training (trained without damaged situations), and 'UMC' is our method. 'BodyTrf' is the abbreviation of BodyTransformer sferrazza2024bodytransformerleveragingrobot, also a baseline structure. 'Failure Rate' refers to the proportion of robots that fall during their actions, and '3 Units Reached' refers to the percentage of robots that are still able to move a distance of 3 units after the added damage. (a) presents the qualitative comparison for the humanoid robot task, while (b) depicts it for the quadruped task. The last comparison set in (b) shows the qualitative comparison between our UMC framework and the SOTA method proposed in hou2024multitasklearningactivefaulttolerant.
• Figure 2: UMC system for transformer-based Actor-Model Architecture. $K$ is the number of encoder layers. For more details of the architecture, please refer to \ref{['sec:umcframework']}.
• Figure 3: Qualitative Comparison Between Methods Under Damaged Scenarios. 'Baseline' refers to robots trained using baseline methods, while 'UMC' denotes robots trained with the UMC method. Figure 'b2)' shows a snapshot of the original trajectory at a specific time point under undamaged conditions, while b1) and b3) are in damaged conditions.
• Figure 4: Comparison of One-Stage and Two-Stage Training in the G1 Task.
• Figure 5: Demonstration of different damage conditions.