← Back ICRA 2026

RENet: Fault-Tolerant Motion Control for Quadruped Robots Via Redundant Estimator Networks under Visual Collapse

Yueqi ZHang, Quancheng Qian, Taixian Hou, Peng Zhai, Xiaoyi Wei, Kangmai Hu, Jiafu Yi, Lihua ZHang

PDF

AI summary

Key figure (auto-extracted from paper)

A dual-estimator framework that automatically switches between vision and proprioception to maintain stable quadruped locomotion during visual sensor failure.

Legged robots fault-tolerant control vision-proprioception fusion reinforcement learning state estimation sim-to-real transfer

Problem

Vision-based quadruped locomotion fails under real-world depth sensor noise and visual degradation, while proprioception-only methods struggle with complex terrain, leaving a critical gap for robust outdoor deployment.

Approach

RENet trains two parallel estimator networks jointly in a single stage and uses an auto-selector module to monitor visual noise, seamlessly switching to a fallback proprioception estimator when vision becomes unreliable.

Key results

Single-stage training framework enabling robust locomotion under visual degradation
Self-adaptive switching mechanism that autonomously detects depth sensor noise
Zero fine-tuning real-world deployment on a Unitree GO1 across complex indoor and outdoor terrains
Lightweight <3MB model deployment with efficient GPU raycasting for rapid sim-to-real transfer

Why it matters

Enables reliable outdoor quadruped robot deployment by ensuring continuous stability when visual sensors fail, bridging the gap between simulation and real-world field conditions.

Abstract

Vision-based locomotion in outdoor environments presents significant challenges for quadruped robots. Accurate environmental prediction and effective handling of depth sensor noise during real-world deployment remain difficult, severely re- stricting the outdoor applications of such algorithms. To address these deployment challenges in vision-based motion control, this letter proposes the Redundant Estimator Network (RENet) frame- work. The framework employs a dual-estimator architecture that ensures robust motion performance while maintaining deployment stability during onboard vision failures. Through an online estima- tor adaptation, our method enables seamless transitions between estimation modules when handling visual perception uncertain- ties. Experimental validation on a real-world robot demonstrates the framework’s effectiveness in complex outdoor environments, showing particular advantages in scenarios with degraded visual perception. This framework demonstrates its potential as a prac- tical solution for reliable robotic deployment in challenging field conditions.

Index terms

Legged Robots Reinforcement Learning