Research Analyzer
← Back ICRA 2026

Knowledge-Based Locomotion Policy for Quadruped Robots under Incomplete Terrain Observation

Taehyeong Kim, Sangmoon Lee

PDF

AI summary

Key figure (auto-extracted from paper)
The recurrent policy actively encodes implicit stair geometry into its hidden state, enabling robust quadruped stair traversal even under occluded LiDAR conditions.
quadruped locomotion partial observability recurrent policy LiDAR occlusion stair traversal implicit terrain knowledge

Problem

Body-mounted LiDAR sensors on quadruped robots suffer from systematic blind spots during stair locomotion, creating a partial observability problem that standard single-step terrain policies cannot resolve.

Approach

A GRU-based recurrent policy fuses pointcloud and proprioceptive inputs over time to build an implicit knowledge representation of stair geometry, trained with an adaptive pointcloud curriculum and auxiliary proprioceptive supervision.

Key results

  • Near-perfect success rates on flat and stair ascent/descent tasks under normal LiDAR conditions
  • Catastrophic performance collapse when LiDAR input is masked at inference, proving recurrent state encodes critical terrain cues
  • Adaptive pointcloud curriculum enabling stable training transition from reference to occluded sensor data
  • Proprioceptive auxiliary supervision improving recurrent state learning under degraded exteroceptive conditions

Why it matters

Enables reliable quadruped navigation in real-world environments where sensor occlusion is unavoidable, advancing robust legged robot deployment.

Abstract

Body-mounted LiDAR sensors suffer from sys- tematic blind spots during stair locomotion, creating a partial observability problem that single-step terrain snapshots cannot resolve. We address this with a recurrent locomotion policy for the Unitree Go2 that builds implicit knowledge of stair geometry through a GRU-based recurrent encoder over pointcloud and proprioceptive inputs, enabling robust stair ascent and descent even under occluded LiDAR conditions. Ablation experiments show that masking pointcloud input at inference time causes catastrophic failure on stair terrain and severe performance degradation overall, confirming that implicit stair knowledge is a critical cue for step negotiation rather than a merely complementary signal.

Index terms

Legged Robots Reinforcement Learning Reactive and Sensor-Based Planning

Related papers