MARG: MAstering Risky Gap Terrains for Legged Robots with Elevation Mapping
YINZHAO DONG, Ji Ma, Liu Zhao, Wanyue Li, Peng Lu
AI summary
Problem
Existing blind locomotion controllers lack environmental perception and fail on risky gaps, while perception-based methods rely on complex multi-sensor setups or expensive computing resources, hindering safe and efficient real-world deployment.
Approach
MARG fuses an asymmetric actor-critic DRL policy with a robot-centered elevation map and estimated privileged states, trained with specialized foothold-safety rewards to dynamically adjust actions and maintain stability.
Key results
- Predicts body velocity and foot contact state to enhance stability in risky gaps
- Introduces three foot-related rewards (air time, stumble, center) to promote safe footholds
- Generates drift-minimized, robot-centered height maps using a single LiDAR
- Achieves zero-shot real-world transfer across 65 cm gaps, 18 cm bridges, and variable beams
Why it matters
Provides a practical, low-cost framework for deploying safe legged locomotion in unstructured environments without expensive hardware or prior mapping.
Abstract
Deep Reinforcement Learning (DRL) controllers for quadrupedal locomotion have demonstrated impressive per- formance on challenging terrains, allowing robots to execute complex skills such as climbing, running, and jumping. However, existing blind locomotion controllers often struggle to ensure safety and efficient traversal through risky gap terrains, which are typically highly complex, requiring robots to perceive terrain information and select appropriate footholds during locomotion accurately. Meanwhile, existing perception-based controllers still present several practical limitations, including a complex multi- sensor deployment system and expensive computing resource requirements. This paper proposes a DRL controller named MAstering Risky Gap Terrains (MARG), which integrates terrain maps and proprioception to dynamically adjust the action and enhance the robot’s stability in these tasks. During the training phase, our controller accelerates policy optimization by selectively incorporating privileged information (e.g., center of mass, friction coefficients) that are available in simulation but unmeasurable directly in real-world deployments due to sensor limitations. We also designed three foot-related rewards to encourage the robot to explore safe footholds. More importantly, a terrain map generation (TMG) model is proposed to reduce the drift existing in mapping and provide accurate terrain maps using only one LiDAR, providing a foundation for zero-shot transfer of the learned policy. The experimental results indicate that MARG maintains stability in various risky terrain tasks.