REASAN: Learning Reactive Safe Navigation for Legged Robots

TL;DR

REASAN tackles the challenge of onboard reactive navigation for legged robots in dynamic environments by decomposing control into four lightweight modules: locomotion, safety shielding, navigation, and a Transformer-based exteroceptive estimator that processes raw LiDAR. The approach trains each module sequentially in simulation with tailored rewards and curricula, avoiding heuristic policy switching, and demonstrates real-time onboard performance in both single- and multi-robot scenarios. Comprehensive simulations show modular learning outperforms end-to-end baselines and validate the exteroceptive estimator's architectural choices; real-world experiments confirm robustness to static and dynamic obstacles, detours, and dead ends. The work provides open-source training and deployment code and highlights practical considerations for sim-to-real transfer and future enhancements such as longer-horizon prediction, uneven terrain handling, and semantic perception with cameras.

Abstract

We present a novel modularized end-to-end framework for legged reactive navigation in complex dynamic environments using a single light detection and ranging (LiDAR) sensor. The system comprises four simulation-trained modules: three reinforcement-learning (RL) policies for locomotion, safety shielding, and navigation, and a transformer-based exteroceptive estimator that processes raw point-cloud inputs. This modular decomposition of complex legged motor-control tasks enables lightweight neural networks with simple architectures, trained using standard RL practices with targeted reward shaping and curriculum design, without reliance on heuristics or sophisticated policy-switching mechanisms. We conduct comprehensive ablations to validate our design choices and demonstrate improved robustness compared to existing approaches in challenging navigation tasks. The resulting reactive safe navigation (REASAN) system achieves fully onboard and real-time reactive navigation across both single- and multi-robot settings in complex environments. We release our training and deployment code at https://github.com/ASIG-X/REASAN.

Figures (10)

• Figure 1: REASAN deployed onboard two quadrupedal robots.
• Figure 2: System pipeline of REASAN with a highly modularized design for both training and deployment.
• Figure 3: Environments for learning the safety-shield (a) and navigation policies (b)-(c). Yellow arrows indicate guidance.
• Figure 4: Network architecture of the exteroceptive estimator.
• Figure 5: Evaluation scenarios in simulation. We simulate ray-based exteroception in (a)-(d) and raw LiDAR scans in (e).