[Paper Review] Resilient Legged Local Navigation: Learning to Traverse with Compromised Perception End-to-End

Jin Jin, Chong Zhang+5 more|arXiv (Cornell University)|Oct 5, 2023

Robotic Locomotion and Control1 citations

TL;DR

This paper proposes an end-to-end reinforcement learning (RL) policy for legged robots that enables resilient local navigation under compromised perception, such as invisible obstacles or pits. By integrating proprioceptive and exteroceptive feedback from the locomotion system into the navigation policy, the method learns to detect and react to perception failures—like collisions or missed steps—without relying on heuristic anomaly detection, achieving a 30% higher success rate than classical planners in perception failure scenarios.

ABSTRACT

Autonomous robots must navigate reliably in unknown environments even under compromised exteroceptive perception, or perception failures. Such failures often occur when harsh environments lead to degraded sensing, or when the perception algorithm misinterprets the scene due to limited generalization. In this paper, we model perception failures as invisible obstacles and pits, and train a reinforcement learning (RL) based local navigation policy to guide our legged robot. Unlike previous works relying on heuristics and anomaly detection to update navigational information, we train our navigation policy to reconstruct the environment information in the latent space from corrupted perception and react to perception failures end-to-end. To this end, we incorporate both proprioception and exteroception into our policy inputs, thereby enabling the policy to sense collisions on different body parts and pits, prompting corresponding reactions. We validate our approach in simulation and on the real quadruped robot ANYmal running in real-time (<10 ms CPU inference). In a quantitative comparison with existing heuristic-based locally reactive planners, our policy increases the success rate over 30% when facing perception failures. Project Page: https://bit.ly/45NBTuh.

Motivation & Objective

• To develop a navigation policy that remains robust when exteroceptive perception fails, such as in the presence of invisible obstacles or pits.
• To eliminate reliance on heuristic-based anomaly detection and replanning for perception failure recovery.
• To enable real-time, end-to-end learning of navigation strategies that react to physical feedback from the robot’s body during locomotion.
• To validate the method in both simulation and real-world deployment on the ANYmal quadruped robot.
• To demonstrate that legged robots can successfully navigate even when fully blind to obstacles, using only body feedback.

Proposed method

• Trains an asymmetric actor-critic deep RL policy that takes corrupted exteroception and proprioceptive feedback as input to generate velocity commands.
• Incorporates locomotion-level exteroception (e.g., foot contact forces) and proprioception (e.g., joint angles, base IMU) as key observations to detect collisions and missed steps.
• Uses a memory-augmented actor to infer environment state from partial or corrupted sensor inputs, improving robustness to perception failures.
• Employs a critic with access to privileged information (e.g., perfect map, external forces) to guide training and stabilize learning.
• Designs a custom simulation environment with controllable perception failure rates (e.g., 0%, 50%, 100% obstacle/pit visibility) to train and evaluate the policy.
• Applies regularization to implicitly infer the true environment state in latent space, preventing suboptimal policy convergence.

Experimental results

Research questions

• RQ1 Can an end-to-end RL policy learn to navigate effectively when perception fails, such as when obstacles or pits are invisible?
• RQ2 How do proprioception and exteroception from the locomotion system contribute to detecting and recovering from perception failures?
• RQ3 Does incorporating memory and latent state inference improve robustness to perception degradation compared to baseline methods?
• RQ4 Can the policy generalize to real-world navigation under full perception failure, such as being completely blind?
• RQ5 How does the policy compare quantitatively to heuristic-based planners under varying levels of perception corruption?

Key findings

• The proposed RL policy increases success rate by over 30% compared to classical heuristic-based planners under perception failures, with a 93.0% success rate on pits with 0% visibility.
• In obstacle scenarios with 50% visibility, the policy achieves 92.0% success rate versus 54.0% for the baseline planner, with a p-value < 0.001.
• Proprioception alone improves robustness, reducing base collisions from 48.3 to 4.6 in 0% visibility obstacle scenarios.
• Locomotion-level exteroception significantly improves pit detection, reducing average time cost from 22.5s to 12.3s in 0% visibility pit scenarios.
• Regularization for implicit state inference is critical—without it, performance drops significantly, especially under high perception corruption.
• The policy successfully navigates real-world environments even when fully blind, reacting to collisions on the base, thighs, and legs, and recovering from invisible pits using lateral velocity commands.