FTR‐Bench: Benchmarking Deep Reinforcement Learning for Flipper‐Track Robot Control

摘要

ABSTRACT Tracked robots equipped with flippers and sensors are extensively employed in outdoor search and rescue scenarios. However, achieving precise motion control on complex terrains remains a significant challenge, often necessitating expert teleoperation. This stems from the high degree of robot joint freedom and the need for precise flipper coordination based on terrain roughness. To address this problem, we propose F lipper‐ T rack R obot Bench mark ( FTR‐Bench ), a simulator featuring flipper‐track robots tasked with crossing various obstacles using reinforcement learning (RL) algorithms. The primary objective is to enable autonomous locomotion in environments that are too remote or hazardous for humans, such as disaster zones or planetary surfaces. Built on Isaac Lab, FTR‐Bench achieves efficient RL training at over 4000 FPS on an RTX 3070 GPU. Additionally, it integrates RL algorithms with OpenAI Gym interface specifications, enabling fast secondary development and verification. On this basis, FTR‐Bench provides a series of standardized RL‐based benchmarking experiments baselines for obstacle‐crossing tasks, providing a solid foundation for subsequent algorithm design and performance comparison. Experimental results empirically indicate that SAC algorithms performs relatively well in single and mixed terrain traversal, but most algorithms struggle with multi‐terrain traversal skills, which calls the RL community for more substantial development. Our project is open‐source at https://github.com/nubot-nudt/FTR-Benchmark .