ROBEL: Robotics Benchmarks for Learning with Low-Cost Robots

• DOI: 10.48550/arxiv.1909.11639
• Publication Date: 2019-01-01
• Authors: Michael J. Ahn; Henry Zhu; Kristian Hartikainen; Hugo Ponte; Abhishek Gupta; Sergey Levine; Vikash Kumar
• Keywords: Leverage (statistics); Benchmark (surveying)

ROBEL is an open-source platform of cost-effective robots designed for reinforcement learning in the real world. introduces two robots, each aimed to accelerate research different task domains: D'Claw a three-fingered hand robot that facilitates dexterous manipulation tasks, and D'Kitty four-legged agile legged locomotion tasks. These low-cost, modular are easy maintain robust enough sustain on-hardware from scratch with over 14000 training hours registered on them date. To leverage this platform, we propose extensible set continuous control benchmark tasks robot. feature dense sparse objectives, additionally introduce score metrics as hardware-safety. We provide scores initial using variety learning-based methods. Furthermore, show these results can be replicated across copies located institutions. Code, documentation, design files, detailed assembly instructions, final policies, baseline details, videos, all supplementary materials required reproduce available at www.roboticsbenchmarks.org.