Last Updated on
Benchmarks made to measure the performance of AI systems that control robots are often limited to expensive hardware designed for industrial environments that can cost tens of thousands of dollars. Researchers from UC Berkeley and Google Brain addressed this problem by introducing Robotics Benchmarks for Learning with Low-Cost Robots (ROBEL), an open source platform designed to encourage rapid experimentation and on-hardware reinforcement learning. ROBEL also comes with benchmark tasks specifically made for tracking the quality of AI systems on lower-cost robots.
Naturally, more affordable robots designed to work with platforms and performance benchmarks make adoption more likely by developers, students, or startups interested in iterating to advance the field.
ROBEL is made to work with D’Claw, a three-fingered robotic hand, and D’Kitty, a four-legged robot. Made by Trossen Robotics, fully assembled versions of these robots sell for $3,200 and $3,700 respectively. D’Claw is a 9-degrees-of-freedom (9DoF) device while D’Kitty is a 12-degrees-of-freedom device (12DoF).
“These low-cost, modular robots are easy to maintain and are robust enough to sustain on-hardware reinforcement learning from scratch with over 14,000 training hours registered on them to date,” ROBEL makers said in a paper published on arXiv.
By comparison, Sawyer and Baxter robots have a reputation in lab environments but can cost well over $15,000, while robotic arms like Franka can cost more than $10,000. Baxter maker Rethink Robotics was acquired by Hahn Group in late 2018.
Lowering costs and increasing affordability have been a focus of robotics researchers in recent months. Earlier this year, UC Berkeley released Blue, a two-armed robot with grippers that was lauded for costing less than $5,000. PyRobot, a robotics framework released this summer by Facebook AI Research, works at launch with LoCoBot, a $5,000 robot.
The ROBEL benchmark is made for carrying out tasks like “turn and screw” for D’Claw and “stand and walk” for D’Kitty. A robotic hand good at screwing or unscrewing things could be used to manipulate valves in a factory or simply to open a jar, while a robot that knows how to get around on four legs can use locomotion to climb over obstacles or travel.
To help users speed development and experimentation, ROBEL also includes a simulator to allow the incorporation of synthetic data.
ROBEL will be presented at the Conference on Robotic Learning (CoRL 2019) being held October 30 at the University of Tokyo.