Mobile ALOHA

Stanford's 2024 project that puts the ALOHA bimanual robot on a mobile base for low-cost whole-body teleoperation and imitation learning.

Mobile ALOHA is work published in January 2024 by Zipeng Fu, Tony Zhao, and Chelsea Finn at Stanford, presented at CoRL 2024. It mounts the previously tabletop-fixed ALOHA bimanual platform onto a Tracer mobile base from AgileX Robotics; the operator's waist is linked to the base so that walking moves the base along, while both hands drive the master arms, teleoperating the arms and base together. The full hardware budget is about $32,000. Actions are 16-dimensional: 14 joint positions for the two arms, plus the base's linear and angular velocity. The problem it targets is data collection and learning for mobile manipulation — using both hands while on the move. Another key finding is co-training: training on the new mobile data together with existing stationary ALOHA data raises success rates by up to 90%, using only about 50 demonstrations per task. The policy itself uses off-the-shelf imitation-learning methods such as ACT and Diffusion Policy.

ExampleMobile ALOHA has autonomously stir-fried shrimp and plated it, opened a two-door cabinet to store a heavy pot, called and boarded an elevator, and rinsed a used pan under a faucet.

Also called

Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation

Related

ALOHA · Action Chunking with Transformers · Mobile Manipulation · Co-training · Whole-Body Teleoperation · AgileX Robotics

Sources

Mobile ALOHA (arXiv 2401.02117)
Mobile ALOHA 项目主页 (Chinese)

As of

2024-01