Is the Mobile ALOHA (Stanford) autonomous?

Mobile ALOHA is a research platform developed at Stanford University (IRIS/REAL Lab) by Zipeng Fu, Tony Z. Zhao, and Chelsea Finn, funded by the Boston Dynamics AI Institute and ONR. It augments the stationary ALOHA bimanual teleoperation system with an AgileX Tracer mobile base and a whole-body teleoperation interface, costing approximately $32,000 for the full build. The system is designed primarily for data collection via human teleoperation, after which imitation learning (Action Chunking with Transformers, ACT) with co-training on existing static datasets enables autonomous task execution at up to 90% success-rate improvement across tasks like cooking shrimp, using elevators, and cabinet manipulation — all demonstrated in a controlled lab setting. The platform is fully open-source and is a research prototype, not a commercial product.

Mobile ALOHA operates in two distinct modes: (1) teleoperation for data collection, where a human physically performs tasks via the leader-follower interface, and (2) autonomous execution of trained tasks using imitation learning (ACT + co-training), where the robot performs the task without a human driving it. The autonomy verdict applies to mode (2), which is the system's stated research goal and is demonstrated in the official project results. The research paper, PMLR proceedings, and official project page all confirm autonomous task completion (cooking, elevator use, cabinet manipulation) after training on ~50 demonstrations. The teleoperation disclaimers on viral videos refer to data-collection mode, not the autonomous policy execution mode — the project website explicitly distinguishes these. Confidence is moderate (not high) because: autonomous results are demonstrated only in controlled lab conditions on specific trained tasks; generalization to novel environments or tasks is not established; and the system is a research prototype, not a deployed product. The human's role in teleoperation is for data collection (setup/training), not for performing the task during autonomous operation — this does not disqualify the Autonomous classification per the definitions provided.

Claimed vs. Real