Teach a robot by moving its arms: Flash ALOHA opens its training code

Fifty demonstrations per task were enough for a two-arm robot to learn tasks including catching a ball and clearing a table, Flash ALOHA's team reports. It claims success rates above 80% at close to human speed. The accompanying code and project materials opened on November 12.

Teaching starts with a person physically guiding the robot's arms while cameras and joint sensors record the demonstration. Gravity and friction compensation make the arms easier to move. The repository connects recording, policy training and execution, with example configurations for one arm or two. An earlier version appeared as ALOHA Lightning in July.

Fast movements also need clean handoffs between successive batches of actions. The code can blend neighboring batches and train a policy to continue an already committed motion plan. Image masking is available to handle the operator visible during teaching. Publishing the machinery behind the demonstration gives other labs a way to inspect, adapt and reproduce the workflow. The repository's software and accompanying documentation are released under the MIT License.

This remains research hardware and software. Training recordings and model checkpoints are not included, and the sample calibration came from one physical setup. Other arms and cameras need fresh calibration, and execution commands really move the equipment. Demonstrated task performance does not establish reliable housekeeping in unfamiliar homes. The nearer-term benefit is a more accessible way for researchers to teach robots and examine where they fail.

Source times: initial code-release commit, November 12, 2026, 09:11:40 Beijing time / November 11, 20:11:40 EST; project-page launch commit, November 12, 2026, 10:28:48 Beijing time / November 11, 21:28:48 EST.