Machenta robots learn a new task from one demonstration

Show a task once, rearrange the objects, and the robot gets to work. Machenta's October 27 demonstration claims new tasks need no additional training.

The company says its simulation-trained model uses a demonstration and camera observations to act. One company demonstration teaches three cooperating robots in two minutes.

There is an academic predecessor: Imperial College's Instant Policy, presented at ICLR 2025, studies immediate imitation after one or two demonstrations. It represents observations, demonstrations and actions as a graph, generating behavior from new examples without retraining weights for every task. That could make changing jobs easier than relying on fixed routines.

Deployment reliability is another question. Machenta reports 97 successful bottle-righting attempts out of 99 in a simple test, not arbitrary workplace performance. The earlier paper is not independent validation of the company's current system.

Source dates: Machenta, October 27, 2026; clock time and timezone unspecified, so Beijing and US Eastern calendar dates cannot be determined. The Instant Policy project identifies ICLR 2025 but gives no exact publication time.