Rhoda AI and the robots that predict their next move

By TheSequence · Fri, 30 Oct 2026 11:02:15 +0000.

Issue #962 of The Sequence Robotics examines Rhoda AI's bet: using short-term sensory future prediction to organize a robot's behavior, while a separate inverse dynamics model converts each predicted future into concrete movements.

According to the newsletter's summary (TL;DR), the company's scaling report, published in September, links better video prediction to stronger performance on a bearing-unpacking task. Its October report on tactile sensing extends that same prediction loop to touch.

The newsletter illustrates the intuition with an everyday example: when watching someone empty a cardboard box, the brain anticipates what comes next (the flap opens, the bag lifts, the contents fall out). For a robot, however, recognizing that sequence leaves an awkward problem unresolved: which motors should move, and what happens if the bag gets stuck. A convincing film of a job well done does not automatically translate into a job well done.

Rhoda AI is building its company on that translation. The appeal, according to the text, is the vast availability of training videos; the challenge is ensuring that the lessons drawn from them survive contact with a real factory.

The open question the newsletter highlights is whether these results, reported by the company itself and obtained on a single task with a single type of robot body (embodiment), carry over to other workflows, to repeated training runs and to sustained industrial operation.

Note: the body of the received email is truncated after the introduction, so no further technical details, figures or specific results from the full analysis are available.