TacSushi keeps failed food-handling attempts as training lessons

A robot learning sushi preparation can keep its mistakes as teaching material without copying the failed movements. TacSushi uses recorded consequences during training. First posted as a preprint on September 17, it appears among the accepted papers for the November 12 dexterous-manipulation workshop at CoRL in Austin, Texas, United States. Mitsubishi Electric Research Laboratories issued a technical report dated November 13.

Author Haodi Hu’s project page groups demonstrations involving rice, seaweed and fish with fingertip-force and teleoperation tests. The manipulation clips are labelled six-times speed. That is an important viewing distinction: completing a movement and completing it quickly are different achievements. Speeding up footage helps a viewer follow an entire sequence, while changing the apparent pace. Watching ingredients being supported and repositioned also reveals more about the task than a photograph of a finished piece of sushi.

The platform is a Shadow Hand. Shadow’s 2022 Lite specification describes a research system with actuation and sensing built into the hand and forearm. It provides joint-position, fingertip-contact and tendon-load information, supports alternative compatible tactile sensors, and allows different control approaches. Hardware supplies measurements and moves the fingers; a learning system still has to decide how those measurements should change the next action. A new policy should therefore be understood separately from the established robotic hand it uses.

Think of teaching someone to shape a rice ball. A successful demonstration shows a movement to repeat; a collapsed ball gives a mistake to discuss. TacSushi’s consequence predictor is used for training and removed for execution. Six methods underwent 600 total trials, 20 per method-task pairing. Reported success falls from 68.3% on familiar tasks to 37.5% on two unfamiliar ingredients. Its outcome measure combines five human and three vision-language-model ratings of appearance, rather than judging taste. A convincing-looking result and a satisfying meal remain different tests.

The workshop lists TacSushi among accepted contributions without guaranteed poster space. Its stated questions include how to compare different robotic approaches and what evidence would establish general dexterity. Those are useful questions for a kitchen demonstration too. One successful example invites attention; dependable help requires performance that survives changes in ingredients and surroundings. Keeping a record of mistakes makes that gap easier to discuss concretely, instead of letting the smoothest clip stand in for the entire experiment.

Original preprint: September 17, 2026, 11:01 Beijing time / September 16, 11:01 PM EDT. MERL technical report dated November 13, 2026.