A robot learned to picture the mess before it happens, and its sushi turned out better for it.
Researchers built TacSushi, a robot control system that combines camera images, language instructions, hand position, and fingertip touch sensors to grip, cut, and plate sushi. During training, the system also predicted what its camera would see next, how close it was to finishing the task, and the risk of losing grip, using data from both successful and failed real-robot attempts. That prediction module gets stripped out once the robot is actually working, but the lessons stick. Trained on 340 successful and 50 failed trials, then tested across 600 rollouts, TacSushi hit 68.3% success on tasks it had seen before and 37.5% on unfamiliar ingredients, versus 36.7% and 10% for a version without the future-prediction training, an outcome that nearly quadruples out-of-distribution success.
The out-of-distribution number is the one worth watching. Most robot manipulation demos look great on the exact task they were trained for and fall apart the moment something changes, like a different cut of fish or an unfamiliar wrapper. Forcing the model to reason about consequences, not just mimic motions, appears to buy real generalization, and the team's decision to grade outcomes with both human raters and vision-language models sidesteps the usual trick of judging food robots by a single geometry threshold.
Don't picture this replacing your neighborhood sushi chef anytime soon. The training set is a few hundred trials, the test kitchen is a lab, and 37.5% success on a new ingredient still means the robot botches the roll more often than not.