Robots learn to grip objects by following high-level instructions instead of precise hand positions

A new system lets robots learn dexterous hand control by following simple grasp categories (like 'pinch' or 'wrap') rather than detailed finger-position targets. In real-world tests, the approach reached 87.9% success on novel objects and remained controllable when interventions were needed, meaning humans could adjust strategy without retraining the entire system.

Robot hands have been hard to train because specifying exactly where each finger should go for each object is tedious and brittle. This work shows that steering robots toward a category of grasp (the taxonomy) lets them figure out the precise finger motions on their own, generalizes better to new objects, and stays interpretable enough that a human can intervene. The practical implication: robot systems become faster to deploy and easier to adjust when they fail, which matters for any company trying to scale manipulation robotics beyond laboratory conditions.