Robot hands just got better at figuring out where and how to touch things, without a human coding in the answer.
Researchers built ContactExplorer, a reinforcement learning method that gives dexterous robot hands a reason to try new ways of touching objects. It tracks contact as which fingertips touch which parts of an object, then keeps a counter of those patterns using hashed representations of the object's state. Two rewards do the work: one for discovering contact patterns the hand hasn't tried, another for pushing the hand toward object regions it hasn't reached yet. Across seven manipulation tasks and five different hand designs, ContactExplorer beat existing exploration methods on both sample efficiency and success rate, and needed fewer task-specific hints to get there. The team also reports it transfers to real hardware, not just simulation.
That generality is the interesting part. Dexterous manipulation has been a weak spot for reinforcement learning precisely because novelty is hard to define when the real signal is finger-by-finger contact, not just where a joint moves. Most existing fixes bolt on priors tuned to one task or one gripper, which stops working the moment the hardware or the object changes. A reward signal that holds up across five hand embodiments points toward training dexterous hands without rebuilding the reward function every time.
None of this means a robot is folding your laundry soon. Five simulated hand types and seven tasks is a useful benchmark, not a warehouse floor, and real-world contact is messier than any hash code.