A new training trick helps robot hands discover coordinated grasps instead of flailing joint by joint.
Researchers describe EigenDEXplore, a method for teaching robotic hands dexterous skills like grasping, in-hand reorientation, and contact-rich manipulation. Instead of perturbing each joint independently, which rarely stumbles onto coordinated motions, it adds noise along directions derived from human hand-motion data on top of ordinary joint-space noise, while leaving the robot's action space untouched. The team tested the approach across multiple robotic hands and training setups, including unstructured reinforcement learning, reference-guided reinforcement learning, trajectory optimization, and sim-to-real transfer. It consistently beat both plain joint-space exploration and earlier methods that learn a separate, low-dimensional action space from human data.
Earlier work used human hand poses to shrink the action space itself, trading away generality for easier search, or stitched human-derived and joint-space actions together into redundant, harder-to-tune systems. This paper's finding is simpler: leave the action space alone and just bias how the robot explores it, a change that costs little and preserves flexibility. The improvement is largest when researchers strip out reward shaping and curricula, suggesting the method is picking up slack engineers currently fill in by hand.
It is still a benchmark win, not a robot doing your dishes, and the real test is whether 'consistently outperforms' survives contact with a hand that has to work every single time.