A team has open-sourced a robot hand designed to learn entirely in simulation, then work on real hardware without extra tuning.
Aero Hand Open is a tendon-driven hand: instead of putting a motor in every joint, it routes force through cables from motors mounted elsewhere, so smaller motors suffice and one motor can drive several joints at once. That is what makes a hand this capable affordable to build, but it also makes the hand brutally hard to simulate, since the cable-driven joints are not independently controllable and the transmission itself resists modeling. The release ships three pieces to solve that: a simulator that reproduces the cable transmission, an identified map that translates between simulated states and real motor commands (including the three-way coupling in the thumb), and a reinforcement learning package for training grasping policies. A policy trained purely in that simulator can then run on the physical hand with no fine-tuning and no added state estimation.
That closes a real gap. Sim-to-real training already works well for direct-drive robot arms, but dexterous hands have lagged because tendon transmissions are so poorly behaved in simulation. The catch is that the actuation map is specific to this hand's own cable routing, so a policy trained here will not automatically transfer to a different tendon-driven design.
Even with that limit, publishing the mechanical design, simulator, mapping and training code together gives dexterous-hand research a shared, reproducible starting point instead of another one-off hand only its own lab can use.