Copying the Human Hand Does Not Automatically Make It Easier to Drive
Build a hand that looks like yours, and control should get easier. arXiv:2609.05206, posted 4 September from Pázmány Péter Catholic University in Budapest, measures that claim instead of assuming it.
The two subjects sit at opposite ends of a four-part taxonomy: joint-axis geometry, actuator-to-DoF ratio, coupling, and how authority is spread across joints. Shadow Dexterous Hand (SDH) is orthogonal axes, slightly underactuated (m/n = 0.82), sparse coupling. The Anatomically Correct, Biomechatronic Hand (ACBH) is oblique axes, overactuated (m/n = 1.50), branching tendons.
Where anatomy helps, and where it does not
They split the mapping. J is joint motion to fingertip. A is actuators to joints. B = JA is what a controller actually drives. Condition number κ near 1 is even; large κ means some directions eat disproportionate effort.
At a mid-range pose, long-finger κ(J) is 3.84 on SDH versus 5.77–9.55 on ACBH. Orthogonal axes win there. The thumb flips: ACBH κ(J) = 4.76, SDH 9.45. Human-like CMC geometry covers opposition better.
Actuation is the other half. SDH’s PIP/DIP 1:1 coupler slightly worsens the index (κ(B) = 4.25 vs κ(J) = 3.84). ACBH’s tendon web does the opposite on every long finger: κ(B) < κ(J). Extra muscles spread authority. The thumb is the outlier. ACBH thumb κ(A) ≈ 30.74 and κ(B) = 20.03, against SDH’s fully actuated 9.45. All 11 thumb muscles hit the CMC yaw/pitch; the IP joint is left with 2. That is opposition anatomy, not a modeling bug.
Does RL notice?
Reach with PPO: SDH hits 0.75 success around 11 million steps (~6 h). ACBH hits 0.65 around 25 million (~19.5 h). BlockRotateZ with DDPG+HER is a tie near 0.67–0.69 at 16 million steps. Full BlockRotateXYZ with TQC+HER: SDH about 0.76 after ~23 million steps, ACBH about 0.25 in the same step budget (and more wall time). DDPG+HER on that task stays under 0.1 for both.
Parameters come from Shadow’s URDF and the ACBH MuJoCo model of Polcz et al. Numbers are at one mid-range pose plus some workspace sweeps, not a factory grip set.
A Human’s Take
I have watched too many hand launches sell “human-like” as if that were a control spec. This paper’s useful sentence is that the ACBH thumb is kinematically nicer and actuator-ugly. If you copy thenar muscles, you inherit their favorites. Design the coupling you can afford to learn, not the one that photographs well.