Boston Dynamics Reveals Redesigned Atlas Hand: 13 DoF, Four Fingers, Built for Mass Manufacturing

Boston Dynamics has offered new detail on its redesigned humanoid hand, a four-fingered gripper with 13 degrees of freedom that is directly actuated and engineered for mass manufacturing. The company said the hand remains capable of handling objects over 100 lb., and that it was built for high-fidelity simulation to support sim-to-real reinforcement learning.

The design marks a visible departure from earlier hands: Boston Dynamics dropped the pinky finger and reduced the size of the gripper. In an interview with The Robot Report, Alberto Rodriguez, director of robot behavior for Atlas at Boston Dynamics, framed the changes as the product of a deliberate roadmap for Atlas, which forced hard decisions about the right level of complexity to pack into a hand and how to navigate tradeoffs among dexterity, strength, ruggedness, cost, reparability, and sensing.

The company said the hand remains capable of handling objects over 100 lb., and that it was built for high-fidelity simulation to support sim-to-real reinforcement learning.

Rodriguez said lessons from Atlas’ body design carried over to the hand. Simplicity, modularity, and actuator maturity may seem sensible in theory, he noted, but their impact on development velocity becomes clear over time. A reliable robot, he said, lets the team capture data faster, run more experiments, and take bets at early deployments, while modularity concentrates effort on fewer things that must be right. The new hand follows that philosophy: it is built from 13 identical actuators, all completely encapsulated, with no fragile cables crossing joints, which Rodriguez said is already paying off.

Strength was not the constraint. Rodriguez said earlier hands were strong enough to lift 100 lb.-plus heavy objects, so there was no need to increase strength, opening the door to considering a reduction. The redesign therefore leans on the existing strength envelope while pursuing other priorities such as manufacturability, simulation fidelity, and reinforcement-learning readiness.

Source: The Robot Report