Boston Dynamics Sacrifices Atlas’ Fifth Finger for Efficiency

WALTHAM, Mass. — Boston Dynamics has redesigned the hand on its Atlas humanoid based on the finding that a robot intended to do human work does not necessarily need five fingers.

The new Atlas hand has four fingers and 13 degrees of freedom, up from seven degrees of freedom in the previous three-finger design.
The Massachusetts-based manufacturer says the new architecture aims to let Atlas manipulate objects and use tools; it deliberately omitted a fifth (pinky) finger, deciding the extra actuators weren't worth the added size, power draw, cost and failure risk.
Engineers even taped their pinkies to their ring fingers for a day while considering the trade-off, according to the company. Alberto Rodriguez, director of robot behavior at Boston Dynamics, told IEEE Spectrum that robotic hands have to be balanced for capability, reliability and manufacturability.

The decision differs from the increasingly humanlike five-finger designs being developed for some competing humanoids. Boston Dynamics' approach is to keep enough similarity to a human hand, to make human demonstration data useful and remove parts it deems unnecessary for the industrial tasks.

The new hand has a four-DOF opposable thumb, while each of the other three fingers has three degrees of freedom. The fingers can spread apart, allowing Atlas to form pinch and three-finger grasps and, according to Boston Dynamics, hold tools while operating their triggers. The company intends drills, grinders, nail guns and welding torches among the applications.

Boston Dynamics has also placed actuators at the joints instead of using tendons or cables. Pressure sensors cover the fingertips and palm, while backdrivable transmissions allow forces acting on the fingers to be sensed through the actuators.

Designing the Hand for Simulation

Simulation influenced choices such as joint placement and sensor layout. Boston Dynamics designed the hand so its kinematics and dynamics can be modeled closely enough for reinforcement-learning policies trained in simulation to transfer to the physical robot.
The company has gotten preliminary results indicating there is sim‑to‑real transfer in dynamic manipulation tasks, but it has not published comparative performance data establishing how those results measure against other dexterous hands.

That requirement creates a different engineering problem from simply maximizing the number of joints. A mechanically complicated hand can offer additional theoretical dexterity but it is harder to simulate, manufacture, and repair, especially in industrial settings where the hand must be protected from damage.
The four-finger design is also part of Boston Dynamics' effort to turn Atlas from research hardware into a machine that can eventually be produced at much larger volumes. Rodriguez told IEEE Spectrum the company is investigating what design changes would be needed to produce 100,000 hands a year. That figure describes a theoretical design study, not current output or a production target.
Independent testing hasn't shown four fingers outperform five. Boston Dynamics says that for Atlas's expected manipulation and tool-use tasks, the fifth finger doesn't add enough to justify the extra hardware. The new hand still needs to prove that balance between performance and manufacturability in continuous commercial factory use.
The new hand has yet to establish that trade-off under sustained commercial factory use.