Summary
Boston Dynamics has detailed a four-finger, 13-degree-of-freedom hand for its Atlas humanoid, designed for dexterous manipulation and tool use. The company says early tests show sim-to-real transfer for dynamic tasks.
Boston Dynamics has detailed a new four-finger hand for its Atlas humanoid, increasing the design from seven degrees of freedom (DOFs) in the previous hand to 13. The company says the hand is built for dexterous manipulation—including using tools—and for training control policies in simulation before running them on the physical robot.
A degree of freedom is an independently controllable movement. The new hand has a four-DOF opposable thumb and three DOFs in each of its other three fingers. Boston Dynamics says this layout supports fingertip movements across the fingers, dexterous pinches and three-finger grasps, as well as holding tools while pressing their triggers.
Designed for manipulation and contact
The hand uses direct actuation at the joints, with a backdrivable transmission. In a backdrivable mechanism, an external force can move the joint rather than being resisted entirely by the drive system. Boston Dynamics says this helps the hand respond to contact, sense forces through actuator feedback and move out of the way during impacts. The company also describes controls designed to compensate for friction and cogging—the unevenness in motor motion that can make precise control harder.
The hand has pressure sensors across its fingertips and palm to detect contact. Together, actuator feedback and tactile sensing give the robot information about both joint motion and where pressure is being applied. That feedback is important for manipulating objects whose position or resistance can change during a task.
The design has four fingers rather than five. Boston Dynamics says its team tested the trade-off and concluded that an additional finger’s dexterity did not justify the added actuators, size, power use and complexity. The company’s stated goal is a hand that can perform useful work while remaining practical to manufacture and maintain, rather than a replica of a human hand in every detail.
Simulation as part of the hand’s design
Boston Dynamics says the hand was designed to be simulated with enough fidelity to support reinforcement learning (RL). In this approach, a robot learns control policies by practising in a simulated environment. The company says its training randomizes factors such as motor torque, surface friction, object geometry and task disturbances, exposing policies to varied conditions.
The company reports initial sim-to-real results: behaviours trained in simulation have been run on hardware, using high-rate actuator feedback. It describes these as dynamic tasks and points to object reorientation as representative of capabilities relevant to tool use and assembly. Boston Dynamics says it is continuing work to scale this training process.
The company’s broader case for dexterity is that many industrial robots rely on carefully arranged parts and specialised grippers. A more capable hand could help a humanoid handle objects in less controlled settings and perform manual tasks involving tools, cables or parts presented in different ways. That is the rationale behind the design; the published results are an early step toward those applications, rather than a demonstration of general-purpose work across industrial settings.