Boston Dynamics Gives Atlas a 13-Motion Hand Built for AI Training
Boston Dynamics unveiled a new 13-DOF hand for Atlas, directly actuated and purpose-built for high-fidelity simulation and sim-to-real reinforcement learning.
- Boston Dynamics revealed new hands for Atlas with 13 degrees of freedom, up from 7.
- Four fingers only, no pinky, after the team tested taping their own pinkies down.
- Direct-drive, backdrivable actuators enable proprioceptive force sensing and impact resilience.
- Hand is explicitly engineered for high-fidelity simulation and sim-to-real reinforcement learning.
- Supports pinch grasps, tripodal grasps, in-hand reorientation, and triggered tool use (drills, grinders).
- Dense tactile sensors on fingertips and palm; similar strength to prior hand with a 100lb+ payload.
Boston Dynamics has unveiled a four-finger hand for its electric Atlas humanoid. According to the company’s technical post, the mechanism was designed for accurate simulation, allowing control policies trained in software to transfer more reliably to the physical robot. That focus shapes its finger count, actuation and sensors.
Thirteen motions expand the job
The previous Atlas hand had seven degrees of freedom and focused on grasping objects of different shapes. A degree of freedom is an independently controlled axis of motion. The new hand has 13: four in the thumb and three in each of the other fingers. The three fingers can also splay, giving the hand more control over an object after it has been grasped.
Independent thumb motion and finger splay support several behaviors required for industrial manipulation:
- Sliding the thumb across the length and width of another finger
- Forming a pinch grasp with any finger
- Using three-point grasps while reorienting an object
- Recovering when an object begins to slip
- Holding and triggering drills, torque drivers, grinders, nail guns and welding torches
Boston Dynamics describes the hand as approximately the size of a large human hand, with somewhat greater average strength. Atlas can carry a loaded minifridge weighing more than 100 pounds, although that figure describes the robot’s overall payload rather than fingertip capacity.
Why Atlas loses the pinky
During development, team members taped their ring and pinky fingers together for a day to assess how much capability a fifth robotic finger would add. They found that four fingers covered the target tasks, including in-hand reorientation, slip recovery and operating tool triggers.
Omitting the pinky saves three actuators along with their wiring, controls and structural volume. It also reduces cost and the number of components that can fail during repeated industrial use. The finished hand uses a single actuator design throughout, which further simplifies manufacturing, maintenance and simulation.
Simulation shaped every joint
Boston Dynamics optimized the mechanism for control policies trained with reinforcement learning in a physics simulator and then deployed on hardware, a workflow known as sim-to-real. A policy maps sensor readings to motor commands. Reinforcement learning improves that policy through trial and error against a defined reward, allowing risky or repetitive training to occur without damaging a physical hand.
Many humanoid teams collect demonstrations with instrumented gloves or handheld UMI devices. Boston Dynamics sees those proxy motion signals as useful for pretraining models on the intuitive physics of manipulation. Its approach uses simulated reinforcement learning to develop the fast, contact-rich control required for dynamic tasks.
The company links transfer performance to three mechanical and control choices:
- Direct joint actuation: No cables span multiple joints, so each commanded movement has a cleaner relationship to the resulting finger motion.
- Backdrivable transmissions: External forces can move a joint through its transmission. This allows the motors to estimate joint position and force, a capability called proprioception, while helping the mechanism yield during impacts.
- Compensated motor behavior: The controls account for friction and cogging, the uneven torque caused by magnetic interactions inside a motor. Modeling those effects narrows the gap between simulated and physical actuators.
For its early dynamic tasks, Boston Dynamics trained policies entirely in simulation with domain randomization, which varies properties such as friction and mass so the controller learns to tolerate modeling errors. The reported hardware rollouts use high-rate actuator proprioception as their feedback signal. Dense pressure sensors in the fingertips and palm can also detect small changes in contact.
Boston Dynamics describes the initial sim-to-real results as promising. The announcement provides no comparative success rates or standardized benchmark results, so the evidence currently consists of the company’s demonstrations and technical description.
Hardware joins the learning stack
The architecture treats mechanical predictability as a machine-learning requirement. Clean kinematics, consistent actuators and measurable joint forces make the simulator easier to calibrate, reducing the corrections required when a policy reaches the robot.
Teams selecting manipulation hardware can evaluate the design through several practical questions:
- Model fidelity: Can the simulator reproduce the hand’s joint motion, friction and contact behavior?
- Feedback rate: Can the controller detect and respond to slip or impact quickly enough?
- Serviceability: How many actuator types and transmission components must be stocked and maintained?
- Tool coverage: Can the hand hold, reorient and operate the tools required by the deployment?
- Transfer evidence: Do simulated policies retain their success rate, speed and stability on physical hardware?
Atlas currently reaches customers through pilot deployments, with broader sales timing, pricing and production-scale reliability data still undisclosed. The Atlas product page presents the hand as part of the full humanoid platform rather than a separately available component.
Comments
