Table of Contents
Boston Dynamics has unveiled a new generation of hands for its electric Atlas humanoid, moving the robot from a 7-degree-of-freedom research gripper to a 13-degree-of-freedom design aimed at real factory work.
The new hand, called GR3, has four fingers, direct joint actuation, a more capable opposable thumb, and tactile sensing across the fingertips and palm. The design is meant to support dexterous behaviors such as pinch grasps, in-hand object reorientation, and tool use while remaining rugged enough to manufacture, repair, and deploy at scale.
Why Boston Dynamics Changed Atlas’ Hand
The redesign was a tradeoff between dexterity, durability, cost, strength, sensing, and manufacturability. The company’s earlier GR2 hand used three fingers and 7 degrees of freedom. The new GR3 hand adds more actuation and a much more capable thumb, bringing the system to 13 degrees of freedom across four fingers.
Each of the three non-thumb fingers has 3 degrees of freedom, while the thumb has 4 degrees of freedom. The added thumb motion allows more human-like opposition, which is central to stable grasping and tool handling. The hand is also sized close to a large human hand, a choice Boston Dynamics said helps reduce the gap between human demonstration data and robot execution.
That does not mean the company is trying to copy a human hand in every detail. Boston Dynamics removed the pinky after internal testing suggested the extra actuators, volume, cost, and failure points were not worth the expected benefit for the tasks Atlas is being built to perform.
The resulting hand is less anatomically faithful than a five-finger replica, but more focused on the company’s priorities: industrial usefulness, robustness, and scalable production.
Built Around Direct Drive, Tactile Sensing, and Sim-to-Real Training
A key design choice is direct actuation. The hand uses fully encapsulated, backdrivable actuators without fragile cables crossing the joints. That architecture is meant to make the hardware more durable and easier to repair, while also improving the quality of force feedback available through the joints.
The hand also includes pressure tactile sensors on the fingertips and palm. That combination of tactile sensing and proprioceptive force feedback gives Atlas more information about contact, pressure, and object state during manipulation.
Boston Dynamics is also designing the hand for high-fidelity simulation. Transparent actuation and controls work are intended to make the hand’s dynamics easier to model, allowing reinforcement learning policies to be trained in simulation and transferred to physical hardware.
That sim-to-real approach has become central to modern robotics. Instead of manually programming every contact-rich behavior, teams can train policies across many randomized simulation conditions and then test whether those policies survive the messier dynamics of real hardware.
Wearable human-demonstration data is useful but not sufficient for agile dexterous control. Hands will need the same kind of reinforcement-learning-driven progress that has already helped whole-body humanoid controllers improve balance, recovery, and dynamic movement.
Tool Use Is the Real Target
The new hand is not being designed only for clean pick-and-place demos. Boston Dynamics specifically pointed to triggered tool grasps involving drills, torque drivers, grinders, nail guns, and welding torches.
The hand is also intended to be strong enough for the broader Atlas platform’s payload ambitions; the need to support heavy manipulation, including examples such as carrying a loaded minifridge, while still preserving enough dexterity for smaller and more precise tasks.
That balance is difficult. More actuators can improve dexterity, but they also add weight, cost, complexity, and potential failure points. More rugged hardware can survive deployment, but may reduce sensitivity or fine control. Boston Dynamics’ GR3 hand represents the company’s latest answer to that engineering compromise.
