Boston Dynamics reveals details of Atlas's new hand in an interview with The Robot Report

The Robot Report has released an interview with Alberto Rodriguez, director of behavior for the Atlas robot at Boston Dynamics. The conversation is about the humanoid's new hand. The published announcement contains no summary of the interview, only the fact that it took place and a link to the piece. So we have not invented any specifics about the design, the number of fingers or the actuators; read the original for those. What we can do is look at why the company returned to the hand at all, and what is known about Atlas around it.

Why the hand is a humanoid's bottleneck

Every humanoid runs into a limit that has little to do with walking: manipulation. Atlas learned to walk and jump a long time ago, and its hydraulic-era parkour and flips became Boston Dynamics' calling card. But on a factory floor, money is made not by acrobatics but by the ability to pick up a part, rotate it, insert it into a slot and not drop it. That takes a hand that is durable, sensitive and simple enough to survive thousands of cycles per shift.

Hence the perennial dilemma. More fingers and joints give more dexterity, but they also add mass, cost, failure points and control complexity. Fewer make the hand more reliable, but the robot copes worse with unusual objects. An interview with the person responsible for the robot's behavior, rather than its mechanics, is interesting for exactly this reason: the hand is designed around what a neural-network control policy can do with it.

What is known about Atlas around this hand

At CES 2026, Boston Dynamics showed the production electric version of Atlas. According to the company, the key specifications are:

• 56 degrees of freedom across the whole robot;
• arm reach of up to 2.3 m;
• lifting capacity of up to 50 kg;
• operating temperature range from −20 to +40 °C;
• tactile sensors in the hands.

The partner ecosystem matters too. Hyundai, which owns Boston Dynamics, plans to use Atlas in its own plants and is building dedicated capacity for humanoids. Google DeepMind works with the company on control models, and those models determine what the robot does with an object in its hand. The hand serves as the interface between the model and the physical world: the richer the signal from the tactile sensors, the better grasping policies train.

What to look for in the full interview

The announcement contains no standalone figures for the hand itself, so here are the questions to keep in mind when reading the original. How many fingers does it have, and which are opposable? How are the actuators built and where are they placed: in the palm or in the forearm? How repairable is the hand in the field? How does data from the tactile sensors get into training? And most importantly: was the hand designed for specific operations at a Hyundai plant, or for the general case?

For engineers and integrators, the practical takeaway is simple. If you are evaluating humanoids for a warehouse or assembly line, do not compare only walking speed and demo videos. Ask the vendor about hand lifespan in cycles, finger replacement time, the range of shapes the hand can grasp and the data the grasps were trained on. For the market as a whole, the fact itself is telling: Boston Dynamics is discussing the hand publicly through its director of behavior, which means it treats the hand as part of the AI system, not just hardware. That is most likely where it will be decided whether humanoids become a working tool or remain a showpiece.