Boston Dynamics drops the pinky: The new Atlas hand has four fingers and 13 degrees of freedom

Boston Dynamics unveiled the new hand of the humanoid Atlas on October 1st, and it has four fingers. Thirteen degrees of freedom instead of seven, four of them in the thumb, pressure sensors across fingertips and palm, a load capacity of over 45 kilos. The pinky finger has been eliminated. The decision was made with tape.

The competition is adding joints—Unitree sells a hand with 22, Tesla also names 22 for Optimus. Boston Dynamics is going in the other direction because each additional finger costs three actuators and increases the number of parts that can break. That's where the problem is: five-fingered hands with around 24 degrees of freedom cost over 70,000 dollars according to Fraunhofer IPA.

Zachary Jackowski, Chief Product and Technology Officer at Boston Dynamics, had his team spend an entire working day with their pinky and ring fingers taped together. By evening, everyone agreed that little was missing in everyday tasks. Dylan Thrush, a mechanical engineer on the project, sums up the result dryly: the additional dexterity does not justify the complexity of three additional degrees of freedom.

Alberto Rodriguez, responsible for robot behavior at Boston Dynamics, describes the development as a mix of intuition and experimentation. One thumb or two? Do you even need a pinky finger? Such questions cannot be decided at a desk because every answer works against another. More fingers means more skill and at the same time more weight, more cables, more failures.

Technically, the hand is built remarkably stubbornly. All joints are driven directly by a single actuator type, the gearboxes are reversible, so the hand feels through the motors themselves what it grasps. The thumb has four degrees of freedom and is arranged in a human-like way, the other three fingers each have three and can be spread. There are closely spaced pressure sensors at the fingertips and in the palm. This allows Atlas to regrip objects in its hand, catch slipping parts, and operate tools with triggers—power drills, torque wrenches, grinders, nailers, welding torches.

The size is deliberately chosen, roughly that of a large man's hand. This is not a matter of taste but a training decision. The less a robot hand differs from a human hand, the more directly recorded human hand movements can be transferred to the machine.

And here things become contradictory. If you want to use human movement data, you record it from five-fingered hands. Boston Dynamics argues for closeness to the human hand and drops a finger in the same breath. The company sidesteps the contradiction by training finger dexterity in the simulator, with randomly varied motor torques, friction values, object shapes and disturbances, and then transfers the learned strategy solely via joint sensors to the hardware. Whether this works equally well on a factory floor with oil, dust, and poorly stacked boxes is not mentioned in any blog post.

The cost calculation behind this is the real story of the industry. Actuators account for around half the manufacturing costs of a humanoid, according to calculations by the Fraunhofer Institute for Production Technology and Automation together with consultant P3 in its study on the value chain of humanoid hardware. A classical industrial robot gets by with a maximum of seven axes. A humanoid has over forty individually driven joints, and each one is a package of motor, gearbox, bearing and sensor. Three fewer actuators per hand are no detail in this calculation.

Sebastian Reitelshöfer, head of the robotics research area at the University of Erlangen-Nuremberg, is skeptical about the cheapest components. In continuous operation, such parts would last one or two weeks. Anyone offering a hand with 24 joints for a few thousand dollars is therefore not selling the same product as Boston Dynamics, even if the datasheets look similar.

For Switzerland, the dispute over the fifth finger is primarily a question of supply. Maxon in Sachseln builds exactly the building blocks in question. The HEJ 70 joint delivers 50 Newton meters of peak torque at little more than one kilogram of weight, runs over EtherCAT and is IP67 protected, the larger HEJ 90 reaches 140 Newton meters. Whoever sells actuators and gearboxes profits regardless of whether the industry ultimately settles on four or five fingers. ETH Zurich showed this year how far basic research here reaches, with a robot hand that caught itself in 21 of 25 falls.

A week before the hand's unveiling, the Hyundai Motor Group opened its Application Center for Robotics in Georgia, where Atlas learns work steps from automotive production. The hand was built for this hall and not for a stage. This also explains the tool list, because nobody needs nailers and welding torches for a demonstration in the living room.

What Boston Dynamics doesn't say is everything that would count for a purchase. No price, no operating life, no timeline for series production, no information on how long a replacement takes if a finger breaks on the night shift. If you're evaluating a humanoid in Switzerland, you should ask early about hand configuration and who supplies the replacement. A broken finger on the factory floor is not a blemish but a standstill.