Boston Dynamics Drops Atlas's Pinky and Targets 100,000 Hands a Year

Boston Dynamics unveiled a four-fingered Atlas hand with 13 degrees of freedom on Oct. 1, and the director of robot behavior for Atlas says the team is working out what must change to build 100,000 of them a year.

The hand has a thumb and three fingers, according to The Robot Report, which put the previous Atlas hand at seven degrees of freedom (the independent joints a hand can move). IEEE Spectrum, reporting the same day, quoted Alberto Rodriguez, director of robot behavior for Atlas, on the production target: "What we're working on right now is figuring out what needs to change in the fine details of our design if we want to make 100,000 of these hands a year."

The number that matters here is the part count. The hand is built from 13 identical actuators, each fully encapsulated, with no cables crossing a joint, according to a second Robot Report piece published Oct. 9. One actuator design repeated 13 times is a manufacturing decision before it is a robotics one.

What the 13-degree hand contains

The thumb carries four degrees of freedom and each of the three fingers carries three, which sums to 13. Boston Dynamics said the joints are directly actuated, with dense pressure-sensing tactile pads on the fingertips and palm, and that the hand is about the size of a large human hand.

Spec	Previous Atlas hand	New Atlas hand
Digits	Three fingers	Thumb plus three fingers
Degrees of freedom	7	13
Actuation	Not given	13 identical, encapsulated, no cables across joints
Load	100 lb or more, per Rodriguez	"Similar" strength, per The Robot Report

Strength did not go up. Rodriguez told The Robot Report that earlier hands could already lift objects of 100 pounds or more, and that strength might even be reduced slightly to make the actuators more transparent to the control software, meaning easier to simulate and to train by reinforcement learning.

The company gave simulation as the reason the hand was designed as it was, and it was specific about the method. Boston Dynamics reported "initial results" of policies trained in simulation, using domain randomization, running on hardware with only actuator proprioception as feedback. It gave no success rate and no task list for those results.

Why the pinky went: three actuators

Dylan Thrush, a mechanical engineer on the Atlas team, said the extra dexterity of a fifth finger "is not worth the extra complexity of three additional degrees of freedom." Chief product and technology officer Zachary Jackowski had the team tape their pinky and ring fingers together for a day, and the team concluded the hand did not need a pinky.

Three degrees of freedom is three actuators, so a pinky would have taken the hand from 13 to 16. Rodriguez also rejected a two-thumb design, saying it "involves extra actuation complexity that ultimately incurred extra DoFs, cost, and volume."

What 100,000 hands a year implies

The arithmetic is ours, not the company's. At 13 actuators a hand, 100,000 hands is 1.3 million actuators a year. At two hands a robot, it is 50,000 robots' worth of hands, before any spares.

Rodriguez framed the target against rivals. He told IEEE Spectrum that "many of the current designs are giving up on things like reliability and manufacturability." IEEE Spectrum's Evan Ackerman, its robotics editor, contrasted the Atlas hand, which "does look awfully clunky," with the humanlike hands of Clone, 1X, Unitree and Figure, and gave no degree-of-freedom counts for any of them.

Boston Dynamics has not said how many Atlas robots it will build, or when. The Robot Report separately reported that the company appointed former Amazon executive Rohit Prasad as chief executive; the piece is a podcast summary, and it gives no start date.

What has not been disclosed

Boston Dynamics gave no unit cost for the hand, no cycle-life or shift-life figure, and no weight. Both Robot Report pieces attribute the manufacturing claim to the company, which says the hand is "designed for mass manufacturing" and gives no volumes, timeline or cost. IEEE Spectrum's account contains no independent test data on reliability or cost.

The hands are also not in the field. Rodriguez said testing at the Robotics Metaplant Application Center, which opened a week before the hand was unveiled at Hyundai Motor Group Metaplant America in Georgia, has not started. The company has run the hand through handheld power tools, cable handling and fishing screws out of a pile, along with juggling, at its own headquarters.

What would change my read is a per-hand cost or a failure count from the Georgia site. Without either, 100,000 a year is a design goal, not a supply chain. Brendan Schulman, Boston Dynamics' vice president of policy and government relations, is due on a reshoring panel at RoboBusiness in Santa Clara on Oct. 20 and 21, the next public occasion the company has listed to be asked.

For how simulation-first training fits into the wider toolchain, see our report on AWS's open-source physical A.I. toolchain, and on the data side, Mecka AI's $60 million round for robot data.