Boston Dynamics Dropped Atlas's Pinky: Inside the New 13-DoF Four-Finger Hand

Boston Dynamics unveiled a new hand for its Atlas humanoid on 1 October 2026. It has four fingers and 13 degrees of freedom (DoF), up from seven in the previous design, and it has no pinky. All specifications below are company-stated, and no independent test has been published.

What is the new Atlas hand?

It is a redesigned hand that moves Atlas from grasping objects to manipulating them. Boston Dynamics says the older hand (called GR2 in a company video transcript) had 7 DoF, and the new one (GR3) has 13. In plain terms: the old hand could hold a wide range of things, while the new one can turn them over in the fingers.

The thumb has 4 DoF and the other three fingers have 3 each, which gives 4 + 3 + 3 + 3 = 13. The company describes the hand as directly actuated and designed for mass manufacturing. It is a companion to the production Atlas shown at CES in January 2026.

What is Boston Dynamics' thesis for robot hands?

The company treats the hand as a ruthless tradeoff, not a race to copy human anatomy. Its blog lists five design goals: many useful tasks including tool use, strength and ruggedness for industrial work, a small gap to human demonstration data, clean simulation for learning, and low-cost manufacturing and repair.

That is a reliability-first argument. In a Q&A published 9 October 2026, Alberto Rodriguez, director of robot behavior for Atlas, said a product roadmap for Atlas guided the tradeoffs among dexterity, strength, ruggedness, cost, reparability and sensing. He also said hands are so central to the humanoid pitch that building them in-house is the only realistic route.

Why did Boston Dynamics drop the pinky?

Because the team judged it not worth the cost. Dylan Thrush, a mechanical engineer on the Atlas team, said the extra dexterity was "not worth the extra complexity of three additional degrees of freedom", and cited size and power use too. Boston Dynamics adds that a pinky would bring three more actuators, more volume and another point of failure.

The decision came from a simple experiment. Zachary Jackowski, described by The Robot Report as Boston Dynamics' chief product and technology officer, asked staff to tape their pinky and ring finger together for a day and report what they could no longer do. The team then agreed on "no pinky". The company's blog says a pinky would not be hard to add later, given the modular design.

A two-thumb design was also considered. Rodriguez said it had early support because one assembly line could build both hands. It was rejected for extra actuation complexity, degrees of freedom, cost and volume.

What are the Atlas hand's specifications?

The company has published the architecture but not a full datasheet. The table marks every figure as company-stated and flags what is missing.

How does the Atlas hand work?

Each joint is driven by an actuator mounted at the joint itself, so there are no tendons or exposed cables. The company says the actuators are rigid-drive with a backdrivable transmission, and its controls compensate for cogging and friction. Backdrivable means the joint can be pushed by the outside world, which helps the robot sense contact.

That sensing works with the touch layer. The hand relies on proprioception, meaning the actuators' own feedback, plus pressure sensors on the fingertips and palm. In practice, it can notice small contact changes, such as a tool starting to slip.

The hand is built for training in simulation. The company says it trains with randomised motor torque profiles, surface friction, object shapes and disturbances, and reports "promising" early sim-to-real results on dynamic tasks, run on hardware using only high-rate actuator proprioception. It argues that gloves and handheld data-collection devices capture only proxy signals, so reinforcement learning in simulation is essential for dexterity beyond pick and place.

What can the Atlas hand do?

The company lists these motions:

• Sliding the thumb fingertip along the length and across the width of the other fingers.
• Pinch grasps between the thumb and any other finger.
• Tripodal grasps.
• Grasping triggered tools, including drills, power torque drivers, grinders, nail guns and welding torches.
• In-hand reorientation and recovery from a slipping grasp.

Rodriguez named handheld power tools, cable handling and fishing for screws as common industrial tasks, and said the team has also played with juggling and speed hand games. These are company claims; no third party has measured success rates.

How does it compare with other robot hands?

Hand makers disagree on how many joints a humanoid needs. The figures below are company-stated and relayed by trade press, and the dates differ, so they show design philosophy, not a benchmark.

The technology split is clear. Unitree describes a 20-DoF hand with 16 active and four passive joints, a fingertip repeat accuracy of +/- 1 mm and a weight of 1,000 g. Tesla says its hand has 22 DoF plus three in the forearm, with all actuators moved to the forearm. Boston Dynamics goes the other way: fewer joints, all in the hand, one actuator type. For a normal user, that means Atlas trades finger count for a hand that is meant to be cheaper to build and easier to repair.

Atlas's pitch is no cables crossing the joints.

Where will people actually use this?

Not at home, at least not yet. The target is industrial work. Engadget reports that Hyundai plans to use Atlas in its car plants from 2028 for tasks such as parts sequencing, with component assembly by 2030. Boston Dynamics' blog says tool use is "at the top of our list", and that many manual processes involve tools, cables or bin picking.

The production Atlas shown at CES has 56 degrees of freedom and human-scale hands with tactile sensing, according to TechCrunch, and, per Engadget, Boston Dynamics claims it can lift 110 pounds. Engadget reports the first deployments go to Hyundai and Google DeepMind.

What's still unclear?

Several things. First, no independent test of the hand has been published, so every capability above is the company's own claim. Second, the sim-to-real results are described only as "initial" and "promising", with no success rates.

Third, the hand has not been proven in a real plant. Asked whether it had been tested at Hyundai's Robotics Metaplant Application Center, Rodriguez said "Not yet", and said fleet-wide rollout needs coordination so it does not disrupt data collection. Fourth, weight, dimensions, sensor specifications, price and a ship date are all unpublished.

The weakest point of the argument is the pinky itself. A tape test measures what staff can do without a finger, not how a hand performs across thousands of hours of factory tasks. The company's own note that a pinky can be added later suggests the question is not closed.

Should you buy it or wait?

You cannot buy it. The hand is part of the Atlas program, which is aimed at industrial customers, and no standalone sale has been announced. For anyone following the field, the useful signal is the design bet: a 13-DoF hand that is simple to build and repair, trained mostly in simulation. Watch for deployment results from the Georgia site and any published datasheet.