Boston Dynamics' New Atlas Hand Has 13 Actuators and a 100,000-a-Year Target, but No Cost Figure
Hardware / analysis
Boston Dynamics' New Atlas Hand Has 13 Actuators and a 100,000-a-Year Target, but No Cost Figure
The company's Oct. 1 announcement dropped the pinky and published a 100-pound carry claim, while leaving out the cost, durability and unit numbers that decide whether a hand can be mass-produced.

Boston Dynamics' new Atlas hand has 13 degrees of freedom, up from seven, and is built from 13 identical actuators. The company has published no cost, durability or unit figures for it.
Those three numbers decide whether a humanoid hand can be made at volume, and the October announcement does not contain them. What it does contain is a design argument, laid out in the company's blog post "Robot Hands for Modern AI and Real Work", and a production target quoted by IEEE Spectrum. The number that matters here is not the degrees of freedom but the actuator count, because each actuator is a part that has to be bought, assembled and repaired.

What Boston Dynamics published on Oct. 1
The hand has a four-degree-of-freedom thumb and three fingers of three degrees each, which sums to 13, according to The Robot Report. There is no pinky. Every joint uses the same type of actuator, directly driving the joint, fully enclosed, with no cable crossing a joint. Dense pressure sensors cover the fingertips and palm.
The previous Atlas hand had seven degrees of freedom and was built to grasp. The company says this one is built to manipulate: sliding the thumb along the other fingers, tripod grasps, and holding a drill while pressing its trigger.
IEEE Spectrum's headline, "Atlas Robot's New Hand May Outperform Humanlike Designs", is a hypothesis rather than a result. Evan Ackerman's piece, dated Oct. 1, 2026, reports no strength, cost or durability data and no count of units built or deployed.
The pinky was three actuators out of 16
Boston Dynamics said it tested a pinky and left it off to avoid three extra actuators and their cost, volume and failure risk. Because the hand uses one actuator per degree of freedom, a pinky version would have needed 16 actuators; dropping it removes about 19 percent of them (the percentage is The Terminal's arithmetic, not the company's).
IEEE Spectrum reports that Alberto Rodriguez, director of robot behavior at Boston Dynamics, said the team reached the decision after spending a day with their pinkies taped to their ring fingers. Rodriguez told The Robot Report that a second thumb was also rejected, because it added actuation complexity, degrees of freedom, cost and volume.
- Previous Atlas hand7 DOF
- New Atlas hand13 DOF
Source: Boston Dynamics blog post, accessed 2026-10-21
That is the trade the company is selling. A hand with fewer than 16 actuators is cheaper and easier to repair than one with 16, but it gives up the fifth finger's reach and the spread of motion a human pinky provides.
100,000 hands a year is 1.3 million actuators
IEEE Spectrum quotes Rodriguez on production: the team is working out "figuring out what needs to change in the fine details of our design" to make 100,000 hands a year. At 13 actuators per hand, that rate is 1.3 million actuators a year. If each robot carries two hands, it corresponds to 50,000 robots a year (the two-hand figure is an assumption, not a company number).
The volume is a target, not a capacity. The Robot Report's second article, published Oct. 9, says the hands have been used extensively at Boston Dynamics' headquarters and have not yet been deployed to the full robot fleet, which requires coordination so that data collection and other operations are not disrupted.
A week before the hand was announced, the company opened a Robotics Metaplant Application Center at Hyundai Motor Group's Metaplant America in Georgia, where staff are trained to integrate Atlas into automotive manufacturing, per The Robot Report. The hand announcement does not say when the new hand reaches that plant.
The strength claim has a baseline but no test
The blog post says the new hand's strength is similar to the previous hand's, which it describes as able to carry a loaded minifridge weighing more than 100 pounds. Rodriguez told The Robot Report that the hand can still handle more than 100 pounds and that strength may be reduced slightly to improve how well the actuators sense force.
That is a vendor-supplied figure with no test method given, duty cycle or number of cycles. A hand that lifts 100 pounds once is a different part from one that lifts it 10,000 times, and the announcement does not say which this is.
The same gap between a maker's number and an outside measurement appears in DeepSeek's Terminal-Bench claim, which Vals AI measured at 31.8 against the lab's 42.3. Robotics has no Vals AI equivalent for hand durability yet.
What the announcement leaves out
| Item | Published? | Where checked |
|---|---|---|
| Degrees of freedom (13) | Yes | Boston Dynamics blog |
| Strength (similar to prior hand) | Vendor claim | Boston Dynamics blog |
| Cost per hand | No | IEEE Spectrum, The Robot Report |
| Durability or cycle counts | No | IEEE Spectrum, The Robot Report |
| Units built or deployed | No | IEEE Spectrum, The Robot Report |
| Production timeline | No | The Robot Report |
Speed figures are also absent from the blog post, and so is any statement of which demonstration behaviours were autonomous and which were remotely operated. Robot makers with humanlike hands have shown more dexterous demonstrations, as Ackerman acknowledges, and he argues those hands tend to be fragile or expensive. A force-sensing robot model such as Eka Robotics' $150 million round shows how much of the value sits in software that needs a hand to be reliable first.
What would change this read
A published cost per hand, even a range, would settle the actuator argument. So would a cycle count at rated load, or a number of hands running in a customer plant rather than at the company's headquarters. Boston Dynamics has given none of the three, and the next dated checkpoint is the full-fleet rollout the company has described but not scheduled.
Sources
More in Hardware
- 01Vulcan Fails 10 Minutes Into Flight, Losing 40 Amazon Leo SatellitesULA has not named the failed stage, and the rocket had two earlier booster nozzle problems, in 2024 and February 2026.
- 02Agility Raised $465 Million Against $174,000 in Half-Year RevenueThe humanoid maker listed on Nasdaq as AGLT on Nov. 5. Two reports give revenue for different periods, and neither gives a fleet count.
- 03SpaceX Catches Starship Ship 42 on Flight 15 but Loses Booster 22The tower caught the ship after 14 orbits and 26 Starlink V3 deployments. The booster aborted its own catch, and SpaceX's commentator ruled out guidance as the cause.
- 04Seagate FireCuda Forge Costs $699 and Leaves Out the SSDThe 8TB and 20TB external drives cache hot files on an M.2 SSD the buyer supplies, and Seagate and Blocks & Files quote different read speeds.