On a factory assembly line, a human holds a power drill only one way: thumb on the trigger, four fingers wrapped around the grip. Swapping a drill bit means setting the tool down, freeing one hand, and re-gripping. Could a robot break this compound motion apart? Boston Dynamics answered on October 1 by fitting Atlas with a hand that has only four fingers — the pinky was deliberately cut.
✋ Five Fingers vs. Four
The human hand has five fingers. Atlas's new hand has four and 13 degrees of freedom: 4 in the thumb, 3 in each of the other three fingers, capable of both pinch grips and three-finger grasps. The previous-generation Atlas hand had 7 degrees of freedom and was positioned for stable grasping and carrying.
The difference is not whether the hand can pick things up, but whether it can change their orientation. In the official demo, Atlas picks up small objects and flips them in its palm, rolls two golf balls simultaneously in one hand, and recovers objects from slipping grasps.
The engineers put it plainly: the old hand was designed to grasp various objects; the new hand is designed to manipulate them.
🔢 Why Cut the Pinky
The team laid out the cost openly: the marginal dexterity and additional tasks a pinky would enable are not worth three extra degrees of freedom, plus the added size and power consumption. Three extra DOF means three more actuators, another drive system, and another potential failure point — a poor trade for a machine meant to be maintained and operated repeatedly on a factory floor.
This is not biomimetic laziness. Finger count is only a means; reliability, consumable cost, and maintainability are the actual optimization targets. Humans have five fingers, but that is irrelevant to this production line.
Before cutting, the team ran experiments: they bound the ring and pinky fingers together and confirmed four fingers covered the expected tasks — pinch, three-finger grasp, and in-palm reorientation. That experiment was the cost-basis for eliminating those three degrees of freedom.
🔩 Direct Drive and Backdrivability: Mechanical Margin Sacrificed for Simulation
All joints use the same actuator type, fully enclosed, with no fragile cross-joint cabling. Two direct benefits: simpler transmission paths and backdrivable joints. Backdrivability means external forces can move the joints rather than being blocked by rigid, high-friction mechanisms. This pays off twice: the robot can estimate its own joint positions and motion via proprioception, and it can yield when contact goes differently than expected instead of pushing back.
The cost is equally clear: a backdrivable joint is less "stiff," with output torque capped by the motor's own torque limit.
The design core lies in simulation. Boston Dynamics says their control compensates for friction and cogging torque, keeping the hand's dynamics highly faithful in simulation. Policies are trained in simulation via reinforcement learning with domain randomization, then deployed on hardware relying only on high-frequency proprioception from the actuators.
The team calls initial sim-to-real results "optimistic" — a claim worth weighing precisely: no trial counts, success rates, or confidence intervals were given, and no third party has replicated the results. Simulation can support a training methodology; it cannot establish routine reliability in a factory.
🛠️ Tool List and Touch
The tool list is the hand's target list: drills, electric torque guns, grinders, nail guns, torches — all sharing a trigger that must be held continuously and a grip that must be held steadily.
Boston Dynamics says the thumb tip can slide along the length and width of the other fingers, letting tools rotate in the palm without dropping.
Touch is the other main thread. Fingertips and palm are covered with dense pressure sensors; the controller adjusts grip force based on measured contact. These sensors complement proprioception: one tells you where the hand is, the other what is happening at the contact surface. Combining both is what lets the robot decide different grip forces for an egg versus a steel plate.
Atlas's platform payload specs are 110 lb instantaneous / 66 lb continuous, and the new hand can hold a loaded mini-fridge over 100 lb. These are reported capabilities, not safety guarantees for arbitrary factory environments.
🏭 From Demo to Production Line: A Gap Remains
| Dimension | Previous hand | New GR3 hand | |---|---|---| | Fingers | Three-finger gripper | Four | | Degrees of freedom | 7 | 13 | | Target | Stable grasp & carry | Fine manipulation & tool use | | Thumb | Non-opposable | Opposable, 4 DOF | | Transmission | Not stated | Single-type direct drive, backdrivable | | Cross-joint cabling | Not stated | None; actuators fully enclosed | | Touch | Not stated | Dense fingertip and palm pressure sensing | | Size | Not stated | Close to a large human hand | | Tool operation | Not stated | Drill, torque gun, grinder, nail gun, torch |
Boston Dynamics has recently established a robotics MegaFactory Application Center, simulating Hyundai Motor Group's actual production lines and collecting on-site Atlas data for repeated training and validation. This step is the real dividing line of the story: up to the announcement, a demo is a demo; the application center turns "doing the same thing repeatedly on a line" into a measurable process. The company's dexterous hands have moved from CES 2026's three-finger gripper to a four-finger, 13-DOF hand along the same path: pushing a demonstrable capability toward a maintainable device.
Also worth noting that same week: at IROS 2026, Agility unveiled Digit 5 and reported 65,000+ cumulative runtime hours for Digit 4 across North American customers including Amazon, Schaeffler, and Toyota Canada. Sanctuary AI reported 99.5%+ success rate and a 2.54-second cycle time on a complex wire-insertion benchmark at a Tier-1 automotive supplier. These figures are deployment-side metrics, categorically different from demo videos.
🧭 How to Read the Number 13
The jump from 7 to 13 degrees of freedom is the most quotable — and most easily misread — number here. Joint count alone guarantees nothing. Sensing quality and mechanical compliance determine how much useful contact these motions produce; control quality determines whether it can be held stable.
Four capabilities are the real acceptance criteria: in-palm reorientation, pinch grasp, three-finger grasp, and trigger operation. Whether four fingers are enough depends on whether these four can be repeated under production-line conditions, not on how anatomically important the missing finger is.
The next concrete milestone: repeat operation counts, per-cycle force-control precision, and maintenance intervals on a real production line. Whoever publishes those three numbers first gets the first verifiable data point for dexterous hands moving from demo to mass production.
---
Sources and Caveats
- Boston Dynamics release material, October 1, 2026; coverage by heraldcorp, roboticfirms, science.report; Hong Kong Economic Journal and RoboRadar.
- Quotes from Alberto Rodriguez (Director of Robot Behavior) and Dylan Thrush (Mechanical Engineer) on the Atlas team.
- Sim-to-real results are the team's self-described optimism, with no trial counts, success rates, or confidence intervals, and no third-party replication.
- Payloads, fingertip sliding, and similar capabilities are reported values, not safety guarantees for arbitrary factory environments.
- Comparative figures (Agility's 65,000 hours, Sanctuary's 99.5% and 2.54 s) come from the companies' own statements; original measurement methods were not independently verified.