New Hands for Atlas | Boston Dynamics

TL;DR

Boston Dynamics unveils a redesigned four-finger hand for the electric Atlas humanoid, moving from 7 DoF to 13 DoF (four in the thumb, three in each of the other three fingers, plus finger splay). The hand drops the pinky — the team found tape-your-pinky-down ablations on their own hands produced no meaningful task-loss — and uses direct-drive, backdrivable actuators to get proprioceptive force sensing and impact tolerance, with dense tactile sensors on fingertips and palm and ~100 lb payload matching the prior hand. The explicit design target is high-fidelity simulation + sim-to-real reinforcement learning: finger count, actuation choice, and sensor layout were picked to make the hand faithful inside a simulator so RL policies trained in sim transfer to the physical robot. No transcript or paper is attached to the video; details come from the Boston Dynamics technical post summarized by secondary coverage.

Key claims

The new Atlas hand has 13 DoF, up from 7 on the previous generation — four in the thumb and three in each of the three remaining fingers, with the three non-thumb fingers also able to splay [video description + accompanying technical post, as summarized in secondary coverage].
The design is four-fingered (no pinky); the team tested the omission by taping down their own pinkies before committing [secondary summary of BD technical post].
Actuators are direct-drive and backdrivable, enabling proprioceptive force sensing and impact resilience [secondary summary of BD technical post].
The hand is explicitly engineered for high-fidelity simulation and sim-to-real RL — simulator fidelity drove the mechanical design, not the other way around [secondary summary of BD technical post].
Dense tactile sensors are placed on both fingertips and palm, with strength roughly matched to the prior hand (~100 lb payload class) [secondary summary of BD technical post].
Capabilities demonstrated or claimed in the video include pinch grasps with any finger, three-point / tripodal grasps with in-hand reorientation, sliding the thumb across the length and width of other fingers, and triggered tool use (drills, grinders) [video + secondary summary].

Method

Not disclosed beyond the video and the brief technical post summary. The video itself has no transcript available via the fetcher; the design rationale above is sourced from BD’s accompanying post as reported in secondary coverage. What the announcement does commit to as design logic: (1) finger count is chosen to match the subspace of human hand motions that humans actually use for the target workplace tasks (hence pinky removal after self-ablation); (2) actuation is direct-drive to make the dynamic model inside the simulator a close match for the physical plant, so sim-trained RL policies transfer; (3) tactile placement spans fingertips and palm rather than fingertips alone, enabling in-hand reorientation feedback that fingertip-only sensing cannot provide.

Results

No quantitative results. Published artifact is a product-reveal video plus a short technical post; no success rates, no sim-to-real transfer numbers, no in-hand manipulation benchmarks, no comparison against the prior 7-DoF Atlas hand. The video demonstrates (not measures) pinch grasps, tripodal grasps, in-hand reorientation, and tool triggering on drills/grinders. Payload is reported as approximately matched to the prior hand’s ~100 lb class.

Why it’s interesting

Boston Dynamics is anchoring the sim-first-tactile-integration bet with real hardware at the top of the humanoid platform market. The explicit “designed for high-fidelity simulation + sim-to-real RL” framing lands in the same design space the Tactile sensing for manipulation concept has been tracking from the simulator-tooling side: Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks argued placement dominates sensor type and that whole-hand coverage (fingertip + palm + proximal phalanges) beats fingertip-only — Atlas’s fingertip-and-palm layout is the hardware instance of that recipe, and the first filed instance where the payload-class industrial platform ships with it. ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning ADEPT provided the sim-trained-RL-transfers-zero-shot existence proof on a 29-DoF Flexiv-Sharpa hand; Atlas pushes the same recipe onto a bipedal humanoid platform in the field.

On the hand-hardware axis the concept also tracks, Atlas’s 13-DoF, direct-drive, backdrivable, four-finger design stakes a distinct position from the three filed anthropomorphic references: Unitree Dex5-S — 22-DoF 1:1 anthropomorphic dexterous hand with backdrivable joints, from $6.5K Unitree Dex5-S (22 DoF, 1:1 size, five-finger, backdrivable, $6.5K starting), Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic hand M1 (15 active DoF, 21 joints, tendon-driven, five-finger, backdrive via tendons), and ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning ORCA (17 DoF, five-finger, open-source, fingertip FSR). Atlas is the lowest-DoF of the group and the only direct-drive-in-hand design — a deliberate trade of per-finger articulation for actuator fidelity inside the simulator, with pinky removal paying back the DoF budget. The load-bearing open question the announcement does not answer: whether the backdrivable actuators are read out as an implicit per-joint force channel (the Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic recipe), or whether force sensing is deferred entirely to the explicit palm+fingertip taxel array.