Boston Dynamics Atlas Humanoid Robot Explained

An electric robot built around handling work

Atlas uses electric joints, cameras, touch sensing and coordinated body control to move objects while standing or walking. Boston Dynamics lists material handling as its initial commercial purpose. Understanding its capabilities requires separating the current product specification from earlier research robots and from individual experiments with different hands or software. [1]

The Atlas name covers several machines

Boston Dynamics' earlier hydraulic Atlas was a research platform. Its jumps and parkour belong to that hardware generation. An all-electric Atlas appeared in April 2024. The company describes factory and laboratory work during 2025, including a Hyundai trial in Georgia, before presenting the product version in January 2026. [2]

At the 5 January 2026 CES announcement, the newly revealed product and the prototype performing on stage were distinct. Boston Dynamics announced manufacturing of the product version and scheduled fleets for Hyundai and Google DeepMind. Those shipment plans cannot serve as evidence of completed work at either destination. At the January launch, Boston Dynamics said its 2026 deployments were fully committed and planned to add customers in early 2027. The current product page invites qualified prospects to discuss an application and does not publish a price. [3] [1]

Published product specifications

The official PDF in the January 2026 upload directory carries a 23 December 2025 revision date. These are manufacturer ratings, checked on 8 October 2026. [4]

[4]

A load held farther from the torso produces more turning force at an arm joint. Keep the instant, sustained and one-handed ratings separate when specifying a grip and carrying posture.

The hands have changed too

Boston Dynamics now describes a GR3 hand with four fingers and 13 directly actuated degrees of freedom. The thumb has four; each other finger has three. Pressure sensors cover the fingertips and palm. The preceding GR2 hand had seven degrees of freedom. This newer hand account does not supply a revised whole-robot count, so its numbers must remain separate from the product table. [5]

Direct actuation places controlled motors at the joints. Tactile measurements report contact that a camera can miss behind a finger. Together, these signals can help a controller adjust a grasp while the object remains in the hand. Motor count alone cannot establish grip reliability on oily metal, flexible packaging or parts of unfamiliar shape. [5]

A documented engine-cover task

In its earlier electric Atlas engine-cover example, Boston Dynamics supplied a list of source and destination bin locations. A vision model located fixtures and bins, while a grasp policy handled the parts. The company states that movements were generated online without teleoperation. Its account also identifies insertion errors, trips and contact with the surroundings, with recovery using vision, force and robot-state measurements. [6]

This task shows the difference between finding a bin and completing an insertion. The robot must keep estimating the part's position after pickup. Contact with a fixture can change that position. The published account does not provide an attempt count or a full-shift success rate. [6]

The coordinate and contact calculations behind this sequence are developed in how humanoids locate and grasp objects.

Manipulation changes the balance problem

The Boston Dynamics and Toyota Research Institute Large Behavior Model project describes an earlier 50-DoF Atlas with seven-DoF grippers. A pair of HDR stereo cameras in the head supplies images. Images, robot-state measurements and language prompts feed a policy that produces action chunks controlling the robot at 30 Hz. Its data-collection system combines a VR interface with model predictive control, which predicts motion under physical constraints. The project includes stepping, squatting and handling Spot components with one policy. [7]

These research details apply to the stated platform. They do not establish the complete software installed on every 2026 product. Physically, moving a heavy part also moves the combined center of mass. The legs and torso must account for the load while the arms follow their target. A correct hand path can still fail when foot contact or balance is lost.

For the operator's role in collecting those examples, see how humanoid teleoperation maps human commands onto a robot.

What the refrigerator experiment establishes

A separate Boston Dynamics experiment used an animated reference and reinforcement learning to teach a refrigerator carry. Training varied load, floor friction and motor strength. The company reports training with 50–70 lb loads and subsequently carrying a loaded refrigerator exceeding 100 lb. The contents could shift during movement. The account identifies older grippers on that robot and does not publish a repeated-trial success rate. [8]

That load result belongs to the documented experiment. It does not replace the sustained product rating or establish daily handling throughput. Shifted contents are useful test conditions because the controller cannot rely on one fixed mass distribution throughout the lift. [8]

Battery exchange still takes time

The sheet gives three minutes for autonomous battery exchange and 1.5 hours for charging. Exchange time and recharge time are separate constraints. Sustained work requires charged packs and access to the station, whose failure rate is not reported. [4]

Factory training is confirmed

On 21 September 2026, Boston Dynamics announced that Atlas robots were training at the operational Robotics Metaplant Application Center on Hyundai's campus near Savannah, Georgia. The work includes preparing automotive parts and placing them in assembly order. This confirms a physical training deployment. The same announcement treats later assembly work and expansion across factories as plans. [9]

The sales sheet names part sequencing, machine tending and order fulfillment as target applications. Each still needs its own record of completed work, operator intervention and uptime. [4]

Control modes and remaining evidence gaps

The product announcement describes autonomous operation, VR teleoperation and tablet control. Boston Dynamics also describes a system that pauses when people enter a surrounding area, plus padding and reduced pinch points. Orbit supports supervision and task assignment. These features do not establish a safe clearance distance for every installed workstation; the reviewed descriptions provide no numeric stopping-distance envelope. [3] [10]

For a deployment decision, request results from the intended part, fixture and shift pattern. Useful records include completed transfers, drops, recovery time, battery-station interruptions and human interventions. The reviewed sources document hardware and specific tasks. They leave sustained factory throughput, lifetime maintenance cost and configuration-specific commercial pricing unanswered.