Atlas (robot)

Atlas is the name of a humanoid robot program developed by Boston Dynamics. It covers several distinct machines: the hydraulically actuated disaster-response platform unveiled for the DARPA Robotics Challenge in 2013, a smaller hydraulic research robot introduced in 2016, an electric research prototype revealed in 2024, and a separate product version announced in January 2026. Treating those machines as one unchanging robot leads to incorrect specifications and deployment claims.

The hydraulic program was used to study bipedal locomotion, whole-body control, perception, parkour, and manipulation. Boston Dynamics retired that hardware in April 2024 and introduced an electric successor the next day. The company then used an electric prototype for factory trials and learning research before unveiling a 1.9 m, 90 kg product with 56 degrees of freedom, a 2.3 m reach, and several load ratings. The product's published maximum instantaneous load is 50 kg; its sustained and one-handed ratings are lower.^[18][26][27]

Atlas is therefore both a research lineage and, since 2026, a manufacturing program. Boston Dynamics said in January 2026 that product manufacturing had begun and that the year's deployments were committed to Hyundai and Google DeepMind. Those announcements did not disclose a unit price or report a general public sale. At the research cutoff of July 28, 2026, the evidence supports a 2025 Hyundai factory pilot, manufacturing in progress, scheduled customer fleets, and a July 2026 public demonstration of the production robot. It does not support the earlier page's claim that more than 25,000 Atlas units had been deployed, or that the robot could turn 360 degrees in 0.5 seconds.^[28][32][35]

What the Atlas name covers

Atlas is a family name, not a single chassis upgraded continuously since 2013. Dimensions, joint counts, actuation, computers, hands, and intended uses differ by generation. Even the electric period contains an important boundary between the research prototype shown from 2024 through 2025 and the product model specified in 2026.

Scroll sideways for more →

Boston Dynamics engineer Josh Reynolds described PetProto, built from BigDog parts, as the company's first humanoid prototype and a precursor to PETMAN. He also identified the late hydraulic Atlas as Humanoid Version D, or HD, and described it as a research platform. That internal history helps explain why the program is older than the Atlas name and why the 2016-2024 robot should not be confused with the DRC chassis.^[12]

PETMAN and the disaster-response origin

PETMAN and PetProto

Atlas grew out of Boston Dynamics' work on PETMAN, short for Protection Ensemble Test Mannequin. PETMAN was designed to test chemical protective clothing in a controlled chamber while walking and performing calisthenics. Its artificial skin could regulate temperature and perspiration, and chemical sensors were intended to detect whether agents penetrated a suit. A tether supplied hydraulic and electrical power, cooling water, and communications.^[1]

The peer-reviewed PETMAN account says design work began in 2009 and describes a three-year development program. Before the complete mannequin was ready, Boston Dynamics assembled PetProto from BigDog hardware in four months to develop dynamic walking control. PetProto used two five-degree-of-freedom legs, reached 7.2 km/h in tests, and carried a 40 kg payload. The finished PETMAN weighed about 80 kg, used 29 actuated joints, contained onboard sensing and control, and demonstrated dynamically balanced heel-to-toe walking up to 4.8 km/h. Those figures describe PETMAN and PetProto, not later Atlas products.^[1]

PETMAN's relevance to Atlas is technical rather than merely visual. The program required a human-scale hydraulic machine with strong, compact actuation, broad joint motion, state estimation, foot-placement control, and recovery from moderate disturbances. Its control architecture separated discrete stepping decisions from continuous regulation of body height, orientation, and center of pressure. Atlas inherited that body of work while taking on a different mission.

The original DARPA platform

DARPA commissioned Atlas as standardized hardware for software teams in a disaster-response competition prompted by the limitations exposed during the 2011 Fukushima Daiichi nuclear accident. DARPA publicly unveiled the robot on July 11, 2013. The agency described a 6 ft 2 in, 330 lb machine with an onboard real-time control computer, a hydraulic pump and thermal-management system, 28 hydraulically actuated joints, a Carnegie Robotics head with lidar and stereo sensors, and two alternative sets of hands supplied by iRobot and Sandia National Laboratories.^[2]

The original machine still depended on a tether for external power. DARPA characterized Atlas as a physical shell whose useful behavior would come from team software and human operators. Seven teams received Atlas hardware after the Virtual Robotics Challenge. This division of responsibility matters: Boston Dynamics built and supported the common platform, while different teams wrote perception, planning, control, and operator-interface systems.

Scroll sideways for more →

The DRC tasks were designed around disaster sites made for people and ordinary tools. They included driving, crossing rough ground, moving debris, opening doors, climbing, cutting through a wall, connecting a hose, and operating valves. A humanoid form was useful because the test environment contained stairs, vehicles, doors, and tools with human proportions. The competition was not a test of an independent general-purpose robot; it evaluated human-robot systems under constrained communications.

Atlas in the DARPA Robotics Challenge

Trials in 2013

The DRC Trials took place at Homestead-Miami Speedway on December 20 and 21, 2013. Sixteen teams attempted eight groups of disaster-response tasks. SCHAFT, using its own robot, led with 27 of 32 possible points. IHMC Robotics scored 20 points with Atlas hardware and received task recognition for doors and wall cutting. DARPA selected eight teams for further funding after the event.^[3]

The results showed why a common hardware platform did not produce common performance. Atlas teams differed in their perception pipelines, step planners, whole-body controllers, task interfaces, recovery procedures, and operator strategies. The Trials also exposed practical limits such as modeling error, uncertain contact, limited force control, slow operation, and vulnerability to falls.

Academic work from participating groups documents the control problem more precisely. Carnegie Mellon researchers described a two-level optimization system in which a high-level planner generated center-of-mass and swing-foot trajectories and a low-level controller solved inverse-dynamics and inverse-kinematics problems. They added the kinematic component because model and torque-measurement errors made pure inverse-dynamics control difficult on the physical hydraulic robot.^[8]

MIT's team described a different optimization-based stack. Its footstep planner used mixed-integer convex optimization, its state estimator combined inertial, kinematic, and lidar information, and a quadratic-program controller stabilized walking and whole-body trajectories. The paper reports reliable balancing and walking but also identifies sensor limitations, model inaccuracies, and torque-feasibility as open problems. It presented planned running and jumping motions as future work rather than accomplishments already transferred to the robot.^[7]

The 2015 wireless upgrade

For the DRC Finals, DARPA required robots to operate without power, communications, or fall-arrest tethers. Communications could be degraded deliberately, and teams could not touch a robot after a run began. Boston Dynamics consequently rebuilt much of Atlas. DARPA said the upgraded machine was 75 percent new, with only the lower legs and feet carried over.^[4]

The upgraded Atlas remained 1.88 m tall but weighed 156.5 kg. It carried a 3.7 kWh lithium-ion battery with the potential for about one hour of mixed activity, used three onboard perception computers, and communicated through a wireless router in the head. A variable-pressure hydraulic pump improved efficiency, and resized hip, knee, and back actuators increased strength. Those specifications apply to the 2015 DRC machine, not to the smaller robot introduced the following year.^[4]

Seven of the 25 teams listed for the Finals planned to use the upgraded Atlas. DARPA emphasized that identical hardware would make software, interfaces, task selection, and strategy the differentiators.^[5] The Finals were held June 5 and 6, 2015, at Fairplex in Pomona, California. Team KAIST won with DRC-Hubo, IHMC Robotics placed second with an Atlas-based system, and Tartan Rescue placed third with CHIMP.^[6]

The second-place result is evidence that Atlas could support a strong disaster-response system, not that the base robot completed the course autonomously. Operators supervised task execution, communications were part of the test, and the team's software supplied the planning and control logic. Falls across the event also illustrated the gap between isolated successful behaviors and reliable field operation.

Research beyond the competition

Atlas remained a widely studied experimental platform after the DRC. Researchers demonstrated methods for walking on partial footholds by estimating available contact area from foot rotation and center-of-pressure measurements, then combining fast stepping with angular-momentum control. Their paper reports experiments on line and point contacts and discusses performance limitations rather than presenting the method as unrestricted terrain autonomy.^[9]

A separate study adjusted both step timing and step location to improve recovery from disturbances. It combined ankle-based balancing, faster swing motion, foot-placement changes, and a momentum objective in a whole-body controller. The authors reported simulation and physical Atlas experiments while identifying robustness improvements still needed.^[10] These papers are useful evidence about Atlas because they expose methods, assumptions, and failure boundaries that edited demonstration videos do not.

The smaller hydraulic research Atlas

The 2016 redesign

Boston Dynamics released "Atlas, The Next Generation" on February 23, 2016. Its description called the robot electrically powered and hydraulically actuated: a battery drove the hydraulic system, so "electric power" did not mean electric joint actuators. The new machine stood about 5 ft 9 in and weighed 180 lb. Body and leg sensors supported balance, while lidar and stereo sensors in the head supported obstacle avoidance, terrain assessment, navigation, and object manipulation.^[11]

The demonstration showed Atlas opening a door, walking through snow and uneven woods, lifting boxes, recovering after pushes, and standing after being knocked down. These were genuine hardware demonstrations, but a short edited video did not establish task success rates, duration, or readiness for unsupervised customer work. Boston Dynamics continued to describe the machine as a research platform.^[12]

The 2016 chassis became the basis for the hydraulic Atlas demonstrations that followed, but it did not remain mechanically or computationally frozen. Boston Dynamics refined hardware, grippers, perception, planning, and controllers throughout the period. Dimensions reported for a particular later demonstration should therefore not be treated as a new product specification unless the company identified them that way.

Backflips, dance, and parkour

In November 2017, Atlas performed a standing backflip and received a Guinness World Records listing for the first backflip by a bipedal humanoid robot. The record is narrowly defined: it documents the demonstrated maneuver, not general athletic superiority across arbitrary settings.^[13]

Boston Dynamics used parkour as a research program for dynamic locomotion, behavior creation, perception, and control. In a 2021 two-robot course, Atlas systems ran over banked panels, jumped gaps, climbed steps, vaulted, and performed backflips. The company's own account notes that the vault succeeded about half the time during filming and that the long routine compounded small failure probabilities. It also distinguishes reactive parkour from earlier floor and dance routines that followed prepared sequences without sensing and adapting to the environment.^[14]

The later parkour stack had three major layers:

1. A depth camera produced point clouds, and perception software extracted and tracked obstacle surfaces.
2. A high-level map supplied approximate obstacle types and desired maneuvers, while live perception supplied geometric details.
3. A library of behaviors built with offline trajectory optimization provided templates that a model predictive control system adjusted online.

The controller repeatedly solved an optimization over predicted motion, adjusting forces, posture, and timing to handle geometry changes and slips. It could interpolate around a behavior template, but it did not invent any possible maneuver from scratch. Boston Dynamics explicitly described limits on how far the controller could depart from its prepared repertoire.^[15]

Dance explored a related but different problem. Choreographers and engineers translated an artistic reference into motions the control system could execute while maintaining balance and respecting the robot's constraints. The well-known 2020 group routine demonstrated precision and sequencing, but Boston Dynamics later contrasted its largely preplanned performance with perception-driven parkour. Claims that the dance itself showed visual understanding or general autonomy are therefore inaccurate.^[14][15]

Whole-body manipulation

The 2023 "Atlas Gets a Grip" demonstration shifted attention from locomotion alone to loco-manipulation. Atlas grasped and moved a plank, carried and threw a tool bag, pushed a wooden block, climbed, and finished with an inverted 540-degree maneuver. Engineers said the asymmetric flip was near the hardware's limit and required avoiding self-collisions. The same account still called Atlas purely a research platform.^[16]

A companion technical article explains the control changes. Earlier parkour and dance controllers used a simplified model centered on total mass and inertia. For manipulation, the model included every link's momentum and forces exchanged with a carried or thrown object. Engineers could layer a locomotion reference with an object-motion reference, leaving the controller to find a feasible combined motion online. The tool bag weighed 7.8 kg, and a separate example used a 35 lb curl bar.^[17]

Those figures are demonstration loads, not maximum hydraulic Atlas payload specifications. Boston Dynamics did not publish a product-style load table for HD Atlas comparable to the 2026 electric product sheet. The 50 kg rating published later belongs to product Atlas.

Retirement

Boston Dynamics announced the retirement of hydraulic Atlas on April 16, 2024 and introduced a fully electric successor on April 17. Its launch article described the new platform as designed for real applications and said the program would begin customer testing with Hyundai. The article also referred to the just-retired machine as HD Atlas, making clear that the next robot was a new design rather than the same chassis with electric pumps.^[18]

The hydraulic platform's legacy lies in the engineering methods it helped develop: high-force actuation, state estimation, dynamic balance, trajectory optimization, predictive control, perception-guided motion, and whole-body manipulation. Its videos made that work visible, but they should be interpreted as selected research demonstrations. Boston Dynamics' own reliability disclosures and the academic papers provide essential context about failures and unsolved problems.

The electric Atlas research prototype

Mechanical reset in 2024

The first electric Atlas video showed a robot rising from the floor by folding its limbs through ranges unlike a human's and then rotating its torso and head. Boston Dynamics said it had designed the platform to complete tasks efficiently instead of preserving human joint limits. The company described the electric robot as stronger and more dexterous than previous generations and said it was exploring different grippers for expected customer applications.^[18]

The early announcement did not publish the later product's complete dimensions, joint count, battery figures, protection rating, or load table. It also did not support several claims that appeared in the earlier article, including an aluminum-titanium construction, a 0.5-second full turn, and the absence of wires across every rotating joint. Those details are omitted unless an archived source specifies them.

The electric prototype and the product share design ideas, including electric actuation, broad joint rotation, bimanual manipulation, and industrial part handling. They are not specification-identical. For example, Boston Dynamics' August 2025 behavior-model article described its research Atlas configuration as having 50 degrees of freedom, whereas the January 2026 product sheet specifies 56. That difference is a reason to attribute every number to its dated configuration.^[22][27]

Factory application and perception

Boston Dynamics selected automotive part sequencing as an early application. In this task, parts arrive in one fixture and must be moved into an ordered rack for the next manufacturing step. The work combines repetitive value with nontrivial perception, grasping, transport, recovery, and precise placement. The company used it to develop both application-specific systems and reusable capabilities.^[25]

The documented perception stack did more than run a generic object detector. A lightweight computer vision model identified fixtures, predicted outer and internal keypoints, and supported pose estimation. A tracker fused visual observations with joint-encoder kinematics and, when useful, force information. Because a grasped part could be occluded, kinematics supplied a prior for its location. Calibration procedures aligned camera and robot geometry, while a fixed-history estimator reconciled high-rate motion information with slower visual pose estimates.^[24]

The system could notice a displaced fixture, revise its estimate, and replan. It could also attempt recovery after a dropped or misinserted part. These were application demonstrations on the prototype; they were not a public benchmark across factories, fixtures, lighting conditions, or part types. Boston Dynamics described moving toward a more unified model, which indicates that the published perception stack remained an active development system.^[24]

Hyundai's CES material says the prototype completed a pilot at Hyundai Motor Group Metaplant America in October 2025 and repeatedly performed sequencing tasks. Boston Dynamics separately describes a fall 2025 metaplant test as Atlas's first customer deployment. The product model shown at CES was a different machine built from that development work.^[32][40]

NVIDIA compute and reinforcement learning

NVIDIA announced Project GR00T in March 2024 as a general-purpose foundation model initiative for humanoid robots, along with the Jetson Thor computer and Isaac simulation and learning tools. That announcement described a platform and an ecosystem, not a finished GR00T policy already operating Atlas.^[19]

In March 2025, Boston Dynamics said it was designing Jetson Thor into electric Atlas and using Isaac Lab for learned dexterity and locomotion policies. The stated objective was to run multimodal models alongside whole-body and manipulation controllers and to train behaviors in large numbers of simulated environments. The announcement also identified functional safety, security, learning, and vision pipelines as collaboration topics, which should not be mistaken for completed certifications.^[20]

Reinforcement learning became one route for whole-body motion. Boston Dynamics later described a system trained to carry a mini-fridge while adapting its posture to an uneven, deformable load. The policy used proprioceptive feedback rather than an external object-motion sensor for this adaptation. The demonstration is evidence of a particular learned behavior and training method, not a published success rate for arbitrary heavy objects.^[38]

For a July 2026 football demonstration, engineers retargeted human motion-capture data to Atlas's nonhuman kinematics and then trained policies in simulation. Reinforcement learning reconciled the reference motion with balance, contacts, and actuator limits. The work produced football-inspired movements, including a "Ghost Rabona," and contributed to the production robot's World Cup routine. It also illustrates a recurring Atlas workflow: capture or design a reference, adapt it to the robot, train at scale in simulation, and validate on hardware.^[37]

Large behavior models with Toyota Research Institute

Boston Dynamics and Toyota Research Institute, or TRI, announced a joint research program in October 2024. TRI contributed its work on large behavior models, while Boston Dynamics contributed the electric Atlas platform and interfaces for programming and teleoperating whole-body behavior. The partnership announcement framed safety, assurance, whole-body sensing, data collection, and evaluation as research questions.^[21]

In August 2025, the partners showed Atlas performing long sequences that combined walking, crouching, grasping, regrasping, sorting, packing, and object articulation. Researchers introduced disturbances by moving objects or closing a container during a task, and the robot adjusted its actions. A single policy controlled hands and feet rather than delegating locomotion to a separate learned system.^[23]

The accompanying technical description gives unusually specific architecture details:

• The research robot was described as having 50 degrees of freedom, with seven-degree-of-freedom grippers.
• Two HDR stereo cameras provided visual input.
• Human demonstrations were collected through a teleoperation system built on Boston Dynamics' predictive controller.
• A 450 million parameter Diffusion Transformer used a flow-matching training objective.
• Images, proprioception, and a language prompt were mapped to whole-body actions at 30 Hz.
• The model predicted 48-action chunks, corresponding to 1.6 seconds, and normally executed 24 actions before the next inference.

The team also reported replaying some policies faster than their demonstration speed. These results are research disclosures, not a public independent benchmark, and the article does not show that every product Atlas customer receives the TRI model. The demonstration ran on a pre-product research configuration and should be described as an electric Atlas experiment.^[22]

Calling this work a robot foundation model does not mean the model had solved general-purpose robotics. The partners presented it as a step toward broader skill learning. They did not publish a standardized task suite, confidence intervals, fleet-scale reliability data, or a safety case for autonomous factory deployment. The most defensible conclusion is that a shared whole-body policy learned multiple long-horizon behaviors and reacted to selected disturbances in laboratory demonstrations.

Product Atlas

CES 2026 announcement

Boston Dynamics unveiled the product version of Atlas on January 5, 2026 during Hyundai's CES media event. Its announcement draws an explicit distinction: the product was unveiled, while a prototype performed the live stage demonstration. Manufacturing of the product version was to begin immediately at the company's Boston headquarters.^[26]

This wording corrects two common errors. First, the robot moving on stage was not proof that the displayed product unit had performed that routine. Second, "production-ready" did not mean thousands of deployed units. At announcement time, Boston Dynamics reported a committed customer schedule and manufacturing start, not completed fleet rollouts.

Published specifications

The product sheet dated December 23, 2025 and released for the January announcement provides the most complete authoritative specification. Where marketing pages use broader phrases such as "up to 50 kg," the sheet separates short-duration, sustained, and one-handed ratings.^[27]

Scroll sideways for more →

The height is 6.2 ft, not 6 ft 3 in as the earlier page stated. The source describes a 360-degree camera view and says most joints are fully rotational; it does not specify a 0.5-second rotation. The hands have tactile fingers and palms, but the sheet does not use the earlier page's unsupported description of a particular four-finger geometry. The battery numbers are nominal product ratings, not guarantees under every payload, temperature, or task cycle.^[27][31]

Boston Dynamics' product page repeats the four-hour nominal battery figure and the 50 kg instantaneous and 30 kg sustained ratings. It also describes automatic battery exchange and Orbit-based fleet oversight. Because both sources are from the manufacturer, these values are published specifications rather than independently verified endurance tests.^[39]

Control, fleet integration, and service

Product Atlas can be operated autonomously, teleoperated through virtual reality, or steered with a tablet. Boston Dynamics says its Orbit software connects robot operations to manufacturing execution systems, warehouse management systems, and other records. The company also says a learned task can be distributed across a fleet. Those are product capabilities and deployment claims, not evidence that any behavior will transfer safely to every facility without validation.^[26][28]

The product design emphasizes serviceability. Boston Dynamics says it reduced the number of unique parts, designed components around automotive supply chains, and made modules field-replaceable. Hyundai Mobis agreed to supply actuators and to work with Boston Dynamics on development, reliability, and scaled production. Hyundai Mobis called Boston Dynamics its first official robotics-components customer.^[26][30]

Published safety-related features include human detection and fenceless guarding. The sources do not provide a complete safety standard, certification record, protective-separation-speed curve, or validated failure rate. "Fenceless guarding" should therefore be reported as a listed system feature, not as proof that Atlas may operate in any shared workspace without site-specific risk assessment.^[27]

Product and prototype evidence boundaries

Several public records in 2026 concern different machines:

• Hyundai showed both a prototype and a product model at CES. Its exhibit article attributes the 2025 HMGMA pilot and sequencing demonstrations to the prototype.^[40]
• Boston Dynamics' launch release says the live CES stage demonstration used a prototype, while the product version was unveiled and entered manufacturing.^[26]
• The product sheet specifies the 56-degree-of-freedom commercial machine.^[27]
• The TRI research article reports a 50-degree-of-freedom research Atlas configuration.^[22]
• The July 2026 World Cup event was identified by both Boston Dynamics and Hyundai as a public appearance of the production version.^[35][36]

Keeping those records separate prevents research results from being presented as standard product functions and prevents planned product capabilities from being backdated to the 2024 prototype.

Deployment and manufacturing status

Hyundai pilot and 2026 customer schedule

Boston Dynamics says Hyundai was the first customer and that a deployment was completed in 2025. Its later retrospective places the first deployment at the Hyundai metaplant in Georgia in the fall of that year. Hyundai's own account dates a prototype pilot to October 2025 and describes part-sequencing work. These records support a customer-site prototype trial, not broad commercial production deployment.^[28][32][40]

At the January 2026 launch, Boston Dynamics said every Atlas deployment slot for 2026 was committed. Fleets were scheduled for Hyundai's Robotics Metaplant Application Center, or RMAC, and Google DeepMind, with additional customers planned for early 2027. "Scheduled to ship" and "planned" are the operative terms. The announcement did not list delivery counts, acceptance results, customer uptime, or commercial prices.^[26]

Boston Dynamics' subsequent evolution article says product units were being manufactured and deployments remained scheduled during 2026. That establishes production activity but does not by itself prove that both customer fleets had completed delivery by the cutoff.^[32]

Google DeepMind research

Boston Dynamics and Google DeepMind announced a partnership to explore Gemini Robotics models on Atlas. Their January release says joint research would begin in the coming months and aims to combine Boston Dynamics' whole-body control with visual-language-action and embodied-reasoning models. It does not say Gemini Robotics was already installed as the product's shipping autonomy stack.^[29]

The partnership should also be distinguished from the earlier TRI project. TRI's 2025 model had already been demonstrated on a research Atlas. The Google effort was a newly announced research program with future integration goals. Referring to both as "Atlas AI" without dates, organizations, and platform boundaries obscures their different status.

First public production demonstration

On July 5, 2026, a production Atlas appeared at halftime during the Norway-Brazil FIFA World Cup round-of-16 match at New York New Jersey Stadium. The robot delivered the ceremonial ball and performed goal-celebration motions. Boston Dynamics called it the first public appearance of production Atlas, and Hyundai called it the first public demonstration of that version's movement in a live match environment.^[35][36]

The routine used whole-body policies developed through the motion-retargeting and reinforcement-learning process described by Boston Dynamics. It is evidence that a production unit executed a prepared live demonstration in a stadium. It is not evidence of industrial throughput, unsupervised generality, or a completed customer fleet rollout.

Hyundai's longer-term plan

Hyundai's January 2026 roadmap calls for RMAC-trained Atlas robots to enter parts-sequencing processes at Hyundai Motor Group Metaplant America by 2028, followed by more complex component assembly around 2030. These are planned phases subject to training, validation, safety, quality, and business decisions. The dates must not be written as completed deployments.^[31]

Hyundai also stated an aim to establish production capacity for 30,000 robot units annually by 2028. That figure refers to a broader robot production system, not 30,000 Atlas units already made or deployed. The same source says Hyundai's network could generate demand for tens of thousands of Atlas units over time. Demand forecasts and capacity targets are not installed-base counts.^[31]

No archived primary source in this review supports the earlier statement that Hyundai expected to deploy more than 25,000 Atlas robots at US factories. The accurate formulation is narrower: the group has announced a staged Atlas plan, a broader 30,000-robot annual capacity target, and anticipated long-term demand. Actual Atlas fleet sizes were not disclosed by the cutoff.

Ownership and supply chain

Boston Dynamics was founded in 1992 by Marc Raibert as a spinout from the Massachusetts Institute of Technology's Leg Laboratory. During the Atlas program, the company passed from independence to Google, then to SoftBank, and finally to Hyundai-led ownership. SoftBank's 2017 announcement says it agreed to acquire Boston Dynamics from Alphabet; it did not disclose transaction terms.^[34]

Hyundai Motor Group completed its controlling acquisition on June 21, 2021. The transaction valued Boston Dynamics at $1.1 billion. Hyundai Motor Group held 80 percent after closing, while a SoftBank affiliate retained 20 percent.^[33] The ownership change gave Boston Dynamics access to an automotive manufacturer, component suppliers, factories, and a prospective first customer.

The relationship does not make every Hyundai production target an accomplished Boston Dynamics result. Corporate roadmaps, customer pilots, supplier agreements, and robot specifications remain distinct types of evidence. Hyundai Mobis' actuator agreement, for example, supports a supply-chain role but does not disclose production yield, actuator lifetime, or delivered robot counts.^[30]

Comparing the main generations

Scroll sideways for more →

Two comparisons require special care. First, hydraulic Atlas was not necessarily "less autonomous" in a single measurable sense. Autonomy depended on the task and software stack: DRC systems were heavily supervised, while later parkour and handling systems combined prepared behaviors with perception and online control. Second, electric actuation does not alone explain every improvement. The product also reflects changes in kinematics, hands, sensing, compute, training, enterprise integration, supply chain, service design, and the intended operating environment.

Boston Dynamics has not published enough comparable data to rank the generations by uptime, energy per task, fall rate, positional accuracy, or total cost of ownership. Statements that electric Atlas categorically exceeds HD Atlas on those measures would go beyond the evidence reviewed here.

Development chronology

Scroll sideways for more →

How to interpret Atlas demonstrations

Atlas has often been presented through short, highly produced videos. Those records are valuable because they show real machines executing dynamic motions, and Boston Dynamics has frequently supplemented them with technical explanations. They do not, by themselves, answer questions about repeatability, intervention rate, time between failures, maintenance, energy use, or performance across unprepared sites.

Three kinds of evidence should be kept separate:

1. Research demonstrations establish that a configuration executed a selected behavior under described conditions. Parkour, the 540-degree maneuver, the TRI task sequence, the heavy-load policy, and football motions belong here.
2. Published product specifications state intended physical and operating limits for the 2026 product. They are manufacturer claims unless an independent test is available.
3. Deployment records establish where and when robots were tried, shipped, or operated. A pilot, a scheduled fleet, a public demonstration, and a scaled production installation are not interchangeable.

This distinction does not diminish the engineering. It makes the record more useful. The academic work shows optimization and state-estimation methods, Boston Dynamics' technical articles disclose behavior-generation and perception architectures, and customer and supplier releases document the path toward manufacturing. Together they show a long transition from common research hardware to an industrial product, while leaving important commercial reliability and scale questions unanswered.

See also

• Robot learning
• Vision-language-action model
• Spot

References

1. ^Nelson, Gabe, et al. PETMAN: A Humanoid Robot for Testing Chemical Protective Clothing. Journal of the Robotics Society of Japan, volume 30, issue 4, 2012.
2. ^Defense Advanced Research Projects Agency. DARPA's ATLAS Robot Unveiled. July 11, 2013.
3. ^Defense Advanced Research Projects Agency. After Impressive Demonstrations of Robot Skill, DARPA Robotics Challenge Trials Conclude. December 26, 2013.
4. ^Defense Advanced Research Projects Agency. Upgraded Atlas Robot to Go Wireless as the Stakes Are Raised for the DARPA Robotics Challenge Finals. January 20, 2015.
5. ^Defense Advanced Research Projects Agency. Twenty-five Teams From around the World to Participate in DARPA Robotics Challenge Finals. March 5, 2015.
6. ^Defense Advanced Research Projects Agency. Robotics Challenge Workshop: From Better Robots to Better Futures. June 8, 2015.
7. ^Kuindersma, Scott, et al. Optimization-based locomotion planning, estimation, and control design for the Atlas humanoid robot. Autonomous Robots, volume 40, 2016.
8. ^Feng, Siyuan, et al. Optimization-based Full Body Control for the DARPA Robotics Challenge. Journal of Field Robotics, 2015.
9. ^Wiedebach, Georg, et al. Walking on Partial Footholds Including Line Contacts with the Humanoid Robot Atlas. 2016.
10. ^Griffin, Robert J., et al. Walking Stabilization Using Step Timing and Location Adjustment on the Humanoid Robot, Atlas. 2017.
11. ^Boston Dynamics. Atlas, The Next Generation. February 23, 2016.
12. ^Boston Dynamics. Ask a Roboticist: Q&A with Josh. Accessed July 28, 2026.
13. ^Guinness World Records. First bi-pedal humanoid robot backflip. November 2017.
14. ^Boston Dynamics. Leaps, Bounds, and Backflips. August 17, 2021.
15. ^Boston Dynamics. Flipping the Script with Atlas. August 18, 2021.
16. ^Boston Dynamics. Sick Tricks and Tricky Grips. January 18, 2023.
17. ^Boston Dynamics. Picking Up Momentum. January 19, 2023.
18. ^Boston Dynamics. An Electric New Era for Atlas. April 17, 2024.
19. ^NVIDIA. NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots and Major Isaac Robotics Platform Update. March 18, 2024.
20. ^Boston Dynamics. Boston Dynamics Expands Collaboration with NVIDIA to Accelerate AI Capabilities in Humanoid Robots. March 18, 2025.
21. ^Boston Dynamics. Boston Dynamics and Toyota Research Institute Announce Partnership to Advance Robotics Research. October 16, 2024.
22. ^Cousineau, Eric, et al. Large Behavior Models and Atlas Find New Footing. Boston Dynamics, August 20, 2025.
23. ^Toyota Research Institute. AI-Powered Robot by Boston Dynamics and Toyota Research Institute Takes a Key Step Towards General-Purpose Humanoids. August 20, 2025.
24. ^Boston Dynamics. Making Atlas See the World. 2025.
25. ^Boston Dynamics. Getting Real with Humanoids. 2025.
26. ^Boston Dynamics. Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry. January 5, 2026.
27. ^Boston Dynamics. Atlas specification sheet. Dated December 23, 2025.
28. ^Boston Dynamics. Enterprise Robotics, Redefined. January 2026.
29. ^Boston Dynamics. Boston Dynamics and Google DeepMind Form New AI Partnership to Bring Foundational Intelligence to Humanoid Robots. January 5, 2026.
30. ^Hyundai Mobis. Hyundai Mobis Forms Strategic Collaboration Framework with Boston Dynamics. January 8, 2026.
31. ^Hyundai Motor Group. Hyundai Motor Group Announces AI Robotics Strategy to Lead Human-Centered Robotics Era at CES 2026. January 5, 2026.
32. ^Boston Dynamics. Atlas' Evolution From Research Robot to Industrial Humanoid. 2026.
33. ^Hyundai Motor Group. Hyundai Motor Group Completes Acquisition of Boston Dynamics from SoftBank. June 21, 2021.
34. ^SoftBank Group. SoftBank Announces Agreement to Acquire Boston Dynamics. June 9, 2017.
35. ^Boston Dynamics. Deploying Robots Into (and Onto) the Field. July 2026.
36. ^Hyundai Motor Company. Hyundai Motor Brings Atlas Humanoid Robot to FIFA World Cup 2026 in First-Ever Live Match Environment Robotics Integration. July 5, 2026.
37. ^Boston Dynamics. Can football teach a robot to move?. 2026.
38. ^Boston Dynamics. Training a Humanoid Robot for Hard Work. 2026.
39. ^Boston Dynamics. Atlas Humanoid Robot. Accessed July 28, 2026.
40. ^Hyundai Motor Group. CES 2026: Hyundai Motor Group brings AI Robotics to life. January 2026.