Boston Dynamics Atlas teleoperation: from the VR rig to the Large Behavior Model

The electric Atlas is Boston Dynamics' production humanoid: 1.9 m tall, 90 kg, 56 degrees of freedom and a 50 kg instant lift, unveiled as a product at CES in January 2026. It is also the humanoid whose maker has been most open about how operators teach it. Boston Dynamics and Toyota Research Institute have described the VR rig, the data pipeline and the policy architecture in detail, and the head of the Atlas program has talked publicly about who the operators are and how long they take to train. That makes Atlas the best case study there is for anyone setting up humanoid teleoperation, even though very few people can buy one yet.

Can you buy an Atlas?

Not in 2026. In its CES 2026 announcement, Boston Dynamics said production was starting at its Boston headquarters, that "All Atlas deployments are already fully committed for 2026" to Hyundai's Robotics Metaplant Application Center and Google DeepMind, and that it plans to add customers in early 2027. The product page describes the current phase as early adoption with "a select number of early adopters". There is no published price and no public Atlas SDK. If you need a humanoid with operators on it before 2027, skip to the alternatives below.

Three ways to drive Atlas

Boston Dynamics says the product Atlas "can be controlled in three different ways: autonomous mode, teleoperated, or by using a tablet steering interface". Each one has a different job in a deployment:

The VR teleoperation rig, step by step

Boston Dynamics and TRI described the rig in their post on Large Behavior Models and Atlas. It sits on top of Atlas' model predictive controller, the same balance and whole-body control layer the robot uses on its own, through a custom VR interface:

From demonstrations to a Large Behavior Model

The published pipeline has four stages: collect teleop data on real hardware and in simulation; process, annotate and curate it, with quality-assurance tooling to review, filter and give feedback on episodes; train; and evaluate. The team wrote that "high quality demonstrations were a critical part of getting successful policies".

One caveat on hardware: the research Atlas in that post is listed at 50 degrees of freedom with 7-DoF grippers, while the CES product Atlas is listed at 56. Policies and datasets are tied to the body they were recorded on.

Who the operators are

In a September 2025 IEEE Spectrum interview, Atlas lead Scott Kuindersma gave the clearest description any humanoid maker has offered of its operator team: "The people doing it now are not roboticists; we have a team of 'robot teachers' who are hired for this." He said novice operators are "probably not generating the highest-quality motions for training policies", and that consistently producing good data over several hours "typically takes a couple of weeks of onboarding".

He also placed teleoperation at the top of the data pyramid, "the highest quality, hardest-to-get data", stressed the importance of a large baseline of actual on-robot data, said co-training with simulation had improved real-robot performance, and admitted that "no one knows the exact distribution" of the right mix. Those points are the planning assumptions for any humanoid data program: hire for physical skill and consistency rather than robotics degrees, budget weeks of onboarding per operator, and treat on-robot teleop as the scarce, expensive layer. Our humanoid teleoperator careers page describes the role from the operator side.

What a good operator does on an Atlas-style rig

Where TRI's LBM paper fits

People often cite TRI's Large Behavior Model study as Atlas research. It is related but not the same: the paper trains on almost 1,700 hours of data, including 468 hours of bimanual real-robot teleoperation, 45 hours of simulation teleoperation, 32 hours of UMI data and about 1,150 hours curated from Open X-Embodiment, and evaluates on bimanual Franka FR3 stations, not Atlas. Its headline result, that pretrained LBMs can learn new tasks with 3 to 5 times less data in challenging settings, is the reason the Atlas team adopted the approach. It also found that data normalization choices often mattered more than architecture changes, a reminder that consistent, well-formatted episodes are worth as much as clever models. Our teleoperation data collection guide covers how to keep episodes consistent from the first one.

Control loop and latency for a remote Atlas operator

Because the VR rig drives targets into MPC rather than raw joint commands, the controller on the robot keeps a 90 kg machine balanced even when the operator's signal arrives late or jittery. That is what makes remote operation of a robot this size plausible at all. The remaining question is task precision. Tire handling and bin moves tolerate a few hundred milliseconds of delay; tight assembly and threading do not, and need an operator close to the robot or a policy doing the fine part with the human supervising. Our teleoperation latency guide breaks budgets down by task.

What happens to Atlas in 2026

The 2026 fleet goes to two places. Hyundai's application center is the industrial test bed, and Hyundai Motor Group said at CES that it is preparing to deploy tens of thousands of Boston Dynamics robots in its facilities, with Hyundai Mobis supplying Atlas actuators. The Google DeepMind partnership will combine DeepMind's Gemini Robotics models with Atlas for industrial tasks, starting with automotive, with joint research due to begin in 2026. Neither announcement says anything about outside data access, so expect no public Atlas datasets in the near term.

If you need humanoid operators before 2027

Teams that want to reproduce the Atlas recipe now (VR teleop, recovery demonstrations, a diffusion or flow-matching policy) usually do it on a buyable research humanoid. The full-size Unitree H2 is the nearest in stature and was added to Unitree's open xr_teleoperate stack in v1.6, which records episodes directly. The smaller Unitree G1 has the largest open teleop and dataset ecosystem of any humanoid.

How RoboPilots supplies robot teachers

Kuindersma's description of the job is close to what we staff. Several humanoid makers have sent RoboPilots robots to test our operating protocols. On a customer program we train a dedicated group of operators on your own unit and task protocol, including recovery demonstrations, so the weeks of onboarding happen once and the same people stay on your robot.