Overview
Boston Dynamics is a robotics company founded in 1992 by Marc Raibert as a spin-off from MIT’s Leg Laboratory research. As a pioneer in the field of dynamic locomotion robots, the company is known for Atlas’s backflips and Spot’s stable locomotion, and has been recognized as a leader in Classical Robotics. After the acquisition announcement in 2020, the company was officially acquired by Hyundai Motor Group in 2021.
| Item | Details |
|---|---|
| Headquarters | Waltham, Massachusetts |
| Founded | 1992 (MIT spin-off) |
| Founder | Marc Raibert |
| CEO | Robert Playter (2019~) |
| Parent Company | Hyundai Motor Group (2021~, 80% stake) |
Product Line
Spot
Quadruped robot. Commercially available for inspection, surveillance, and data collection.
Atlas
Bipedal humanoid robot. Used for research and development. Redesigned with electric actuation in 2024.
Stretch
Warehouse logistics robot. Specialized for box unloading tasks.
Classical Robotics Approach
Boston Dynamics has long been known for its modular approach:
- Perception: Understanding environment through sensors
- Planning: Path and motion planning
- Control: Precise control via Model Predictive Control (MPC)
This approach produced impressive demos like Atlas’s backflips, parkour, and Spot’s stable locomotion.
Transition to Physical AI
Collaboration with Toyota Research Institute (TRI)
In October 2024, Boston Dynamics and TRI announced a joint research partnership on Large Behavior Models (LBMs). In August 2025, they unveiled a demonstration of the Atlas humanoid robot powered by LBM.
“Boston Dynamics is developing Large Behavior Models (LBMs) for Atlas as part of a collaboration between AI research teams at Toyota Research Institute (TRI) and Boston Dynamics.” — The Robot Report
LBM Architecture
| Item | Specification |
|---|---|
| Parameters | 450M |
| Architecture | Diffusion Transformer |
| Output Frequency | 30Hz |
| Control DoF | 50 (full Atlas) |
“The specific architecture used for Atlas is a 450-million-parameter diffusion transformer… It outputs a continuous stream of actions at 30Hz to control all 50 of Atlas’s degrees of freedom.” — IEEE Spectrum
Three Learning Approaches
Boston Dynamics uses three parallel approaches to build Atlas’s intelligence:
- Teleoperation: Human operators guide the robot using VR
- Reinforcement Learning: Millions of motion practice runs in simulation
- Observation Learning: Learning physical intuition from human action videos (long-term goal)
Hybrid Approach
Boston Dynamics hasn’t adopted full end-to-end (pixels-to-torques), instead using a hybrid approach with learning-based high-level decision-making while maintaining MPC for low-level control.
Nevertheless, the fact that the leader of Classical Robotics is introducing learning-based models demonstrates the transition to the Physical AI paradigm.
Key Milestones
- 1992: Spun off from MIT (Founder: Marc Raibert)
- 2005: BigDog unveiled (DARPA-funded)
- 2013: Acquired by Google (December)
- 2017: Acquired by SoftBank (June)
- 2019: Robert Playter becomes CEO (October)
- 2020: Hyundai Motor Group acquisition announced (December)
- 2021: Hyundai acquisition completed (June, 80% stake)
- 2024: Electric-powered Atlas unveiled (April), hydraulic Atlas retired
- 2024: LBM research partnership with TRI announced (October)
- 2025: Atlas LBM demonstration unveiled (August)
References
- Boston Dynamics Official Site
- Large Behavior Models and Atlas Find New Footing
- Boston Dynamics Wikipedia