LBM (Large Behavior Model)
Boston Dynamics and Toyota Research Institute's Whole-Body Control Model for Atlas
Author’s Note
- LBM is the first commercial humanoid system to integrate locomotion and manipulation control in a single model.
- The approach of “if you can demonstrate it, the robot can learn it” is impressive. Deformable object manipulation like rope tying and cloth spreading uses the same training pipeline.
- A powerful collaboration combining Boston Dynamics’ hardware with TRI’s AI capabilities.
Key Significance
- Whole-Body Single Model Control: Integrates locomotion, balance, and manipulation in one policy without separation
- 450M Diffusion Transformer: Flow Matching based, 30Hz image input, 48 timestep action chunks
- Language-Conditioned Multi-Task: Single policy performs diverse tasks including rope tying, tire manipulation, cloth spreading
- VR Teleoperation: Intuitive data collection with stereo HMD + bimanual mapping + foot trackers
- Cross-Embodiment: Shared training across Atlas, Atlas MTS, and TRI Ramen platforms
Overview
| Item | Details |
|---|
| Announced | August 20, 2025 |
| Companies | Boston Dynamics + Toyota Research Institute (TRI) |
| Blog | bostondynamics.com/blog |
| Robots | Atlas (50 DoF), Atlas MTS (29 DoF) |
| Leaders | Scott Kuindersma, Russ Tedrake |
LBM (Large Behavior Model) is a humanoid whole-body control model jointly developed by Boston Dynamics and Toyota Research Institute. Following the partnership announcement in October 2024, they revealed results on Atlas in August 2025.
Architecture
Model Specifications
| Item | Spec |
|---|
| Parameters | 450M |
| Architecture | Diffusion Transformer |
| Objective | Flow Matching |
| Input Frequency | 30 Hz (images) |
| Action Chunk | 48 timesteps (1.6 seconds) |
| Execution | ~24 actions per cycle |
- Images: HDR stereo head-mounted cameras (30Hz)
- Proprioception: Joint states, force/torque sensors
- Language Prompts: Task objective specification
Whole-Body Integrated Control
While existing humanoids separated locomotion/balance control from manipulation control, LBM treats hands and feet almost identically, controlling the entire body with a single model.
Training Pipeline
4-Stage Training Process
| Stage | Description |
|---|
| 1. Data Collection | VR teleoperation + MPC controller |
| 2. Processing | Data annotation, QA, curation for ML pipelines |
| 3. Training | Multi-task, language-conditioned neural network policies |
| 4. Evaluation | Systematic testing and iterative improvements |
VR Teleoperation System
| Feature | Description |
|---|
| Stereo HMD | Head-mounted camera feeds for spatial awareness |
| Bimanual Mapping | 1:1 intuitive bimanual control |
| Foot Trackers | Dynamic repositioning and stepping enabled |
| Haptic Feedback | Real-time tactile feedback + AR overlays |
“Fluid, dynamic, and dexterous” control across both stationary and mobile manipulation tasks.
Capabilities
Whole-Body Coordination
- Locomotion, stepping, stance configuration
- Crouching, balance maintenance
- Lifting 22lb (10kg) objects
Dexterous Manipulation
- Grasping, regrasping, articulating objects
- Deformable objects: Rope tying, cloth spreading, tire manipulation
- Tasks extremely difficult with traditional robot programming
Reactive Recovery
- Intelligent responses to unexpected disturbances
- Automatic adaptation to fallen parts, closed boxes, etc.
- Emerges automatically from training examples without algorithm changes
Speed Adaptation
- 1.5x-2x speed adjustment at inference time
- Adjusts action timing predictions without retraining
Hardware: Atlas
Atlas (Full Body)
| Item | Spec |
|---|
| DoF | 50 DoF |
| Grippers | Dual 7-DoF |
| Cameras | HDR stereo head-mounted |
Atlas MTS (Upper Body)
| Item | Spec |
|---|
| DoF | 29 DoF |
| Purpose | Upper-body manipulation testing |
Both platforms share identical hardware/software, enabling cross-embodiment learning.
Cross-Embodiment Learning
LBM performs shared learning across multiple robot platforms:
| Platform | Description |
|---|
| Atlas | Boston Dynamics full-body humanoid |
| Atlas MTS | 29-DoF upper-body variant |
| TRI Ramen | Toyota Research Institute platform |
Multi-task batching enables shared policy improvements.
Core Philosophy
“If you can demonstrate it, the robot can learn it”
- Same training process whether stacking rigid blocks or folding a t-shirt
- Developing new manipulation behaviors no longer requires “an advanced degree and years of experience”
- Data-driven approach applicable to virtually any downstream task that can be demonstrated via teleoperation
Boston Dynamics & TRI Partnership
| Date | Event |
|---|
| 2024.10 | BD-TRI partnership announced |
| 2025.08 | LBM + Atlas demo released |
Leadership
- Scott Kuindersma: Boston Dynamics AI Institute
- Russ Tedrake: Toyota Research Institute
Comparison with Other VLAs
| Model | Parameters | Architecture | Whole-Body | Cross-Embodiment |
|---|
| LBM | 450M | Diffusion Transformer | Yes | Yes |
| π0 | 3.3B | Flow Matching + VLM | No (upper) | Yes |
| Figure Helix 02 | - | System 0/1/2 | Yes | No |
| GR00T N1 | 2.2B | Dual-System | Yes | Yes |
References
See Also