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

ItemDetails
AnnouncedAugust 20, 2025
CompaniesBoston Dynamics + Toyota Research Institute (TRI)
Blogbostondynamics.com/blog
RobotsAtlas (50 DoF), Atlas MTS (29 DoF)
LeadersScott 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

ItemSpec
Parameters450M
ArchitectureDiffusion Transformer
ObjectiveFlow Matching
Input Frequency30 Hz (images)
Action Chunk48 timesteps (1.6 seconds)
Execution~24 actions per cycle

Input Modalities

  • 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

StageDescription
1. Data CollectionVR teleoperation + MPC controller
2. ProcessingData annotation, QA, curation for ML pipelines
3. TrainingMulti-task, language-conditioned neural network policies
4. EvaluationSystematic testing and iterative improvements

VR Teleoperation System

FeatureDescription
Stereo HMDHead-mounted camera feeds for spatial awareness
Bimanual Mapping1:1 intuitive bimanual control
Foot TrackersDynamic repositioning and stepping enabled
Haptic FeedbackReal-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)

ItemSpec
DoF50 DoF
GrippersDual 7-DoF
CamerasHDR stereo head-mounted

Atlas MTS (Upper Body)

ItemSpec
DoF29 DoF
PurposeUpper-body manipulation testing

Both platforms share identical hardware/software, enabling cross-embodiment learning.


Cross-Embodiment Learning

LBM performs shared learning across multiple robot platforms:

PlatformDescription
AtlasBoston Dynamics full-body humanoid
Atlas MTS29-DoF upper-body variant
TRI RamenToyota 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

DateEvent
2024.10BD-TRI partnership announced
2025.08LBM + Atlas demo released

Leadership

  • Scott Kuindersma: Boston Dynamics AI Institute
  • Russ Tedrake: Toyota Research Institute

Comparison with Other VLAs

ModelParametersArchitectureWhole-BodyCross-Embodiment
LBM450MDiffusion TransformerYesYes
π03.3BFlow Matching + VLMNo (upper)Yes
Figure Helix 02-System 0/1/2YesNo
GR00T N12.2BDual-SystemYesYes

References


See Also