Atlas LBM

Atlas LBM (Large Behavior Model on Atlas)

Boston Dynamics·Toyota Research Institute · 2025-08

VLAHumanoid VLA450Mflow matching

A single policy directly outputs whole-body actions including gripper joints·neck yaw·torso pose·hand·foot poses.

Paper

Backbone LLM: None (450M DiT, language-conditioned)
Vision encoder: —
Action head: flow matching DiT 450M (flow matching)
Chunk: 48 steps (24 executed)
Control rate: Not disclosed (images 30 Hz)
Architecture (original summary): 450M Diffusion Transformer + flow matching; conditioned on proprioception · images (30 Hz input) · language, predicts a 48-step (1.6 s) action chunk from observation history, typically executes 24 steps per inference (0.8 s at 1x speed)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed
Data type: Real-robot teleoperation / simulation·synthetic
Data scale: Not disclosed (hours·episodes Not disclosed)
Main datasets: VR whole-body tracking teleoperation (real robot·sim), reviewed·filtered with QA tools
Tasks & evaluation: Spot Workshop long-horizon tasks (picking·folding·regrasping·articulated manipulation·placing·pushing), rope tying, bar stool flipping, tablecloth manipulation, 22 lb tire; success rate not disclosed
Robot: Boston Dynamics Atlas (50 DoF), Atlas Manipulation Test Stand (upper body 29 DoF)

Inference device & speed: Hardware and latency not disclosed; control rate (Hz) not disclosed (30 Hz is the image input rate)
Inference memory: Not disclosed
Weights bf16 (calc.): 0.9 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Imitation learning on VR teleoperation·sim data. Low-level control method not disclosed
Base control: Humanoid legged (Direct whole-body joint/pose output) · Policy outputs foot pose, torso pose, and gripper joints → executed via the same MPC control interface as teleoperation (stepping, squatting); manipulation test stand uses upper body only

Key features: A single policy directly outputs whole-body actions including gripper joints·neck yaw·torso pose·hand·foot poses. A multi-embodiment policy covering both Atlas and MTS, it can run 1.5~2× faster without performance loss.
Notes: Verification corrected '30 Hz' to image input rate rather than control rate. Sim is a co-training source. TRI's general LBM 1.0 (bimanual, non-humanoid) is in the scope of the industry notes. Blog·press release 2025-08-20.
Verification: Corrections applied