BD

Atlas Large Behavior Model

Boston Dynamics, United States 路 August 2025

Boston Dynamics and Toyota Research Institute

Robot foundation model 路 Closed weights 路 Also written Atlas LBM, Boston Dynamics LBM

Parameters
450 M
450M parameters according to the Boston Dynamics blog post.
GPU memory
0.8 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

Boston Dynamics and Toyota Research Institute built a language-conditioned Large Behavior Model that drives the whole Atlas humanoid, walking, crouching and manipulating with one policy. It is a 450M parameter diffusion transformer trained with flow matching on head camera images, proprioception and a language prompt, predicting 48 actions (1.6 s) per chunk and usually executing 24. Pretraining pools data from Atlas, the upper-body Atlas Manipulation Test Stand and TRI Ramen data, with simulation as a major co-training source. Policies could be run 1.5 to 2 times faster without significantly hurting performance, but no success rates, weights or code are published.

Architecture

Backbone
Not detailed. A diffusion transformer that builds on TRI's Large Behavior Models; no vision language model backbone is described.
Action head
Flow matching on a diffusion transformer, 48-step action chunks (1.6 s), with about 24 actions executed per inference call.
Action chunk
48 steps per forward pass
Parameters
450M parameters according to the Boston Dynamics blog post.
Pretraining data
Teleoperation data from Atlas and the Atlas Manipulation Test Stand plus TRI Ramen data, co-trained with simulation.
Embodiments
Boston Dynamics Atlas humanoid (50 DoF), Atlas Manipulation Test Stand (upper body, 29 DoF)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 450 M parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp321.7 GBTraining master weights, rarely used for inference
bf16 or fp160.8 GBThe usual way these checkpoints are served
int80.4 GBQuantised, expect some loss of precision on fine motions
int40.2 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Not published.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Demonstrations on Atlas include a long packing, sorting and organizing sequence with walking and lifting, and recovery from disruptions such as a closed box lid. On the test stand one policy handles rope tying, flipping a barstool, unfurling a tablecloth and moving a 22 lb tire. No success rates are given.

Fine tuning it yourself

Not available to the public.

Where it helps, where it does not

Strengths

  • One policy controls the whole humanoid, treating hands and feet almost alike, instead of separate locomotion and arm controllers.
  • Multi-embodiment training across a full humanoid and an upper-body stand.
  • Can be executed 1.5 to 2 times faster than recorded without a significant performance drop, according to the authors.

Limits

  • Closed: no weights, no code and no Atlas-specific paper.
  • No quantitative results are published.
  • Tied to Boston Dynamics hardware.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-12.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.