TRI

TRI Large Behavior Model

Toyota Research Institute, United States 路 July 2025

Toyota Research Institute (TRI LBM Team)

Robot foundation model 路 Closed weights 路 Also written LBM, Large Behavior Model, TRI LBM

Parameters
not published
Not stated in the paper or on the project page.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

TRI's Large Behavior Model extends Diffusion Policy into a multitask diffusion transformer pretrained on about 1,700 hours of robot data: 468 hours of bimanual teleoperation, 45 hours of simulation, 32 hours of UMI data and about 1,150 hours curated from Open X-Embodiment. It reads wrist and scene cameras, proprioception and a language prompt through CLIP encoders, not through a vision language model, and predicts 16 steps of 20-dimensional actions for bimanual Franka stations at 10 Hz. The study stands out for its evaluation, 1,800 blind, randomized real-world trials and over 47,000 in simulation: fine-tuned LBMs beat single-task policies more often, especially under distribution shift, and learned new tasks with less than 30 percent of the data in simulation, while the pretrained model without fine-tuning did not consistently beat single-task training. The paper and project page link no weights or code.

Architecture

Backbone
No VLM backbone. CLIP ViT-B/16 image encoder (fine-tuned) and frozen CLIP text encoder feeding a diffusion transformer of eight DiT blocks with embedding size 768, conditioned through AdaLN.
Action head
Diffusion denoising head predicting 16 timesteps of 20-dimensional bimanual actions; 8 are executed before the next prediction.
Action chunk
16 steps per forward pass
Control rate
10 Hz
Parameters
Not stated in the paper or on the project page.
Pretraining data
About 1,700 hours (1,695 in the paper): TRI-Ramen real teleoperation 468 hours (362 tasks, 46,063 demonstrations), TRI-Ramen simulation 45 hours, TRI-Ramen UMI 32 hours, and OXE-Ramen, a curated Open X-Embodiment subset of about 1,150 hours.
Embodiments
Bimanual Franka Panda FR3 stations, real and in Drake simulation, Open X-Embodiment robots in pretraining

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.

Published figures. These come from the model authors, not from this site.
Control rate10 Hz

Not published.

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

Real tasks used 50 rollouts per task, policy and condition. Fine-tuned LBMs were statistically the same or better than single-task baselines on seen real tasks and statistically better in aggregate on unseen tasks; on SetBreakfastTable, which single-task training never completed, an LBM fine-tuned on 15 percent of the data beat the single-task baseline.

Fine tuning it yourself

Fine-tuning starts from the pretrained LBM on task demonstrations; in simulation the authors needed less than 30 percent of the data that training from scratch requires. No public checkpoints or code are linked.

Where it helps, where it does not

Strengths

  • Unusually rigorous evaluation: blind, randomized trials with statistical tests, 1,800 real and over 47,000 simulated rollouts.
  • Shows where pretraining helps: fine-tuned models gain most under distribution shift and on new tasks.
  • Performance rises as pretraining data grows in scale and diversity.

Limits

  • No released weights or code.
  • Without fine-tuning the pretrained model does not consistently beat single-task policies.
  • Evaluated on one station type, the bimanual Franka FR3.
  • Results are reported as statistical comparisons and figures more than as single success rates.

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.