Two cobots using autonomous evaluation rollouts from finetuned LBMs to perform long-horizon behaviors, like installing a bike rotor. | Source: Toyota Research Institute
Toyota Research Institute (TRI) this week released the results of its study on Large Behavior Models (LBMs) that can be used to train general-purpose robots. The study showed a single LBM can learn hundreds of tasks and use prior knowledge to acquire new skills with 80% less training data.
LBMs are pretrained on large, diverse manipulation datasets. Despite their growing popularity, the robotics community knows surprisingly little about the nuances of what LBMs actually offer. TRI’s work aims to shed light on recent progress in algorithm and dataset design with this study.
In all, TRI said its findings largely support the recent surge in popularity of LBM-style robot foundation models, adding to evidence that large-scale pretraining on diverse robot data is a viable path towards more capable robots, though with a few points of caution.
General-purpose robots promise a future where household robots can provide everyday assistance. However, we’re not at the point where any robot can tackle average household tasks. LBMs, or embodied AI systems that take in robot sensor data and output actions, could change that, TRI said.
In 2024, TRI won an RBR50 Robotics Innovation Award for its work building LBMs for fast robot teaching.
An overview of TRI’s findings
TRI trained a series of diffusion-based LBMs on almost 1,700 hours of robot data and conducted 1,800 real-world evaluation rollouts and over 47,000 simulation rollouts to rigorously study their capabilities. It found that LBMs:
- Deliver consistent performance improvements relative to from-scratch policies
- Enable new tasks to be learned with 3-5× less data in challenging settings requiring robustness to a variety of environmental factors
- Improve steadily as pretraining data increases
Even with just a few hundred diverse hours of data, and only a few hundred demos per behavior, performance jumped meaningfully, TRI said. Pretraining provides consistent performance uplifts at earlier than expected scales. There is not yet an internet worth of robot data, but benefits appear far before that scale — a promising sign for enabling virtuous cycles of data acquisition and bootstrapped performance, TRI claimed.
TRI’s evaluation suite includes several novel and highly challenging long-horizon real-world tasks; finetuned and evaluated in this setting, LBM pretraining improves performance despite these behaviors being highly distinct from the pretraining tasks.
Inside the architecture and data of TRI’s LBMs
The LBM architecture is instantiated as a diffusion transformer which predicts robot actions. | Source: Toyota Research Institute
TRI’s LBMs are scaled multitask diffusion policies with multimodal ViT vision-language encoders and a transformer denoising head conditioned on encoded observations via AdaLN. These models consume wrist and scene cameras, robot proprioception, and language prompts and predict 16 timesteps (1.6 second) action chunks.
The researchers trained the LBMs on a mixture of 468 hours of internally collected bimanual robot teleoperation data, 45 hours of simulation-collected teleoperation data, 32 hours of Universal Manipulation Interface (UMI) data, and roughly 1,150 hours of internet data curated from the Open X-Embodiment dataset.
While the proportion of simulation data is small, its inclusion in TRI’s pretraining mixture ensures that it can evaluate the same LBM checkpoint in both sim and real.
TRI’s evaluation methods
TRI evaluates its LBM models on a bimanual platform across a variety of tasks and environmental conditions in both simulation and the real world. | Source: Toyota Research Institute
TRI evaluates its LBMs on physical and Drake-simulated bimanual stations employing Franka Panda FR3 arms and up to six cameras — up to two on each wrist, and two static scene cameras.
It evaluates the models on both seen tasks (present in the pretraining data) and unseen tasks (which TRI uses to fine-tune its pretrained model). TRI’s evaluation suite consists of 16 simulated seen-during-pretraining tasks, 3 real-world seen-during-pretraining tasks, 5 previously unseen long-horizon simulated tasks, and 5 complex previously unseen long-horizon real-world tasks.
Each model was tested via 50 rollouts for each real-world task and 200 rollouts for each simulation task. This enables a high level of statistical rigour in our analysis, with the pretrained models evaluated on 4,200 rollouts across 29 tasks.
TRI said it carefully controls initial conditions to be consistent in both the real world and simulation. It also conducts blind A/B-style testing in the real world with statistical significance computed via a sequential hypothesis testing framework.
Many of the effects the researchers observed were only measurable with larger-than-standard sample sizes and careful statistical testing that is non-standard for empirical robotics. It’s easy for noise due to experimental variation to dwarf the effects being measured, and many robotics papers may be measuring statistical noise due to insufficient statistical power.
SITE AD for the 2025 RoboBusiness registration open.
TRI’s top takeaways from the research
One of the team’s main takeaways is that finetuned performance smoothly improves with increasing pretraining data. At the data scales we examined, TRI saw no evidence of performance discontinuities or sharp inflection points; AI scaling appears alive and well in robotics.
TRI did experience mixed results with non-finetuned pretrained LBMs, however. Encouragingly, it found that a single network is able to learn many tasks simultaneously, but it doesn’t observe consistent outperformance from scratch single-task training without fine-tuning. TRI expects this is partially due to the language steerability of its model.
In internal testing, TRI said it has seen some promising early signs that larger VLA prototypes overcome some of this difficulty, but more work is required to rigorously examine this effect in higher-language-capacity models.
When it comes to points of caution, TRI said subtle design choices like data normalization can have large effects on performance, often dominating architectural or algorithmic changes. It’s important that these design choices are carefully isolated to avoid conflating the source of performance changes.