A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation
Abstract
Robot manipulation has seen tremendous progress in recent years, with imitation learning policies enabling successful performance of dexterous and hard-to-model tasks. Concurrently, scaling data and model size has led to the development of capable language and vision foundation models, motivating large-scale efforts to create general-purpose robot foundation models. While these models have garnered significant enthusiasm and investment, meaningful evaluation of real-world performance remains a challenge, limiting both the pace of development and inhibiting a nuanced understanding of current capabilities. In this paper, we rigorously evaluate multitask robot manipulation policies, referred to as Large Behavior Models (LBMs), by extending the Diffusion Policy paradigm across a corpus of simulated and real-world robot data. We propose and validate an evaluation pipeline to rigorously analyze the capabilities of these models with statistical confidence. We compare against single-task baselines through blind, randomized trials in a controlled setting, using both simulation and real-world experiments. We find that multi-task pretraining makes the policies more successful and robust, and enables teaching complex new tasks more quickly, using a fraction of the data when compared to single-task baselines. Moreover, performance predictably increases as pretraining scale and diversity grows. Project page: https://toyotaresearchinstitute.github.io/lbm1/
Cite
@article{arxiv.2507.05331,
title = {A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation},
author = {TRI LBM Team and Jose Barreiros and Andrew Beaulieu and Aditya Bhat and Rick Cory and Eric Cousineau and Hongkai Dai and Ching-Hsin Fang and Kunimatsu Hashimoto and Muhammad Zubair Irshad and Masha Itkina and Naveen Kuppuswamy and Kuan-Hui Lee and Katherine Liu and Dale McConachie and Ian McMahon and Haruki Nishimura and Calder Phillips-Grafflin and Charles Richter and Paarth Shah and Krishnan Srinivasan and Blake Wulfe and Chen Xu and Mengchao Zhang and Alex Alspach and Maya Angeles and Kushal Arora and Vitor Campagnolo Guizilini and Alejandro Castro and Dian Chen and Ting-Sheng Chu and Sam Creasey and Sean Curtis and Richard Denitto and Emma Dixon and Eric Dusel and Matthew Ferreira and Aimee Goncalves and Grant Gould and Damrong Guoy and Swati Gupta and Xuchen Han and Kyle Hatch and Brendan Hathaway and Allison Henry and Hillel Hochsztein and Phoebe Horgan and Shun Iwase and Donovon Jackson and Siddharth Karamcheti and Sedrick Keh and Joseph Masterjohn and Jean Mercat and Patrick Miller and Paul Mitiguy and Tony Nguyen and Jeremy Nimmer and Yuki Noguchi and Reko Ong and Aykut Onol and Owen Pfannenstiehl and Richard Poyner and Leticia Priebe Mendes Rocha and Gordon Richardson and Christopher Rodriguez and Derick Seale and Michael Sherman and Mariah Smith-Jones and David Tago and Pavel Tokmakov and Matthew Tran and Basile Van Hoorick and Igor Vasiljevic and Sergey Zakharov and Mark Zolotas and Rares Ambrus and Kerri Fetzer-Borelli and Benjamin Burchfiel and Hadas Kress-Gazit and Siyuan Feng and Stacie Ford and Russ Tedrake},
journal= {arXiv preprint arXiv:2507.05331},
year = {2025}
}