PaperlayerView exact revision on arXiv

Source-linked AI summary

A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation

TRI LBM Team, Jose Barreiros, Andrew Beaulieu, Aditya Bhat, Rick Cory, Eric Cousineau, Hongkai Dai, Ching-Hsin Fang, Kunimatsu Hashimoto, Muhammad Zubair Irshad, Masha Itkina, Naveen Kuppuswamy, Kuan-Hui Lee, Katherine Liu, Dale McConachie, Ian McMahon, Haruki Nishimura, Calder Phillips-Grafflin, Charles Richter, Paarth Shah, Krishnan Srinivasan, Blake Wulfe, Chen Xu, Mengchao Zhang, Alex Alspach, Maya Angeles, Kushal Arora, Vitor Campagnolo Guizilini, Alejandro Castro, Dian Chen, Ting-Sheng Chu, Sam Creasey, Sean Curtis, Richard Denitto, Emma Dixon, Eric Dusel, Matthew Ferreira, Aimee Goncalves, Grant Gould, Damrong Guoy, Swati Gupta, Xuchen Han, Kyle Hatch, Brendan Hathaway, Allison Henry, Hillel Hochsztein, Phoebe Horgan, Shun Iwase, Donovon Jackson, Siddharth Karamcheti, Sedrick Keh, Joseph Masterjohn, Jean Mercat, Patrick Miller, Paul Mitiguy, Tony Nguyen, Jeremy Nimmer, Yuki Noguchi, Reko Ong, Aykut Onol, Owen Pfannenstiehl, Richard Poyner, Leticia Priebe Mendes Rocha, Gordon Richardson, Christopher Rodriguez, Derick Seale, Michael Sherman, Mariah Smith-Jones, David Tago, Pavel Tokmakov, Matthew Tran, Basile Van Hoorick, Igor Vasiljevic, Sergey Zakharov, Mark Zolotas, Rares Ambrus, Kerri Fetzer-Borelli, Benjamin Burchfiel, Hadas Kress-Gazit, Siyuan Feng, Stacie Ford, Russ Tedrake

arXiv:2507.05331v1cs.RO

TL;DR

Reliable evaluation of generalist robot policies remains limited despite growing interest in Large Behavior Models. This paper rigorously compares multitask-pretrained LBMs with single-task baselines using controlled simulation and real-world trials, finding improved data efficiency, performance, and robustness under distribution shift. The study also reports smoothly improving finetuned performance as pretraining data increases, while identifying statistical power and measurement noise as important considerations.

  • Problem

    Real-world evidence about the capabilities and robustness of Large Behavior Models remains limited, making rigorous evaluation of these multitask robot policies important.

  • Method

    The paper evaluates LBMs trained on diverse robot data against single-task policies using controlled simulation, blind A/B-style real-world trials, and success-rate and task-completion metrics.

  • Results

    LBM pretraining reduces finetuning data requirements, improves aggregate performance over single-task models, and increases robustness under distribution shift.

  • Takeaways & Limitations

    Finetuned LBM performance smoothly improves with increasing pretraining data, supporting large-scale diverse pretraining as a viable path toward more capable robots.

  • Takeaways & Limitations

    The analysis does not explicitly account for stochasticity across training when comparing policy architectures, and real-world noise may obscure small effects.

Abstract

from arXiv · show

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/

I. Introduction

Large Behavior Models address brittleness in single-task robot policies by using diverse multitask pretraining, but their real-world capabilities require rigorous evaluation. This paper evaluates LBMs against single-task baselines across controlled simulation and real-world experiments, testing data efficiency, performance, and robustness to distribution shift.

  • Motivation: Single-task behavior-cloned policies remain brittle when task variations or environments differ from their training distributions.LBMs use large-scale multitask datasets of action-level demonstrations to improve reliability through broader learned visual and sensory representations.
  • Evaluation scope: The study trains LBMs on roughly 1,700 hours of demonstrations and evaluates them across simulation and 1,800 controlled real-world trials.The evaluation includes complex multi-step tasks involving tool use and precise manipulation.
  • Main findings: LBM pretraining reduces task-specific data requirements, improves aggregate performance over single-task training, and increases robustness under distribution shift.Finetuned specialists can match single-task performance with fewer demonstrations and outperform single-task models given the same task-specific data.
  • Evaluation protocol: The protocol uses blind A/B testing, large trial sizes, statistical analysis, qualitative and quantitative metrics, and controlled initial conditions.Performance is measured with success rate and task completion, with task completion capturing intermediate milestones that success rate alone misses.
  • Interpretation: Evaluation emphasizes relative success rates because absolute rates depend strongly on task difficulty, initial-condition distributions, and the number of demonstrations.Experiments target policy success rates around 50% to make relative differences informative, although observed rates vary substantially.
  • Statistical reporting: The reported violin plots summarize means and uncertainty or task-completion distributions, while Compact Letter Displays mark statistically distinguishable policies.The plotted distributions reflect evaluation uncertainty for policy checkpoints but do not capture randomness from training.

A. LBM performance on “seen” tasks

On seen tasks, finetuned LBMs generally outperform single-task baselines across real-world and simulation evaluations, including distribution shift. They also show data efficiency, while performance varies with task difficulty, demonstrations, and simulation conditions.

  • Seen-task comparisons: Finetuned LBMs statistically outperform single-task baselines in aggregate under nominal and distribution-shift conditions in simulation and the real world.The finetuned LBM is statistically distinguishable from single-task policies in all aggregate plots.
  • Seen-task comparisons: 3/3 real-world tasks and 15/16 simulation tasks show finetuned LBMs statistically equal to or better than single-task policies under both conditions.Under distribution shift, finetuned LBMs statistically outperform single-task policies in 2/5 real-world tasks and 10/16 simulation tasks.
  • Task variability: Performance is uneven across Breakfast simulation tasks because some tasks have only 49 demonstrations, while PushCoasterToMug requires obstacle removal and height-sensitive pushing.The scenario has higher cross-task variance and three tasks with substantially lower performance.
  • Task variability: Simulation timeouts can constrain apparent performance: TurnMugRightsideUp succeeds more often in the real world than simulation across policies.A short simulation timeout is proposed as one hypothesis for this discrepancy.
  • Unseen-task comparisons: Finetuned LBMs outperform single-task baselines on unseen tasks across real-world and simulation evaluations, measured by success rate and task completion.The aggregate comparison is statistically better under nominal conditions and distribution shift.
  • Unseen-task comparisons: 4/5 real-world tasks and 4/5 simulation tasks show statistically higher task completion for finetuned LBMs under nominal and distribution-shift conditions.Task-completion comparisons make the advantage clearer because finetuned policies achieve more steps on the task.
  • Data efficiency: Less than 30% of the task-specific simulation data is needed for LBM finetuning to achieve similar aggregate performance to training from scratch.For task completion, finetuned LBMs are statistically better at every tested data fraction; for success rate, the advantage begins at 50% of the data.
  • Data efficiency: 15% of SetBreakfastTable data is sufficient for a finetuned LBM to statistically outperform a single-task baseline trained with all the data.This real-world result further supports the simulation findings on data efficiency.

C. Pretraining scaling laws

The paper examines how pretraining scale and diversity affect LBM performance on unseen simulation tasks. More extensive pretraining consistently improves task completion, especially when task-specific finetuning data is limited.

  • C. Pretraining scaling laws: The study measures pretraining scaling through task completion on five unseen simulation tasks under nominal conditions.Success rates are low and not statistically distinguishable for these complex tasks, so the analysis focuses on task completion.
  • C. Pretraining scaling laws: With 15% finetuning data, all five models are statistically separable, and performance steadily increases as more pretraining tasks are added.The same trend holds with 50% and 100% pretraining data.
  • C. Pretraining scaling laws: All models achieve their best performance when pretrained and finetuned with all available data.The comparison includes the full dataset, TRI-Ramen alone, and reduced subsets of TRI-Ramen.
  • C. Pretraining scaling laws: More pretraining data benefits finetuned LBM performance when task-specific finetuning data is limited.Conversely, when task-specific data is plentiful, LBMs pretrained with less data might suffice.
  • C. Pretraining scaling laws: The experiment identifies a tradeoff between pretraining data and task-specific finetuning data.This conclusion concerns the amount of data used in the two training stages.

1) Policy comparison protocol:

The evaluation protocol compares policies under controlled, repeatable conditions while quantifying both complete success and partial task progress. It combines matched initial conditions, randomized blind trials, automated or rubric-based assessment, and statistical uncertainty analysis.

  • 1) Policy comparison protocol:: Simulation comparisons use identical parameters and seeded initial conditions, with new samples rather than reused training scenes.Out-of-distribution tests explicitly depart from the training distribution.
  • 1) Policy comparison protocol:: Figure 6 compares single-task baselines with finetuned LBMs under distribution shift using success rate in the top row and task completion in the bottom row.The x-axis identifies the task, scenario, and number of demonstrations; different letters indicate statistically distinguishable policies.
  • 1) Policy comparison protocol:: Real-world evaluations use policy-sized test bundles with randomized ordering and reset scenes to reduce environmental and ordering effects.The evaluator remains unaware of which policy is being tested during each run.
  • 1) Policy comparison protocol:: Real-world task completion assigns one credit per completed milestone and divides total credits by the number of milestones.Milestones represent necessary steps toward a successful rollout.
  • 2) Rubrics and predicates:: Simulation success is determined by predicates over simulator state, which also expose partial success and policy failure modes.A rollout succeeds when the required predicates are true.
  • 1) Policy comparison protocol:: Rubric quality assurance on approximately 27% of nearly 2,700 real-world rollouts found 2.31% success-rate discrepancy and 6.25% overall question discrepancy.A separate QA group reviewed the subset, and five rollouts were corrected before reporting.
  • 4) Statistical Analysis: Performance Characterization of Individual Policies:: Success-rate uncertainty treats outcomes as Bernoulli trials, while task-completion analysis presents raw categorical distributions when the mean does not capture the full distribution.Bayesian analysis and hypothesis tests supplement individual-policy uncertainty and comparisons.

5) Statistical Analysis: Performance Comparison of Multiple Policies:

The evaluation compares multiple policies with pairwise hypothesis tests, multiplicity correction, and compact letter displays. Policies use a diffusion-based visuomotor architecture that denoises actions conditioned on observations.

  • Statistical comparison: k(k−1)/2 pairwise tests compare k policy models, with Bonferroni correction maintaining a global 95% confidence level.The Compact Letter Display algorithm summarizes the resulting comparisons.
  • Statistical comparison: Binary success/failure outcomes use sequential hypothesis testing across paired evaluation trials.The procedure continues until reaching a decision or exhausting all trials.
  • Statistical comparison: When multiple tasks are plotted together, individual tests are not further adjusted, so Type-I error is not globally controlled across tasks.Overall multitask performance is nevertheless aggregated across tasks, under an exchangeability approximation that was empirically verified not to affect validity.
  • Diffusion policy: The policy generates actions by iteratively denoising noisy action samples conditioned on visual, proprioceptive, and language observations.The noise-prediction network is parameterized as a Diffusion Transformer.
  • Diffusion policy: The Diffusion Transformer conditions on observation features and diffusion timestep to predict actions.This network provides the learned noise prediction used in the denoising process.

2) Policy Architecture:

The LBM uses a Diffusion Transformer that combines visual, language, proprioceptive, and diffusion-timestep features to predict action sequences. Policies are pretrained broadly, then finetuned on individual tasks for deployment on bimanual tabletop robots.

  • Policy architecture: Visual features come from a pretrained CLIP ViT, language features from a CLIP text encoder, and both are combined with proprioception and the diffusion timestep.The combined features condition action prediction at each observation timestep.
  • Policy architecture: The model has eight DiT blocks with embedding size 768 and predicts 16 timesteps of 20-dimensional actions.Its total output size is 320 action values.
  • Training and deployment: The training recipe first pretrains on the full data mixture and then finetunes on narrower task-specific subsets.This follows a foundation-model-style pretraining and finetuning pattern.
  • Training and deployment: Pretraining uses 48k steps with global batch size 2560 and learning rate 3e-4, while the vision encoder uses one tenth of that rate.Images are resized and augmented with random crops and color jitter during pretraining.
  • Training and deployment: Finetuning runs for 30k real-task steps or 10k simulated-task steps with global batch size 320 and learning rate 2e-5.Checkpoint selection differed between simulated and real tasks, and several alternative training strategies remain future work.
  • Training and deployment: During deployment, the policy runs at 10 Hz and executes eight predicted action timesteps before recomputing actions.Training predicts 16 action steps, but deployment uses only the first eight before replanning.
  • Robot platform: Real-world evaluation uses two Franka FR3 arms with parallel grippers and TRI finray-style fingers across nine robot stations.All reported real-world evaluations use the newer hardware platform after a gripper and wrist-camera upgrade.
  • Evaluation platform: The simulation benchmark supports higher-throughput, controlled evaluation because relying only on real-world trials is expensive and time-consuming.The benchmark is built on Drake with curated assets and human-authored scenarios.

2) Simulation:

The simulation benchmark evaluates policies across kitchen-oriented scenarios that vary in complexity, objects, hardware, and task novelty. It combines TRI and external robot data while testing both seen and unseen tasks under controlled conditions.

  • Simulation scenarios: Four simulation scenarios vary in task complexity, object types, and object count within workspaces modeled after real robot stations.All scenarios focus on food preparation or organization and include task-relevant manipulands plus distractors.
  • Simulation scenarios: DryingRack has 13 tasks, Shelf 12, Breakfast 18, and Kitchen 5 tasks.Kitchen is absent from pretraining, contains only long-horizon tasks, and includes nonprehensile manipulation or semantic-attribute requirements.
  • Pretraining data: Ramen contains approximately 1695 hours of demonstrations, combining approximately 545 TRI-Ramen hours with approximately 1150 OXE-Ramen hours.The mixture combines internally collected and curated external robot data.
  • Pretraining data: TRI-Ramen contains 545 hours of real data across 532 tasks and 64,262 demonstrations.Its components include real, simulated, and Universal Manipulation Interface data collected across multiple stations or devices.
  • Evaluation design: For each of 40 simulation tasks in scenarios D, S, and B, corresponding real-world demonstrations were collected in similar environments.Demonstrations were distributed across two to four robot stations per task, with one station matching the simulated environment’s modeled station.
  • Pretraining data: OXE-Ramen is a heuristically selected OpenX-Embodiment subset mapped into the TRI-Ramen data format.The mapping standardizes coordinate frames, units, end-effector poses, gripper widths, image resizing, and cropping.
  • Pretraining data: Ramen batches are balanced using empirically determined dataset weights so every batch contains samples from all datasets.This controls the mixture presented during pretraining.

2) Data Normalization:

The evaluation spans nominal and shifted conditions in simulation and the real world, including appearance, station, and object changes. Data processing includes percentile normalization and filtering decisions that affect different policy types differently.

  • Data normalization: Timestep-specific normalization preserves more resolution for near-future actions, whose distributions are narrower and more important to predict accurately.Actions are represented relative to current observations, so farther-future actions have wider spread.
  • Data normalization: An implementation error caused some pretraining datagrams to use normalization parameters from a different data source.A smaller simulation experiment found little difference under nominal conditions, while correct parameters performed better under distribution shift.
  • Data preprocessing: Low-motion frames at demonstration starts were filtered using a threshold of more than 5 cm translation or 15° rotation from the starting pose.The filtering targeted operator or logging delays before meaningful motion began.
  • Data preprocessing: Filtering improved simulated single-task performance but decreased pretrained-LBM performance by increasing uncommanded task commitments.The authors therefore pretrained LBMs on unfiltered data while finetuning LBMs and training single-task policies on filtered simulation data.
  • Evaluation conditions: Simulation evaluation uses nominal and distribution-shift conditions, with 200 initial conditions per task and 50 per real-world task.The task sets include seen and unseen tasks selected separately for simulation and real-world evaluation.
  • Evaluation conditions: Simulation distribution shifts alter lighting, environment maps, camera parameters, object textures, colors, and distractor objects.These changes primarily test robustness to appearance variation, with additional distractors in scenario B.
  • Evaluation conditions: Real-world evaluation covers three seen short-horizon tasks and five unseen long-horizon multistep tasks requiring diverse manipulation sequences.The CutAppleInSlices example requires tool use, retrieval, unsheathing, coring, and sequential slicing.
  • Evaluation conditions: Station shift evaluates tasks on robot stations absent from the task’s finetuning data while preserving nominal object and distractor initial conditions.This isolates cross-station transfer as the shift factor.

V. Discussion and Conclusion

The study finds that large-scale, diverse pretraining improves task-specific manipulation performance, data efficiency, robustness, and scaling behavior, while highlighting important evaluation and modeling limitations.

  • Evaluation and findings: 1,800 blind A/B-style real-world rollouts and over 47,000 simulation rollouts support rigorous evaluation of LBMs.The models were trained on roughly 1,700 hours of heterogeneous demonstration data.
  • Evaluation and findings: Finetuned LBMs consistently outperform from-scratch training or achieve equivalent performance with 3-5x less data.The advantage increases when deployment conditions differ from training conditions.
  • Scaling: Finetuned performance improves smoothly as pretraining data increases, with no observed discontinuities or sharp inflection points at examined scales.This describes the measured scaling regime rather than an extrapolation beyond it.
  • Model limitations: Nonfinetuned LBMs learn many tasks simultaneously but do not consistently outperform from-scratch single-task training.The authors associate this difficulty partly with language-steering brittleness in models using small language encoders.
  • Evaluation implications: Many reported effects required larger-than-standard samples and careful statistical testing, while data normalization could dominate architectural or algorithmic changes.The authors warn that insufficient statistical power may make robotics results reflect noise.
  • Limitations: The analysis does not explicitly account for stochasticity across training when comparing policy architectures.Confidence intervals cover evaluation variability but omit the training-level stochasticity term.
  • Limitations: Real-world measurements may miss small effects because setup, scoring, initial-condition, and hardware variability add noise.The evaluation uses reproducible initial conditions and 50 rollouts per task, policy, and condition, but these controls reduce throughput.
  • Scope: The study uses modestly sized CLIP-pretrained language encoders, so some findings may differ for larger VLAs, particularly language steerability.The authors expect broader findings may generalize but identify language steerability as a setting-sensitive aspect.

A. Robot Hardware

The supplementary evaluation spans upgraded hardware and matched simulation platforms, randomized initial conditions, distribution shifts, and predicate-based task completion measurements.

  • Robot Hardware: The newer hardware platform uses Schunk grippers and four wrist cameras, while the older riverway platform uses a Franka Hand and two wrist cameras.The platform transition also changed camera and finger hardware.
  • Robot Hardware: Simulation platforms are behaviorally matched to hardware stations without attempting exact sim-to-real correspondence.The simulation models station behavior and camera effects as a best-effort approximation.
  • Robot Hardware: Scenarios D and S model the older riverway station, while scenarios B and K model the newer salem station.The corresponding simulated platform is cabot for salem.
  • Evaluation Scope: The simulation evaluation uses 16 seen and 5 unseen tasks out of 44 total tasks.Task-specific demonstration counts and shift types are summarized in the supplementary tables.
  • Simulation Conditions: Simulation initial conditions are sampled deterministically from predefined distributions using the simulation seed.The distributions randomize manipuland count, shape, texture, color, and pose relative to arbitrary frames.
  • Simulation Conditions: The scenarios are visually ambiguous so policies cannot infer the task solely from appearance.Sample initial conditions illustrate nominal and shifted distributions across the scenarios.
  • Evaluation Metrics: Simulation success and task completion are computed from task-specific Boolean predicates logged densely throughout each rollout.Predicate trajectories support offline analysis of stateful task milestones.

D. Missing simulation rollouts

The supplementary evaluation documents incomplete simulation rollouts and practical sources of variability in real-world setup, monitoring, scoring, and throughput.

  • Missing simulation rollouts: Each simulation task targets 200 rollouts per policy and condition, but cloud execution can produce missing data.Table S4 lists task, policy, and condition combinations with fewer than 200 rollouts.
  • Real-world evaluation: Real-world rollout initialization uses camera overlays and homographic projection to align scenes across hardware stations.The method works best for objects near the table plane.
  • Real-world evaluation: Tall objects can produce severe overlay artifacts, and visibility differences across stations can make some configurations unobservable.These issues complicate manual alignment and reproducible scene setup.
  • Real-world evaluation: Operators terminate real-world rollouts after success, dangerous behavior, prolonged lack of progress, or repetitive looping.These criteria rely more on human judgment and are more forgiving than simulation criteria.
  • Real-world evaluation: Real-world evaluation is repetitive and laborious, with room for human error, although QA found an overall low discrepancy rate.Success flags and rubric questions were compared against a secondary set.
  • Long-horizon tasks: The benchmark includes five long-horizon tasks, including BikeRotorInstall, CutAppleInSlices, and SetBreakfastTable.Successful hardware rollout film strips are provided for these examples.

C. Sample real-world initial conditions

The real-world benchmark varies task scenes, stations, and objects, and evaluates complex behaviors using manually scored milestone rubrics alongside qualitative rollout inspection.

  • Sample real-world initial conditions: Real-world distribution shifts include station changes and object changes, with sample conditions shown for PutKiwiInCenterOfTable and TurnMugRightsideUp.The supplementary figures compare nominal and shifted initial conditions.
  • Distribution shifts: Object-centric shifts vary color, texture, and shape, and are especially challenging for mug and coaster tasks.Kiwi tasks exhibit relatively less variation in the shifted setting.
  • Task scoring: Unseen real-world tasks are scored with manually evaluated Yes/No rubric questions, each representing a milestone toward completion.The corresponding quantitative results are reported elsewhere in the paper.
  • Task examples: The benchmark includes complex bimanual tasks such as clearing a kitchen counter, setting a breakfast table, and installing a bike rotor.Their rubrics enumerate sequential manipulation milestones.
  • Task examples: CutAppleInSlices requires sequential coring, knife slicing, wiping, and sheathing actions.The rubric records these intermediate steps as task-completion milestones.
  • Distribution shifts: Simulation samples initial conditions from nominal, station-shift, and object-centric-shift distributions for the PutKiwiInCenterOfTable task.Overlays visualize the aggregate initial-condition distributions.
  • Seen-task results: For seen tasks, success rate correlates with demonstration count: 19.5% and 13.5% with 49 demonstrations versus 50% and 68.5% with 196 and 490 demonstrations.The reported tasks are PutBananaOnSaucer, PutKiwiInCenterOfTable, PlaceCupByCoaster, and TurnCupUpsideDown, respectively.
  • Seen-task analysis: Terminal frames and first-motion timing provide qualitative and quantitative views of execution quality and delayed task initiation.Delayed first motion can lead to timeouts.

A. PutKiwiInCenterOfTable and PutBananaOnSaucer

These rollout analyses show that single-task policies usually attempt the correct task but fail during grasping or placement, while LBMs commonly remain stationary or manipulate the wrong object. Finetuning reduces these LBM failure modes in some cases, but motion initiation and task-selection errors remain important limitations.

  • Failure modes: Single-task policies consistently attempt the correct task, but often miss the target or fail to place it correctly.These failures are reported across the PutKiwiInCenterOfTable and PutBananaOnSaucer rollouts.
  • Failure modes: Pretrained and finetuned LBMs commonly fail by remaining at the starting pose or performing a different task.Most rollouts with either failure mode end unsuccessfully.
  • Finetuning effects: Finetuned LBMs stay static less often and perform the correct task more often than pretrained LBMs.The comparison is based on inspection of rollout behavior rather than a reported aggregate metric.

XIII. Additional pretraining dataset details

The additional experiments examine training-data preprocessing and normalization effects on LBM behavior. Filtering low-motion frames improves motion initiation, while the normalization error has little nominal effect but affects distribution-shift comparisons.

  • Normalization: For nominal seen-task simulation evaluations, the normalization error did not significantly affect corrected pretrained-LBM performance.The comparison used the corrected model against the other policies.
  • Normalization: Under distribution shift, the pretrained LBM with the normalization error outperformed the corrected version on 4/16 tasks and in aggregate.This comparison differs from the nominal-condition result.
  • Low-motion data: After low-motion filtering, both pretrained and finetuned LBMs initiate motions much faster across 16 simulation seen tasks.The improvement concerns time to first motion, not necessarily overall task performance.
Loading 2507.05331v1…