SourcesTRI
publisher · 3 sources
TRI
3 sources read from this publisher, newest first, with the methods each one describes.
VLA Foundry: A Unified Framework for Training Vision-Language-Action Models
research paper · technical report · 2026-04-21 · 58 methods
Fully Sharded Data ParallelGradient checkpointingMixed-precision trainingCheckpointing complete data-loader stateDiffusion-policy action headExponential moving average of model weightsGradient accumulation6D continuous rotation representationAbsolute and relative action representationsAction and proprioceptive-state normalizationAction chunking over past and future time stepsAnchor-relative action representationAutoregressive discrete action-tokenization headBalanced per-task sample sizes for unbiased aggregate policy comparisonsCausal past-and-current proprioception conditioningCausal restriction of proprioceptive observationsComposable YAML-based configurationConfigurable proprioceptive input fieldsConfigurable sequence paddingConversion to WebDataset tar shardsDecorator-based model registrationDeterministic per-rank random-number-generator seedingDistributed parallel preprocessing with RayEpisode-boundary padding threshold filteringExplicitly weighted batch balancing across datasetsFlow-matching action prediction objectiveFrozen configuration dataclassesGlobal, per-timestep, and percentile-based robotics data normalizationGram–Schmidt orthogonalizationHierarchical configuration precedenceMerging normalization statistics across datasetsMerging t-digest sketches to estimate pooled percentilesModel-agnostic training loopMulti-stage robotics data preprocessingMulti-task training on real and simulated robotics dataNormalize and concatenate configured proprioceptive fieldsParallel preprocessing with RayPast/future action-window samplingPer-modality preprocessing into WebDataset tar shardsPer-model batch handlers for batching, loss construction, and output reduction
TRI-ML/Foundry-VLA-1.7B-full · Hugging Face
first party release · model card · 2026-04-20 · 4 methods
VLA Foundry
technical report · official blog · 2026-01-01 · 6 methods