Open Models Almanac
ModelsVLA Foundry (Foundry-VLA-1.7B)

Toyota Research Institute

VLA Foundry (Foundry-VLA-1.7B)

What VLA Foundry (Foundry-VLA-1.7B)’s own documents say it is built from — every method with the sentence that describes it.

Parameters
1.7B
Attention
LLM -> VLM -> VLA trained from scratch in one framework; diffusion/flow-matching policy; Qwen3-VL variant (2.1B)
Input
text+image -> action
Licence
Apache-2.0
Released
2026-04-17

Against the consensus recipe

Its documents state 0 of the 18 methods the field agrees on.

The recipe from the overview: the methods the most labs adopt, per stage. A missing one is something its documents do not say, not something the model lacks.
  1. Data curation

  2. Training objective

  3. Model architecture

  4. Optimization

  5. Post-training

  6. Inference & serving

prediction head 2

Flow-matching action head core

“Foundry-VLM-1.3B vision-language backbone + (condition on last 4 layers) flow-matching diffusion action head (24 layers, 1024 hidden dim…”

Diffusion-policy action head optional

“while additional heads such as a diffusion policy are already implemented in the codebase”

multimodal architecture 2

Last-four-layer action-head conditioning core

“(condition on last 4 layers)”

Pretrained Qwen3-VL backbone initialization used

“we also train a VLA with the pretrained Qwen3-VL 2B model Bai et al., 2025 as backbone.”

1 more mentioned or evaluated
Autoregressive discrete action-tokenization head mentioned

prediction head “integrating further variants – for example, autoregressive discrete action tokenizations – requires only a new head module”

filed at the root 3

Sample-based training budgets core

“It expresses training budgets as a number of samples instead of a number of steps, so that runs at different batch sizes or GPU counts…”

Gradient checkpointing used

“The training loop supports the standard levers for scaling distributed training – FSDP (with optional CPU offloading), mixed precision…”

Gradient accumulation used

“The training loop supports the standard levers for scaling distributed training – FSDP (with optional CPU offloading), mixed precision…”

training precision 1

Mixed-precision training used

“The training loop supports the standard levers for scaling distributed training – FSDP (with optional CPU offloading), mixed precision…”

training stability 1

Action and proprioceptive-state normalization core

“Actions and proprioceptive states are normalized during dataloading time and denormalized at inference time.”

training parallelism 2

FSDP2 with WebDataset streaming core

“Built on FSDP2 with WebDataset streaming.”

Fully Sharded Data Parallel optional

“The training loop supports the standard levers for scaling distributed training – FSDP (with optional CPU offloading)”

training runtime 3

Checkpointing complete data-loader state core

“dataloader state checkpointing for exact restarts”

Deterministic per-rank random-number-generator seeding core

“We rely on deterministic per-rank RNG seeding, dataloader state checkpointing for exact restarts”

Model-agnostic training loop core

“The training loop is deliberately model-agnostic. Unpacking a batch into inputs and targets, masking the loss, and reducing model outputs…”

1 more mentioned or evaluated
Exponential moving average of model weights unclear

“torch.compile, and exponential moving average of weights”

filed at the root 2

Past/future action-window sampling core

“During dataset preprocessing, each training sample is constructed around an anchor timestep within an episode.”

Standardized robotics data conversion core

“Robotics raw data can come from any source (simulation logs, real-robot recordings, etc.), as long as a converter knows how to read it and…”

data sourcing 2

Conversion to WebDataset tar shards core

“VLA Foundry has custom scripts to convert raw datasets to the WebDataset tar shards described above.”

Resampling training data used

“Training data: 102M samples from simulated and real-world bimanual manipulation tasks (with resample)”

data filtering 2

Normalize and concatenate configured proprioceptive fields used

“the fields listed in proprioception_fields are each extracted, normalized, and concatenated along the feature dimension”

Episode-boundary padding threshold filtering optional

“To avoid degenerate samples with excessive padding, samples whose required padding exceeds configurable thresholds (max_padding_left…”

data mixture & curriculum 7

Dataloader-time dataset mixing core

“Specify dataset sources and ratios at dataloader time for easy dataset mixing and batch balancing across modalities.”

Probabilistic multi-modal dataset mixing core

“Flexible multi-modal training with probabilistic dataset mixing and dataloaders that support text, image-caption, and robotics data”

Weighted dataset-proportion mixing core

“each dataset contributes its own shards, statistics, and modalities with weighted dataset proportions”

Weighted multi-dataset mixing core

“VLA Foundry natively supports dataset mixing with command-line arguments.”

Multi-task training on real and simulated robotics data default

“Unless otherwise noted, Foundry-VLA-1.7B and Foundry-Qwen3VLA-2.1B-MT are trained on a multi-task mixture of both real and simulation data”

Training on simulated and real-world manipulation data used

“Trained on both simulated and real-world manipulation data.”

Explicitly weighted batch balancing across datasets optional

“the --data.dataset_weighting parameter, which handles the batch balancing ratios; a 1:2:1 weighting corresponds to 25%/50%/25% of each…”

sequence packing 3

Per-modality preprocessing into WebDataset tar shards core

“Raw data is converted to WebDataset Breuel, 2020 tar shards through a per-modality preprocessing stage.”

Multi-stage robotics data preprocessing used

“For robotics datasets, this is done in multiple stages. First, it creates a frames folder in the output directory, where each sample is…”

Configurable sequence padding optional

“At episode boundaries, sequences are padded using a configurable padding strategy (copy, zero, or reflect).”

mid-training & continual pretraining 1

Sequential LLM-to-VLM-to-VLA training pipeline used

“the first trained fully from scratch through our LLM→VLM→VLA pipeline”

1 more mentioned or evaluated
Task-specific fine-tuning of a multitask policy evaluated

supervised fine-tuning “finetuning the multi-task model on single seen tasks further improves performance in aggregate.”

context management 1

Temporal action chunking with past-context conditioning used

“Including past timesteps allows the model to condition on recent action history and proprioceptive context; predicting multiple future…”

filed at the root 2

Runtime model and dataset registry core

“At runtime, create_model() looks up the registry by model.type and dispatches to the correct factory.”

Sequential composable WebDataset pipelines core

“The steps of the WebDataset pipeline are defined sequentially, and notably support both WebDataset built-in functions (e.g.…”

training framework 2

Shared model-agnostic training loop core

“Training in VLA Foundry is done in a single training loop that is shared across every model stage in the pipeline”

Training with pretrained Hugging Face backbones optional

“VLA Foundry supports both from-scratch training and pretrained backbones from Hugging Face.”

infrastructure service 1

Parallel preprocessing with Ray core

“Preprocessing runs in parallel with Ray Moritz et al., 2018”

filed at the root 5

6D continuous rotation representation core

“Rotations are represented in the 6D continuous rotation format Zhou et al., 2019 throughout, with conversion to and from matrices…”

Gram–Schmidt orthogonalization used

“with conversion to and from matrices performed via Gram–Schmidt orthogonalization.”

Absolute and relative action representations optional

“VLA Foundry supports both absolute and relative action representations, which are stored as separate fields in the dataset.”

Anchor-relative action representation optional

“Relative actions are computed with respect to the robot’s actual pose at the anchor timestep, i.e., the frame at which a prediction is made.”

Selectable action and proprioceptive-state normalization optional

“supports four normalization methods: standard deviation, min-max, and two percentile-based variants (percentile_1_99 and percentile_5_95).”

unfiled 23

Action chunking over past and future time steps core

“Actions are chunked Zhao et al., 2023 in a configurable window of past and future time steps around an anchor: the future portion…”

Causal past-and-current proprioception conditioning core

“proprioception uses only the past and current timesteps within the window (i.e., indices ), whereas actions span the full past-and-future…”

Causal restriction of proprioceptive observations core

“Proprioceptive observations are causally restricted to past and current time steps.”

Composable YAML-based configuration core

“Presets themselves are composable – a YAML file can inherit from others, so experiments are expressed by combining building blocks rather…”

Decorator-based model registration core

“Each model is registered via a @register_model decorator on its factory function”

Decorator-based self-registration of models and batch handlers core

“Self-registering models and batch handlers via decorators.”

Flow-matching action prediction objective core

“The output action sequence is trained with the flow-matching objective Lipman et al., 2022.”

Frozen configuration dataclasses core

“immutable frozen dataclasses that prevent hidden configuration changes at runtime”

Global, per-timestep, and percentile-based robotics data normalization core

“our RoboticsNormalizer supports global and per-timestep schemes, including percentile-based variants”

Hierarchical configuration precedence core

“Configurations follow a three-level precedence: command-line arguments override YAML preset files, which override dataclass defaults.”

Per-model batch handlers for batching, loss construction, and output reduction core

“Each model selects a BatchHandler – typically one of the shared handlers defined per modality or model type (LLM, VLM, and DP-VLA) – which…”

Unified LLM-VLM-VLA training stack core

“a shared training stack with end-to-end control, from language pretraining to action-expert fine-tuning”

Unified training framework for LLM, VLM, and VLA models core

“VLA Foundry instead provides a shared training stack with end-to-end control, from language pretraining to action-expert fine-tuning.”

Balanced per-task sample sizes for unbiased aggregate policy comparisons used

“we balance the per-task sample size for each policy to ensure that the aggregate represents an unbiased estimate of the policy’s…”

Distributed parallel preprocessing with Ray used

“We use ray Moritz et al., 2018 to parallelize data preprocessing.”

Merging t-digest sketches to estimate pooled percentiles used

“each dataset retains a serialized t-digest sketch Dunning and Ertl, 2019 during preprocessing, and the sketches are merged at training…”

PyTorch compilation with torch.compile used

“The training loop supports the standard levers for scaling distributed training – FSDP (with optional CPU offloading), mixed precision…”

t-digest percentile estimation and merging used

“For percentile estimation and merging, we use t-digest Dunning and Ertl, 2019.”

Configurable proprioceptive input fields optional

“Proprioceptive state is specified via a separate --proprioception_fields parameter, distinct from --action_fields.”

Merging normalization statistics across datasets optional

“Since datasets are processed individually with their own per-dataset stats.json files, we support merging multiple stats.json files…”

Per-timestep normalization optional

“In per-timestep scope, each timestep within the action window has its own mean and scale derived from statistics computed at that relative…”

Registering custom batch handlers for new training paradigms optional

“Entirely new training paradigms can register an additional handler via @register_batch_handler.”

1 more mentioned or evaluated
PyTorch compiler-based optimization unclear

“gradient checkpointing, torch.compile, and exponential moving average of weights”

Sources

The curated source list for this model. "not fetched" means the URL is recorded but its text was not read in the current run.

SourceKindPublisherStatus
huggingface.co/TRI-ML/Foundry-VLA-1.7B-fullmodel cardTRIread
arxiv.org/abs/2604.19728technical reportTRIread
tri-ml.github.io/vla_foundryofficial blogTRIread
↑