VLAOpen

SpatialVLA

Shanghai AI Laboratory / collaborators

A spatially enhanced 4B VLA pretrained on 1.1 million real-robot episodes with explicit geometry-aware representations.

Code: Available
Weights: Available
Checkpoint: Available
License: mit

Weights Code Paper

Decision summary

Best for

Spatial grounding

Main blocker

Hardware, latency, dependencies, and checkpoint loading are not pipeline-tested.

Next check

Run a checkpoint loading smoke test before marking the model pipeline-tested.

Evidence profile

A visual read of adoption evidence. Scores describe catalog evidence readiness, not task performance.

Published evaluation

5/6

dimensions scored

Unknown evidence remains visible and is never treated as a zero.

Access and governance75conf. 85

Artifact availability90conf. 90

Training and loading reproducibility70conf. 65

Training data requirements76conf. 70

Evaluation evidenceNot scoredconf. 10

Deployment readiness70conf. 55

Why these scores

The strongest decision reasons behind the evidence profile.

Top 3 adoption signals

Artifact availability

Useful

Checkpoint files were verified in the official repository metadata.

90 · confidence 90

Training and loading reproducibility

Useful

Configuration files are present; runtime loading is not tested.

70 · confidence 65

Evaluation evidence

Unknown

No structured evaluation evidence has been verified.

— · confidence 10

Data requirements

Declared loop-data needs, missing evidence, and the next checks that matter.

4 required categories

Observation / ego video

observation · depth

required

Language intent / task phase

language intent

required

Action / robot state

actions · robot state

required

Sim-real / embodiment metadata

depth · robot state

required

Critical gaps

Core categories are represented. Interface alignment and data quality still require verification.

Linked datasets

Signal-level links only. Verify runtime interfaces before use.

Browse datasets

Open X-Embodiment

Cross-embodiment VLA pretraining

Useful

provisional fit 100 · confidence 60

4 matched · 0 unknown

RH20T

human_demo

Useful

provisional fit 100 · confidence 60

4 matched · 0 unknown

BridgeData V2

scripted

Useful

provisional fit 88 · confidence 45

3 matched · 1 unknown

More evidenceShow

Additional decision dimensions

Access and governanceUseful

Access and license are declared by the source.

75 · c85

Training data requirementsUseful

Required signal categories are structured; exact tensor and action interfaces still need verification.

76 · c70

Deployment readinessUseful

Hardware, latency, dependencies, and checkpoint loading are not pipeline-tested.

70 · c55

Artifact facts

release_year: official_claim; 2025; curated official source year
model.required_signals: official_claim; observation · depth · language intent · actions · robot state; Official model documentation and curated signal mapping
repository.code: metadata_verified; repository: SpatialVLA/SpatialVLA · default branch: main · archived: false · code file count: 25 · sample files: data/dataset.py · data/obs_transforms.py · data/oxe/__init__.py · data/oxe/configs.py · data/oxe/mixtures.py · data/oxe/transforms.py · data/rlds.py · data/traj_transforms.py · data/utils/data_utils.py · data/utils/droid_utils.py · data/utils/goal_relabeling.py · data/utils/task_augmentation.py · model/__init__.py · model/action_tokenizer.py · model/configuration_spatialvla.py · model/modeling_gemma2.py · model/modeling_spatialvla.py · model/processing_spatialvla.py · pyproject.toml · requirements.txt; GitHub repository metadata and recursive tree
model.hub: official_claim; declared; curated model hub link
model.weight_files: metadata_verified; repository: IPEC-COMMUNITY/spatialvla-4b-224-pt · count: 2 · sample files: model-00001-of-00002.safetensors · model-00002-of-00002.safetensors; Hugging Face model metadata siblings
model.checkpoint_files: metadata_verified; repository: IPEC-COMMUNITY/spatialvla-4b-224-pt · count: 2 · sample files: model-00001-of-00002.safetensors · model-00002-of-00002.safetensors; Hugging Face model metadata checkpoint files
model.config_files: metadata_verified; repository: IPEC-COMMUNITY/spatialvla-4b-224-pt · count: 5 · sample files: config.json · generation_config.json · preprocessor_config.json · processor_config.json · tokenizer_config.json; Hugging Face model metadata siblings
license.weights: metadata_verified; mit; Hugging Face model card license metadata
license: metadata_verified; mit; Hugging Face model card license metadata
release_timing: metadata_verified; 2025-01-27T07:34:33.000Z; arXiv 2501.15830 published
discovery.source: secondary_claim; name: worldbench/awesome-embodied-data-pyramid · trust: discovery_only · revision: 0568599d0619f20946e8089f6a941d0d9e30b690; #table-embodied-foundation-models-vla-wam, table 17, row 9
discovery.raw_fields: secondary_claim; Time: 2025.1 · Method: SpatialVLA · Institution: Shanghai AI Lab · Project: [![Website](https://img.shields.io/badge/Website-0A7DBD)](https://spatialvla.github.io/) · Model: VLA · Data: ![Real][data-real]; #table-embodied-foundation-models-vla-wam, table 17, row 9
model.reported_release_time: secondary_claim; 2025.1; #table-embodied-foundation-models-vla-wam, table 17, row 9
model.type: secondary_claim; VLA; #table-embodied-foundation-models-vla-wam, table 17, row 9
model.training_data_layers: secondary_claim; real_robot; #table-embodied-foundation-models-vla-wam, table 17, row 9
model.institution: secondary_claim; Shanghai AI Lab; #table-embodied-foundation-models-vla-wam, table 17, row 9
catalog.curation: secondary_claim; tier: editorial_focus · collection: WorldBench Awesome Embodied Data Pyramid · policy: human_curated_priority; #table-embodied-foundation-models-vla-wam, table 17, row 9

Additional references

ProjectProject

OpenBot notes

Official training references Open X-Embodiment and RH20T.

Related models

being_h05_baseVLA

lerobot

being_h05_robocasaVLA

lerobot

pi052_baseVLA

lerobot

lingbot-video-dense-1.3bVLA

robbyant