Open Source RoboticsRobots, software & learning

SpatialVLA

Robotics model

4B-parameter spatial-enhanced Vision-Language-Action model built on PaLiGemma2-3B with Ego3D Position Encoding (via ZoeDepth) and Adaptive Action Grids, pretrained on 1.1M real robot episodes and loadable directly via Hugging Face Transformers. [1] [2] [3]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotics researchers; VLA fine-tuning engineers; 3D-aware manipulation developers [2]
Parameters
4,027,854,731 (~4.03B BF16 parameters per safetensors metadata) [2]
Embodiments
WidowX 250 (Bridge/SimplerEnv), Google Robot (SimplerEnv), Franka Panda (LIBERO / Open X-Embodiment), … [2] [4]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

SpatialVLA (IPEC-COMMUNITY/spatialvla-4b-224-pt) is a 4,027,854,731-parameter (~4.03B) Vision-Language-Action model built on google/paligemma2-3b-pt-224 (SigLIP vision encoder and Gemma2-2B causal LM) augmented with a ZoeDepth monocular depth module (use_vision_zoe: true), Ego3D Position Encoding, and Adaptive Action Grids. It is pretrained on 1.1M real robot episodes from Open X-Embodiment and RH20T and loads through Hugging Face Transformers. [2] [3] [4]

Model card summary

Architecture

PaLiGemma2-3B (SigLIP vision encoder + Gemma2-2B causal LM) augmented with ZoeDepth monocular depth module (use_vision_zoe: true), Ego3D Position Encoding, and Adaptive Action Grid discrete spatial action tokens (spatial_token_num: 8194).

[2] [3]

Base model

google/paligemma2-3b-pt-224

[2]

Parameter count

4,027,854,731 (~4.03B BF16 parameters per safetensors metadata)

Verified via Hugging Face API safetensors metadata.

[2]

Input modalities
rgblanguage

Single RGB image (224x224; depth estimated internally via ZoeDepth) + English language prompt. Multi-view and state inputs are noted as under development on the lerobot branch.

[2] [3]

Output modalities
ee-pose-deltagripper-command

Discrete Adaptive Action Grid spatial action tokens decoded via dataset unnorm_key into continuous end-effector deltas and gripper commands.

[2] [3]

Action space

Discrete Adaptive Action Grid tokens (spatial_token_num: 8194 starting at vocab index 257153) decoded to continuous 7-DoF end-effector delta + gripper actions using per-dataset unnormalization keys (e.g., bridge_orig/1.0.0).

[2] [3]

Trained / supported embodiments
  • WidowX 250 (Bridge/SimplerEnv)
  • Google Robot (SimplerEnv)
  • Franka Panda (LIBERO / Open X-Embodiment)
  • RH20T multi-robot embodiments

[2] [4]

Training datasets

Pretrained on 1.1M real robot episodes from Open X-Embodiment and RH20T.

[2]

Interfaces
  • Hugging Face Transformers AutoModel/AutoProcessor (trust_remote_code=True)
  • Python training/fine-tuning scripts

[2] [3]

Checkpoints

CheckpointFormatSizeAccessPinned revision
IPEC-COMMUNITY/spatialvla-4b-224-ptsafetensors7.54 GBopen886c4557dfde

Ungated BF16 safetensors checkpoint (7.54 GB) verified via Hugging Face API.

[2]

Is this for you?

Use cases
  • Zero-shot evaluation on pretraining embodiments (e.g., Bridge/SimplerEnv)
  • Fine-tuning on custom robot manipulation datasets
  • Research on 3D spatial tokenization for VLAs

[2]

Known limitations
  • Does not zero-shot generalize to unseen robot embodiments not present in the pretraining mixture
  • Requires trust_remote_code=True in Hugging Face Transformers
  • GitHub repository root lacks a standalone LICENSE file (MIT is declared on the Hugging Face model card, while underlying PaLiGemma2-3B carries Gemma Terms of Use)

[1] [2]

Safety and limitations
  • Does not zero-shot generalize to unseen robot embodiments; requires fine-tuning on target hardware demonstrations.
  • Loading weights in Hugging Face Transformers requires trust_remote_code=True (config.json defines custom auto_map implementations); inspect repository code at the pinned revision before enabling.
  • GitHub repository lacks a root LICENSE file; separate upstream terms for google/paligemma2-3b-pt-224 and training datasets (Open X-Embodiment, RH20T) require independent verification.

[1] [2]

Compute & fine-tuning

Inference compute
  • Minimum VRAM: 8.5 GB
  • Precision: bf16
  • Edge / consumer capable: No

Model card explicitly reports 8.5 GB GPU memory requirement for BF16 inference; latency_ms not stated in model card.

[2]

Fine-tuning

[4]

Host computer

Requires 8.5 GB of GPU memory for BF16 inference per model card and README.

[2]

Operating system
  • Linux

README documents Linux apt/tcmalloc tip and CUDA inference.

[4] · Editorial assessment

Dependencies
  • Python
  • PyTorch
  • transformers >= 4.47.0
  • PIL

[2]

Skills
  • Python
  • PyTorch
  • Hugging Face Transformers

Inferred from documented Python/PyTorch/CLI setup and training dependencies.

[2] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [2] · Model card / documentation [1] [2]

  • Install transformers >= 4.47.0 and PyTorch
  • Load IPEC-COMMUNITY/spatialvla-4b-224-pt with AutoModel/AutoProcessor (trust_remote_code=True, bfloat16)
  • Call model.predict_action(inputs) and processor.decode_actions(outputs, unnorm_key=...)
  • Clone SpatialVLA/SpatialVLA for pretraining or fine-tuning

[2]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
LIBERO (Spatial, Object, Goal, Long — fine-tuned)88.2% Spatial, 89.9% Object, 78.6% Goal, 55.5% Long (78.1% avg)Upstream-reported simulation scores in SpatialVLA README; not independently verified by OSR.

[4]

Project notes

Spatial tokenization and runtime requirements

At pinned Hub revision 886c4557dfde, config.json defines SpatialVLAForConditionalGeneration with spatial_token_num: 8194 starting at vocabulary index 257153, decoding discrete spatial tokens into continuous 7-DoF end-effector delta and gripper commands using per-dataset unnormalization keys (such as bridge_orig/1.0.0). The model card documents a PyTorch/CUDA BF16 inference path requiring transformers >= 4.47.0 and 8.5 GB of GPU memory (separate from the repository's Python >= 3.10 training environment prerequisites). Because config.json registers custom configuration and model implementations via auto_map, loading requires trust_remote_code=True, which executes repository-provided Python code; as editorial guidance, inspect that code at the pinned revision before enabling the flag. Unseen robot embodiments require fine-tuning on target demonstrations rather than zero-shot execution. [2] [3] [4]

Upstream-reported simulation evaluations

The pinned GitHub README reports fine-tuned LIBERO simulation success rates of 88.2% Spatial, 89.9% Object, 78.6% Goal, and 55.5% Long (78.1% average), alongside links to SimplerEnv evaluations. All reported metrics are upstream-reported simulation figures; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for SpatialVLA and reports no independently measured latency or task accuracy. [4]

What we checked

None. Prepared from public upstream sources without hands-on hardware execution.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Jun 23, 2025 [1]
Latest release
Mar 16, 2025 [2]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
May 21, 2027
Reviewed revision
SpatialVLA/SpatialVLA@18fd74b2a633ec8d9ec7aadcd803969555cc9fbd; HF IPEC-COMMUNITY/spatialvla-4b-224-pt@886c4557dfde53b44c0f74fde7d89bf83a26c3cd · software repository

Separate license matrix

Inspectable 4B-parameter 3D-spatial VLA with ungated safetensors weights and Hugging Face Transformers integration; marked include_partial because GitHub repo lacks a root LICENSE file and weights inherit PaLiGemma2/Gemma Terms.

LayerAvailability / license
Codeavailable · MIT (declared on HF model card; no root LICENSE file in GitHub repo)
Verified via GitHub API that SpatialVLA/SpatialVLA has no root LICENSE file (license_spdx: null); HF model card states 'All SpatialVLA checkpoints, as well as our training codebase are released under an MIT License.' [1] [2]
Model weightsavailable · MIT (HF model card declaration; separate upstream terms of base model google/paligemma2-3b-pt-224 require independent verification)
HF model card declares license: mit; base_model is google/paligemma2-3b-pt-224, whose upstream terms were not excerpted in this review and require verification. [2]
Training datasetsavailable · Mixed upstream (Open X-Embodiment + RH20T; individual dataset terms not excerpted in this review)
[2]
Documentationavailable · MIT (card declaration)
[2]

Separate code, weights, base-model, and dataset licensing

The Hugging Face model card declares MIT licensing for both checkpoints and training code, but the SpatialVLA/SpatialVLA GitHub repository root lacks a LICENSE file (license_spdx: null via GitHub API). In addition, weights derive from google/paligemma2-3b-pt-224 and pretraining draws on Open X-Embodiment and RH20T; those upstream base-model and dataset license terms were not excerpted in this review and require independent verification (without assuming that dataset terms automatically transfer to model weights). [1] [2]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model draft prepared from pinned GitHub commit 18fd74b2a633 and HF revision 886c4557dfde.