Shanghai AI Laboratory et al.
SpatialVLA
4.03B Commercial use OK 8/8 verified
Facts
- Released
- 2025-01 1 Source
- Parameters
- 4.03 B 1 Source
- Architecture
- Monolithic autoregressive 0 Source
- License
- MIT (both code README and model card tagged MIT) 2 Source
- Commercial
- Commercial use OK 0 Source
- Weights
- huggingface.co ↗ 1 Source
- Code
- github.com ↗ 1 Source
- Paper
- arxiv.org ↗ 0 Source
- Input modalities
- rgb · language 0 Source
- Action space
- End-effector pose delta + gripper (adaptive discrete grid) 0 Source
- Cross-embodiment
- Open X-Embodiment + RH20T, 1.1M episodes in total 1 Source
- Pretraining data (hours)
- Unverified 0 Source
- Target forms
- Single arm 0 Source
Architecture · PaliGemma2-3B backbone + spatial position encoding and adaptive action grid; pretrained on 1.1M real-robot episodes
Sources · 2 sources
release_date,license,cross_embodiment,code_url https://github.com/SpatialVLA/SpatialVLA Official page · Fetched 2026-09-04 · No snapshot stored yet README: released 2025-01-29; MIT license; 1.1 Million real robot episodes from Open X-Embodiment and RH20T; Google Robot / WidowX / Franka
params_b,weights_url,license https://huggingface.co/api/models/IPEC-COMMUNITY/spatialvla-4b-224-pt Model hub page · Fetched 2026-09-04 · No snapshot stored yet safetensors 4,027,854,731；license: MIT；base_model google/paligemma2-3b-pt-224
Activity signals
Public counters fetched daily from GitHub, Hugging Face and ModelScope. Numbers only; no judgement. · fetched 2026-10-31
GitHub stars
734
+7 / 30 d
Commits · 90 d
0
last push 496 d ago
HF downloads · 30 d
543
all-time 157,745 · 14 likes
Weights last modified
2025-03-16
595 d ago
ModelScopeKoorye/spatialvla-4b-224-pt · 15 downloads · updated 2025-05-13 · same-name mirror, not the vendor org
Hugging FaceIPEC-COMMUNITY/spatialvla-4b-224-pt
stars · last 30 d
Public benchmark scores
As printed in the source; protocols differ between papers. Full table with notes on the benchmarks page.
| Benchmark | Subset | Score | Source |
|---|---|---|---|
| LIBERO | avg | 78.1 | paper · Table III, row 'SpatialVLA fine-tuned', LIBERO-Average |
| LIBERO | goal | 78.6 | paper · Table III, row 'SpatialVLA fine-tuned', LIBERO-Goal |
| LIBERO | long | 55.5 | paper · Table III, row 'SpatialVLA fine-tuned', LIBERO-Long |
| LIBERO | object | 89.9 | paper · Table III, row 'SpatialVLA fine-tuned', LIBERO-Object |
| LIBERO | spatial | 88.2 | paper · Table III, row 'SpatialVLA fine-tuned', LIBERO-Spatial |
| Real-robot evaluation (self-reported) | Close the microwave | 100 | paper · Table XI, row 'SpatialVLA', column 'Close the microwave' (Success) |
| Real-robot evaluation (self-reported) | Lift the red chili pepper | 72.7 | paper · Table XI, row 'SpatialVLA', column 'Lift the red chili pepper' (Success) |
| Real-robot evaluation (self-reported) | Put eggplant in the basket | 72.7 | paper · Table XI, row 'SpatialVLA', column 'Put eggplant in the basket' (Success) |
| Real-robot evaluation (self-reported) | Put green cup on the pink cloth (cup on stove) | 81.8 | paper · Table XI, row 'SpatialVLA', column 'Put green cup on the pink cloth (cup on stove)' (Success) |
| Real-robot evaluation (self-reported) | Put purple cup on pink cloth (cup in sink) | 90.9 | paper · Table XI, row 'SpatialVLA', column 'Put purple cup on pink cloth (cup in sink)' (Success) |
| Real-robot evaluation (self-reported) | Put purple cup on white plate | 45.5 | paper · Table XI, row 'SpatialVLA', column 'Put purple cup on white plate' (Success) |
| Real-robot evaluation (self-reported) | Put the carrot in the plate | 45.5 | paper · Table XI, row 'SpatialVLA', column 'Put the carrot in the plate' (Success) |
| SimplerEnv · Google RobotSpatialVLA (fine-tuning) | variant_aggregation | 70.7 | paper · Table I, row 'SpatialVLA (fine-tuning)', Variant Aggregation #Average |
| SimplerEnv · Google RobotSpatialVLA (zero-shot) | variant_aggregation | 68.8 | paper · Table I, row 'SpatialVLA (zero-shot)', Variant Aggregation #Average |
Deployment recipe
Environment, weights, run command and the pitfalls we hit, exactly as executed on rented GPUs (venv spatialvla, 8.5 GB of weights, 3 pitfalls).
Open recipe →Record history
2026-09-04
Sources fetched
release_date,license,cross_embodiment,code_url · params_b,weights_url,license
Latency & cost
cost = p50 ms ÷ 1000 × rental ÷ 3600 × 1000
| Hardware | p50 ms | p95 ms | Peak VRAM | Cost per 1k inferences ¥ | samples |
|---|---|---|---|---|---|
NVIDIA RTX 4090 ¥1.68/h · GongJi Compute (2026) | 376 | 405 | 7.82 GB | 0.176 | 1 Preliminary (n<3) Maintainer measured |
NVIDIA RTX 5090 ¥2.4/h · Runyun 2.4 / GongJi 2.5 / Paratera Cloud 2.98 (2026) | 362 | 370 | 7.82 GB | 0.242 | 1 Preliminary (n<3) Maintainer measured |
NVIDIA A100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA Jetson AGX Thor no rental price | — | — | — | — no latency data | No data |
NVIDIA H100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA A800 80GB no rental price | 451 | 473 | 7.82 GB | — no rental price for this hardware | 1 Preliminary (n<3) Maintainer measured |
Compatibility
Colored only where we hold evidence. An amber dot means the vendor declares this form as a target; that is a claim, not a verified compatibility.
| Unitree G1 | Unitree H1 | AgiBot Genie G1 | AgileX Piper arm | SO-101 | Franka FR3 / Panda | Fourier GR-2 | WidowX 250 S | Google Robot（Everyday Robots） | ALOHA 2 / Cobot Magic (dual-arm) | Galaxea R1 Pro | Franka Emika Panda | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SpatialVLA |
Legend Officially supportedCommunity reproducedFine-tunable in theoryUnsupportedUnknown Vendor-declared target form 3 / 12 cells with a definite status
Reproduction records
Did this model run on your robot? Report it with a link (log, video or repo). Records are reviewed by the maintainers; a verified “works” or “works after fine-tuning” upgrades the matrix cell to “community verified” with your link as evidence. Anonymous is fine; nothing but the fields below is stored.
No verified record yet for this model.
Clear direction, grounded judgement.