WorldFoundry
English中文

Embodied · vla · wam · vla.policy_rollout · wam.world_action_modeling

StarVLA

StarVLA

WorldFoundry runtime executed official StarVLA QwenOFT predict_action with local Qwen3-VL-OFT-LIBERO-4in1 checkpoint and Qwen3-VL-4B base VLM on an 80GB CUDA GPU.

Runner verifiedUnified environmentPinned sourcestarvla

Build your run

Choose a recorded variant, then copy the exact setup or inspection command.

worldfoundry.pipeline
Taskvla.policy_rollout
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
worldfoundry-eval evaluate \
  --mode model \
  --model-id starvla-qwen3-vl-oft-libero-4in1 \
  --model-runner worldfoundry:pipeline \
  --model-manifest-dir worldfoundry/data/models/catalog \
  --requests-path tmp/requests.jsonl \
  --output-dir tmp/model_eval/starvla-qwen3-vl-oft-libero-4in1 \
  --metric artifact_count \
  --json
01

Compatibility & versions

Manifest-backed recipe

Integration
integrated
Runner evidence
verified
Environment
worldfoundry-unified-cu128
Python
3.10
CUDA
CUDA 12.8
PyTorch
torch
Source revision
1947454be8c0f4de315f4cbde96b874819a6dab3
Checkpoint revision
1947454be8c0f4de315f4cbde96b874819a6dab3
Runtime profile
starvla
Pipeline binding
starvla
Runner
worldfoundry.pipeline
Pipeline target
worldfoundry.pipelines.component_pipelines:StarVLAPipeline
Backend stage
in_tree_checkpoint_runtime
Runtime status
in_tree_qwen3_vl_oft_libero_predict_action_gpu_ready
Driver status
compatible
Environment kind
Unified environment
02

Install environment

The environment resolver reads the recorded profile and chooses the unified or dedicated environment shown here.

bash scripts/setup/model_env_install.sh --model starvla-qwen3-vl-oft-libero-4in1
Package constraints 14
  • numpy
  • Pillow
  • opencv-python
  • torch
  • torchvision
  • transformers
  • accelerate
  • safetensors
  • einops
  • qwen-vl-utils
  • decord
  • omegaconf
  • diffusers
  • rich
Conda packages 2
  • python
  • pip
03

Checkpoints & assets

Run the local check before allocating compute. Gated, private, and license fields below come directly from the checkpoint manifest.

worldfoundry-eval zoo model-download --model-id starvla --check-local --json
StarVLA/Qwen3-VL-OFT-LIBERO-4in1
Revision
1947454be8c0f4de315f4cbde96b874819a6dab3
License
—
Gated
—
Private
—
StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1
Revision
e431aa1d059fd61864db707f3c178673af01c9d4
License
mit
Gated
—
Private
—
Qwen/Qwen3-VL-4B-Instruct
Revision
ebb281ec70b05090aa6165b016eac8ec08e71b17
License
—
Gated
—
Private
—
Wan-AI/Wan2.2-TI2V-5B-Diffusers
Revision
b8fff7315c768468a5333511427288870b2e9635
License
—
Gated
—
Private
—
04

Launch & outputs

The generated command uses the shared evaluation boundary and writes durable result manifests and artifacts.

worldfoundry-eval evaluate \
  --mode model \
  --model-id starvla-qwen3-vl-oft-libero-4in1 \
  --model-runner worldfoundry:pipeline \
  --model-manifest-dir worldfoundry/data/models/catalog \
  --requests-path tmp/requests.jsonl \
  --output-dir tmp/model_eval/starvla-qwen3-vl-oft-libero-4in1 \
  --metric artifact_count \
  --json

Input contract

FieldRecorded contract
actionsrobot_action, world_action, latent_action_tokens, action_space
imageRequired
instructionRequired
promptRequired
videoRequired

Artifact contract

Artifact kindFilename / path
action_trace—
action_tracestarvla_action_trace.json
05

Evidence & sources

Catalog integration, native-demo parity, and runner parity are independent records.

Integration
integrated
Runner evidence
verified
Native demo evidence
Not recorded
Validation imports
numpy, PIL, cv2, transformers, decord, diffusers, rich, qwen_vl_utils, omegaconf
Recipe provenance
worldfoundry/data/models/catalog/vla_va_wam/starvla.yamlGitHubhttps://github.com/starVLA/starVLAHugging Face1947454be8c0f4de315f4cbde96b874819a6dab3Hugging Facee431aa1d059fd61864db707f3c178673af01c9d4Hugging Facehttps://huggingface.co/Qwen/Qwen3-VL-4B-Instruct