GraphVLA V7 Stack Bowls — 6,000 steps
Dual-arm real-world Stack Bowls checkpoint. This package follows the W2-VLA release directory convention, with a GraphVLA-specific loader and joint14 mean/std statistics.
checkpoints/
graphvla_stackbowls_v7_step006000.pt
config.yaml
dataset_statistics.json
load_policy.py
code/
assets/Qwen3-VL-4B-Instruct/
release_manifest.json
normalization.json
Download and load
hf download yuuu94/GraphVLA-StackBowls-V7-6k --local-dir ./graphvla-stack-6k --exclude "checkpoint_step_006000.pt" "*.tar.gz"
python ./graphvla-stack-6k/load_policy.py --release-dir ./graphvla-stack-6k --device cpu
Use the matching GraphVLA Python dependencies from code/requirements.txt and the archived training environment. load_policy.py constructs the exact graph/action architecture, then strictly loads a model-only PyTorch state_dict (weights_only=True). The included Qwen3-VL tokenizer/processor/config metadata avoids downloading base weights during model construction. CPU loading validates parameters; it does not connect to or actuate a robot.
The original W2-VLA model class and its LIBERO normalization defaults are different. This release uses the included GraphVLA loader, not the original W2-VLA model loader.
Data and normalization
dataset_statistics.json uses dataset key stack_bowls. action and state each contain the actual training-split mean/std for 14 dimensions. All 14 action dimensions were normalized; action mask is therefore all true. Min/max or quantiles are not fabricated.
- State and action order: left 6 joints, left gripper, right 6 joints, right gripper.
- Joint units: radians. Gripper units: total aperture in metres.
- Actions: absolute joint14 targets, prediction horizon 16, deployment executes the first 8.
- State:
(raw - mean) / max(std, 1e-6). - Action output:
normalized * max(std, 1e-6) + mean, exactly once.
normalization.json is retained as the original training/deployment-format alias (action, observation.state). Its numbers match dataset_statistics.json. Training used 72 trajectories; 9 validation trajectories were excluded from these statistics.
config.yaml is the model-construction/inference configuration with relative release paths. Its execution.mode: test_only setting is inherited from the validated architecture constructor; it does not describe the original training run. effective_config.yaml records the actual training run.
Architecture and deployment
Qwen3-VL-4B + GR00T DiT-B; four RGB views and raw joint14 state; V7 mask-coverage-weighted graph pooling and edge GRU. Online deployment also needs the matching YOLOE/SAM3/Cutie/DA3 weights, robot assets/calibration and camera/CAN frontend. Those large perception weights and robot execution interfaces are not bundled here.
Model-only checkpoints do not contain optimizer state or an in-progress episode memory. They are for inference/initialization, not an exact optimizer resume. No real-world success rate is claimed for this release.
Versions
The 6k repository retains its original root-level DeepSpeed checkpoint, original source archive and original manifests for existing download links. The standardized inference entry is checkpoints/ plus config.yaml, dataset_statistics.json and load_policy.py.
Upstream
Built on Qwen3-VL-4B-Instruct, Apache-2.0. Use is subject to the applicable upstream model and software licenses.
- Downloads last month
- 11