yuuu94/GraphVLA-StackBowls-V7-4k

GraphVLA V7 Stack Bowls — 4,000 steps

Real-world dual-arm Stack Bowls model checkpoint at training step 4,000.

Files

• checkpoint_step_004000.pt: original DeepSpeed model-state checkpoint (model parameters and model metadata). No optimizer shards are uploaded.
• source_code.tar.gz: matching GraphVLA training/model code and internal Python dependencies, preserving project paths under code/.
• code_manifest.json: SHA256 for included source files.
• effective_config.yaml: effective training configuration; server-specific paths must be adapted.
• normalization.json: training-split state/action normalization statistics.

Model and data contract

Qwen3-VL-4B backbone, GR00T DiT-B action head and V7 graph conditioning. Four synchronized RGB views, 14-dimensional joint state and actions, action horizon 16. Graph visual features use mask-coverage-weighted pooling and edge GRU memory. Two GPUs, batch 24 per GPU (global batch 48). Dataset: 72 training and 9 validation trajectories from Stack_bowls_lerobot.

The code archive includes the verified four tiny-mask overlays used by this run. The original perception cache is not bundled. Dataset, robot assets, base-model tokenizer/processor files and perception model assets are not included. Obtain these dependencies separately and adapt paths before use. This is a research checkpoint, not a ready-to-run hardware deployment package.

Loading the checkpoint

Load the checkpoint only from this trusted repository with torch.load(path, map_location="cpu", weights_only=False). The model parameter dictionary is under the DeepSpeed module key. Construct the matching model from the archived code/configuration before loading it. The model checkpoint is not a Hugging Face Transformers from_pretrained export and does not contain the full optimizer state required for exact continuation.

Training snapshot

Step 4,000; BF16 training with ZeRO-2 and standard AdamW. Geometry targets remain in metres. Cached input geometry is estimated rather than simulator ground truth. This release reports no real-world deployment success rate.