Single-run supervised fine-tuning of a QwenGR00T Vision-Language-Action (VLA) model on the Robocasa365 target/atomic split, released with the AlphaBrain framework. Not a continual-learning run — all atomic-task data is trained jointly.

The nostate suffix indicates this variant was trained without feeding the robot proprioceptive state into the action decoder (only visual observations + language instruction). This matches the include_state: false training setup used for language-conditioned evaluation where privileged state should not leak into the policy.

Overview


Architecture	QwenGR00T (Qwen3-VL-4B + Flow-Matching DiT head)
Base VLM	`Qwen/Qwen3-VL-4B-Instruct`
Parameters	~4 B
Action head	DiT-B, `hidden_size = 2560`, 16 layers
Embodiment	panda_omron · action dim 12 · state dim 16
State input	disabled (`include_state: false`)
Training data	Robocasa365 `target/atomic` — all atomic tasks, jointly trained
Training type	Supervised fine-tuning (single run, not continual learning)

Files

├── README.md                     model card
├── framework_config.yaml         AlphaBrain framework configuration
├── dataset_statistics.json       action normalisation (12-dim action, 16-dim state)
├── model.safetensors             full VLA weights (~9.4 GB)
└── qwen_pretrained/              Qwen3-VL tokenizer + preprocessor configs

Usage

bash

git clone https://github.com/AlphaBrainGroup/AlphaBrain.git
cd VLA-Engine-Developer
pip install -e .

export PRETRAINED_MODELS_DIR=/path/to/models   # must contain Qwen3-VL-4B-Instruct/

huggingface-cli download AlphaBrainGroup/qwengr00t-robocasa365-atomic-nostate \
    --local-dir ./qwengr00t_robocasa_atomic

python deployment/model_server/server_policy.py \
    --ckpt_path ./qwengr00t_robocasa_atomic --port 10093 --use_bf16

For Robocasa365 evaluation, see benchmarks/Robocasa365/eval/ in the framework.

Notes on scope

This release is an SFT baseline on Robocasa365 atomic tasks, for users wanting a ready-to-evaluate Robocasa365 policy.
A companion with-state variant trained on the same data (later checkpoint, Apr 2026) exists locally and may be released later; this nostate release is the Apr 15 frozen version.
The continual-learning suite on Robocasa365 — which uses the same framework but a different schedule (sequential task stream) — is a separate set of releases coming soon.

License

MIT — see the parent repository.

Citation

bibtex

@misc{alphabrain2026,
  title  = {AlphaBrain: A Modular Open-Source Framework for Embodied Intelligence Research},
  author = {AlphaBrain Team},
  year   = {2026},
  url    = {https://github.com/AlphaBrainGroup/AlphaBrain}
}