AlphaBrainGroup/qwengr00t-cl-lora-libero-goal

QwenGR00T-CL-LoRA (LIBERO-Goal)

Continual-learning checkpoint released with the AlphaBrain framework. Provided for direct download and evaluation — no retraining needed.

A QwenGR00T Vision-Language-Action (VLA) model fine-tuned sequentially over the 10 LIBERO-Goal tasks with Low-Rank Adaptation (LoRA, r=32) on the VLM backbone and Experience Replay (ER, buffer 1000/task) to mitigate catastrophic forgetting. Ships the final task checkpoint only — the weights you would use to evaluate the method end-of-stream.

Overview

Architecture	QwenGR00T (Qwen2.5-VL-3B + Flow-Matching DiT head)
Base VLM	Qwen/Qwen2.5-VL-3B-Instruct
Parameters	3.8 B total · ~82 M trainable (2.2 %)
LoRA	r = 32, α = 16, dropout 0.05, target_modules: all-linear
Continual-learning	Experience Replay, buffer 1000/task, replay ratio 0.5
Task stream	LIBERO-Goal · 10 tasks · 10 000 steps/task

Results

Evaluated with the final checkpoint on all 10 LIBERO-Goal tasks, 10 rollouts per task.

Metric	Value
Average Success Rate (Avg SR)	~48 %
Negative Backward Transfer (NBT, ↑ better)	+0.15
Naive sequential fine-tuning baseline (no ER)	< 10 %

Numbers are conservative estimates over our internal runs; per-run variance is a few percentage points depending on seed, simulator state, attention implementation, and hardware. Reproduction numbers higher or lower than reported are expected — please file an issue / PR with details.

Files

├── README.md                                       model card
├── config.yaml                                     training config (OmegaConf)
├── dataset_statistics.json                         action normalisation (required for inference)
├── task_9_id9_steps_100000_lora_adapter/           LoRA adapter weights + config
└── task_9_id9_steps_100000_action_model.pt         non-VLM weights (DiT head, DINO encoder)

Usage

Clone the AlphaBrain framework first (the inference server lives there).

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

# Set the directory holding the base VLM checkpoint
export PRETRAINED_MODELS_DIR=/path/to/models   # must contain Qwen2.5-VL-3B-Instruct/

# Download this release
huggingface-cli download AlphaBrainGroup/qwengr00t-cl-lora-libero-goal \
    --local-dir ./qwengr00t_cl_lora

# Merge LoRA + non-VLM weights into a deploy-ready .pt
python -m AlphaBrain.training.trainer_utils.peft.merge_lora_checkpoint \
    --base_config      configs/continual_learning/qwengr00t_cl_lora_libero.yaml \
    --lora_adapter_dir ./qwengr00t_cl_lora/task_9_id9_steps_100000_lora_adapter \
    --action_model_pt  ./qwengr00t_cl_lora/task_9_id9_steps_100000_action_model.pt \
    --output_path      ./qwengr00t_cl_lora_final.pt

# Launch the WebSocket inference server
python deployment/model_server/server_policy.py \
    --ckpt_path ./qwengr00t_cl_lora_final.pt --port 10093 --use_bf16

For a full 10×10 matrix evaluation against LIBERO-Goal, see the CL scripts README.

Reproduction

# One command — defaults match this release
bash scripts/run_continual_learning_scripts/run_cl_train.sh

The default wrapper loads configs/continual_learning/qwengr00t_cl_lora_libero.yaml, which is exactly the training config shipped here as config.yaml. Expect ~15 h on 2 × NVIDIA A800 80 GB.

License

MIT — see the parent repository.

Citation

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