CORAL: Scalable Multi-Task Robot Learning via LoRA Experts

A backbone- and embodiment-agnostic framework that resolves multi-task interference in VLA deployment through strict parameter isolation via LoRA experts.

Installation

conda create -n simvla python=3.10 -y
conda activate simvla

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install transformers>=4.57.0
pip install peft accelerate fastapi tensorboard uvicorn json_numpy safetensors scipy einops timm mmengine pyarrow h5py mediapy num2words av wandb websockets msgpack_numpy
pip install flash-attn==2.5.6 --no-build-isolation
pip install tensorflow tensorflow-datasets

Important: Use transformers>=4.57.0.

Training (LIBERO Dataset)

1. Prepare LIBERO Dataset

Download LIBERO dataset, and place it in ./datasets/metas/.

2. Train CORAL LoRA Experts

Train per-task LoRA experts (one per LIBERO evaluation task). The SimVLA base model is automatically downloaded from HuggingFace.

CUDA_VISIBLE_DEVICES=0,1,2,3 SUITES="libero_10" bash train_coral_libero_all_r16.sh

3. Evaluation

cd evaluation/libero

See evaluation/libero/README.md for detailed instructions.

4. Results

CORAL achieves a 99.3% overall average success rate on the LIBERO benchmark.

Reference

If you find our codes useful, please consider citing our work

@article{luo2026coral,
  title={CORAL: Scalable Multi-Task Robot Learning via LoRA Experts},
  author={Luo, Yuankai and Chen, Woping and Liang, Tong and Li, Zhenguo},
  journal={arXiv preprint arXiv:2603.09298},
  year={2026}
}