Embodied Reinforcement IntelligenCe (ERIC) is a framework that provides high-quality single-file implementations for finetuning vision-language-action (VLA) models via reinforcement learning. Following the design philosophy of CleanRL, ERIC is clean and simple, accelerating your research with user-friendly features. The highlight features of CleanRL are:
- 📜 Single-file implementation
First time in this area?
Don't worry, we provide a great notebook that helps you understand this area and build your project step by step!
See Fintuning VLA with RL from Scratch.
- Python: 3.10 (recommended)
- CUDA: 11.8+ or 12.1+
- GPU: NVIDIA GPU with 8GB+ VRAM (16GB+ for training)
# 1. Create conda environment
conda create -n eric python=3.10
conda activate eric
# 2. Install LIBERO from source
git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
pip install -e LIBERO/
# 3. Clone ERIC and install other dependencies
git clone https://github.com/RLE-Foundation/ERIC.git
cd ERIC
pip install -r requirements.txt
# 4. Install Flash Attention (performance critical)
pip install flash-attn==2.5.5 --no-build-isolationIf Flash Attention installation fails due to CUDA compilation issues, use this alternative method:
# Alternative: Download pre-compiled wheel
wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.5/flash_attn-2.5.5+cu122torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
# Install the downloaded wheel
pip install flash_attn-2.5.5+cu122torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whlNote: This wheel is for:
- CUDA 12.2 (compatible with CUDA 12.1+)
- PyTorch 2.2
- Python 3.10
- Linux x86_64
import torch
import numpy as np
from prismatic.vla.action_tokenizer import ActionTokenizer
from libero.libero import benchmark
print(f"PyTorch: {torch.__version__}")
print(f"CUDA available: {torch.cuda.is_available()}")
print("✓ ERIC components loaded successfully")