🤖 VLA_RL_Reasoning: AIRI 2025

🎯 Developed as part of DL Summer School 2025 research initiative.

🏆 Key Results

✅ Closed issue in ManiSkill repository
📊 Implementation: Located in /src folder, accessible via *.ipynb notebooks
🤗 Models & Datasets: Available on HuggingFace

🚀 Quick Start

# Create conda environment
conda create --name vlarlr python=3.10
conda activate vlarlr

# Clone the repository
git clone https://github.com/dancher00/VLA_RL_Reasoning
cd VLA_RL_Reasoning

...and then enjoy our notebooks!

🚀 Train

Install lerobot

https://github.com/huggingface/lerobot

Run the train script

python3 lerobot/scripts/train.py \
  --output_dir=./outputs/train \
  --policy.path=lerobot/pi0 \
  --dataset.root=/path \
  --dataset.repo_id=dancher00/maniskill-panda-pickcube \
  --policy.repo_id=name \
  --wandb.enable=true \
  --wandb.project=project \
  --optimizer.type=adamw \
  --optimizer.lr=2.5e-05 \
  --optimizer.weight_decay=1e-10 \
  --save_freq=100

📁 Project Structure

📂 File/Folder	📝 Content
data/	📊 Training and testing datasets
src/	💻 Core source code and implementations
tools/	🔧 Utilities for dataset conversion and preprocessing

🔗 Resources

• 📊 Presentation - Project overview and findings
• 📄 Technical Report - Detailed methodology and results
• 🤗 Models & Datasets - Pre-trained models and datasets

👥 Contributors

• 👨‍💻 Vakhitov Rodion
• 👨‍💻 Belov Danil
• 👨‍💻 Ivanov Leonid
• 👨‍💻 Kachaev Nikita

---

🎉 AIRI 2025 Research Project 🎉

Advancing the frontiers of AGI