← Back to Home

🤖 Reinforcement Learning for VLA

My research in embodied AI focuses on developing efficient reinforcement learning frameworks for training and fine-tuning Vision-Language-Action (VLA) models, enabling robots to learn complex manipulation tasks.

First-Author Papers

### π_RL: Online RL Fine-tuning for Flow-based Vision-Language-Action Models **arXiv 2024** We introduce the first open-source framework for efficient RL fine-tuning with flow-based VLA models. The framework achieves 40%+ improvement over behavior cloning on manipulation tasks with successful sim-to-real transfer. **Key Contributions:** - First open-source RL fine-tuning framework for flow-based VLAs - Scalable training infrastructure for large VLA models - Novel policy gradient methods for flow-matching architectures \[[Paper](https://arxiv.org/abs/2510.25889)\] \[[Code](https://github.com/RLinf/RLinf)\]

Research Vision

The intersection of large language models and robotics is rapidly evolving. My research aims to:

1. Democratize Robot Learning: Make RL fine-tuning accessible without massive compute
2. Bridge Sim-to-Real: Develop methods that transfer effectively to real robots
3. Enable Generalization: Train robots that adapt to new tasks with minimal data