MINT-SJTU/Evo-1

What it solves

Evo-1 is a lightweight Vision-Language-Action (VLA) model designed for robotic manipulation. It addresses the challenge of creating efficient models that can translate visual and textual instructions into precise robot actions while maintaining semantic alignment, enabling it to work across various simulation benchmarks and real-world robot embodiments.

How it works

Evo-1 uses a two-stage training paradigm to learn robotic control. In the first stage, it trains an integration module and an action expert. In the second stage, it performs full-scale training, including the VLM (Vision-Language Model) and the action head. It supports a flow-matching action head and integrates with the LeRobot framework for data handling and deployment. The model operates via a server-client architecture where a client sends observations (images, robot state, and prompts) and the server returns action chunks.

Who it’s for

This project is for robotics researchers and developers who need a lightweight VLA model for embodied intelligence, specifically those working with simulation environments like Meta-World, LIBERO, and RoboTwin, or deploying to real hardware such as xArm6, SO100/SO101, and Jetson Orin.

Highlights

• Broad Benchmark Support: Validated on Meta-World, LIBERO, LIBERO-plus, and RoboTwin (50 bimanual tasks).
• Lightweight Architecture: Designed for efficiency, with a dedicated evo1-flash branch for faster training and reduced GPU memory usage.
• Framework Integration: Fully integrated into the Hugging Face LeRobot framework and supported by the RLinf framework for SFT and GRPO fine-tuning.
• Cros-Embodiment Deployment: Provides inference scripts for various hardware including xArm6 and Jetson Orin Docker support.