About the Project

Building the Complete Ecosystem
for Real-World RL

EvoData is an open data platform developed by EvoMind Tech and MINT Lab, SJTU. We are dedicated to building a complete ecosystem — from real-world RL methodology to data collection and community sharing — to accelerate intelligent embodied AI.

The Problem

Why Does This Platform Exist?

Real-world reinforcement learning faces unique data challenges that existing platforms are not designed to address.

Real-World RL Data is Scarce

Unlike simulation data, high-quality real-world RL trajectory data is extremely rare. Collecting it requires physical robots, safety-critical RL training, and significant engineering effort.

VLA Models Need RL-Quality Data

Vision-Language-Action models rely on reinforcement learning for fine-grained, stable control. Without diverse real-world RL data, VLA policies fail to generalize beyond toy tasks.

No Unified Standard

Datasets are scattered across institutions with incompatible formats, making reproducibility and cross-dataset training nearly impossible.

Platform Features

What EvoData Offers

A comprehensive data infrastructure platform purpose-built for real-world reinforcement learning research.

  • Datasets collected via Evo-RL — our real-world RL training framework
  • LeRobot-compatible format for direct integration with modern RL pipelines
  • Unified metadata schema across all datasets
  • Built-in trajectory visualization and data exploration tools
  • Versioned datasets with full provenance tracking
  • Open Apache 2.0 licensing for research and commercial use
  • Community-driven dataset curation and quality control
  • Benchmarking suite for policy evaluation on real robots
Development Roadmap

Building a Complete Ecosystem

From a static showcase to a full real-world RL ecosystem — spanning methodology, data collection, and community sharing.

v0

Static Showcase

In Progress

Static website showcasing the platform vision, dataset previews, and the Evo-RL project roadmap.

  • Platform overview & vision
  • Dataset catalog with preview data
  • Design system & component library
v1

Dataset Browsing & Preview

Upcoming

Full dataset browsing with real metadata, trajectory previews, and search functionality.

  • Real dataset metadata API
  • Full-text search & tag filtering
  • Trajectory visualization
  • Dataset detail pages
v2

Upload & Download

Planned

Enable the community to upload and download real-world RL datasets with authentication and cloud storage.

  • User authentication system
  • Dataset upload pipeline
  • Chunked download with resume
  • Alibaba Cloud OSS integration
v3

Community & Ecosystem

Future

A full ecosystem integrating real-world RL methods, data collection, sharing, and community benchmarking.

  • Community dataset contributions
  • Policy evaluation benchmarks
  • Evo-RL method integration
  • Open API access
Our Team

Built by EvoMind Tech & MINT Lab, SJTU

EvoData is jointly developed by EvoMind Tech and MINT Lab at Shanghai Jiao Tong University — combining industry engineering with academic research to advance real-world robot learning.

EvoMind Tech
SJTU MINT Lab
Open Source