SOLE-R1 is a VLM-based dense reward model for robotic manipulation. Given a natural-language task description and a robot video, it predicts per-timestep task-progress percentages that can be used as reward signals for reinforcement learning.
Tested with Python 3.10 on Ubuntu 22.04.
reward_server/ VLM inference server (SOLE-R1, GPT, or Gemini)
reward_client/ CLI client that queries the server and annotates videos
videos/ Sample robot videos
See reward_server/README.md for full setup, usage, and API documentation.
See reward_client/README.md for full setup and usage documentation.
This repository is released as-is to accompany a paper submission.
If you find any bugs, corrections, or issues that should be resolved for anyone looking to reproduce the results in this repository, please file an issue and we will look at it as soon as we can.
For other improvements, including new features, we recommend creating your own fork of the repository.