An open framework for building embodied agents that use LLM planners, VLA manipulation models and memory to run robot tasks in simulators and on real robots.
1.6K136PythonApache-2.0Updated yesterday
Gen-Verse/RSI-Anything
A research framework that uses world-model predictions and interaction feedback to drive recursive self-improvement across code repair, GUI tasks and lab planning.
- Stars
- 804
- Forks
- 63
- Open issues
- 4
- Last commit
- 7 days ago
What is it?
RSI-Anything is a domain-general framework built on JEPA-Anything. It connects candidate generation, prediction-guided selection, external evaluation and learning in one loop: a policy or workflow proposes actions, a world model predicts their outcomes, selected actions receive real feedback, and retained improvements guide the next round.
The project applies this cycle to five domains: code repair (CodeContests, APPS Introductory, LiveCodeBench), GUI interaction (MiniWoB++ and OSWorld), laboratory planning, Suzuki reaction optimization, and immune science workflows. Each domain adapter defines its own states, actions, prediction targets and feedback signals.
The repository provides a shared Python CLI (`rsi`) with commands to plan, run, resume and report on experiments, plus task-specific recipes under `configs/` and guides for environment setup and asset downloads.
Key features
- 01Shared CLI for planning, running, resuming and reporting experiments
- 02Predictive world model (JEPA-based) guides candidate selection
- 03Domain adapters for code repair, GUI, lab planning and reaction optimization
- 04Per-task recipes and setup guides, with assets hosted on Hugging Face
Use cases
- Researchers studying self-improving agents with world models
- Reproducing code-repair or browser-agent experiments from the paper
- Exploring prediction-guided action selection in scientific workflows
Quick start
- 1
Clone the repository
git clone https://github.com/Gen-Verse/RSI-Anything.git - 2
Enter the project directory
cd RSI-Anything - 3
Create and activate a Python 3.10+ virtual environment
python3.10 -m venv .venv && source .venv/bin/activate - 4
Install the shared command-line workflow and list tasks
python -m pip install -e . && rsi tasks
Things to know
- Full runs need the task-specific environments, assets and GPU setup described in the guides.
- Results depend on external benchmarks and large Hugging Face resources, so reproduction takes significant setup.
- The README covers research experiments rather than a packaged end-user tool.
Topics
FAQ
- What is RSI-Anything?
- A research framework that uses world-model predictions and interaction feedback to drive recursive self-improvement across code repair, GUI tasks and lab planning. RSI-Anything is a domain-general framework built on JEPA-Anything. It connects candidate generation, prediction-guided selection, external evaluation and learning in one loop: a policy or workflow proposes actions, a world model predicts their outcomes, selected actions receive real feedback, and retained improvements guide the next round.
- Is RSI-Anything free to use? What license does it have?
- RSI-Anything is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
- Is RSI-Anything still maintained?
- Actively maintained. The last commit to RSI-Anything was on Oct 13, 2026.
- What can I use RSI-Anything for?
- Researchers studying self-improving agents with world models Reproducing code-repair or browser-agent experiments from the paper Exploring prediction-guided action selection in scientific workflows
- How do I install RSI-Anything?
- Clone the repository: git clone https://github.com/Gen-Verse/RSI-Anything.git Enter the project directory: cd RSI-Anything Create and activate a Python 3.10+ virtual environment: python3.10 -m venv .venv && source .venv/bin/activate Install the shared command-line workflow and list tasks: python -m pip install -e . && rsi tasks
- Where can I get RSI-Anything?
- The source code and installation instructions are on GitHub: https://github.com/Gen-Verse/RSI-Anything
Summary generated by AI from the project's README. GitHub stats refresh daily.
Related projects
Unity ML-Agents Toolkit lets Unity games and simulations serve as environments for training AI agents with reinforcement and imitation learning.
AI & Machine Learning#reinforcement-learning#unity#deep-reinforcement-learning#imitation-learning#game-ai
19.8K4.5KC#OtherUpdated 7 days ago
A Python framework for recursive harness optimization that refines the prompts, tools and orchestration around a fixed LLM agent on unseen tasks.
9.2K802PythonApache-2.0Updated 6 days ago
A framework from the Terminal-Bench team for evaluating and optimizing AI agents and language models in container environments.
6.3K2KPythonApache-2.0Updated 17 hours ago
A browser-based control room for running LLM agent societies, letting you pause, question, steer, and replay simulated towns step by step.
AI & Machine Learning#ai-agents#multi-agent-simulation#llm-agents#social-simulation#human-in-the-loop
1.2K136PythonApache-2.0Updated 18 days ago
Turns any Agent Skill into RL-ready terminal tasks in the Harbor format, using Codex agents inside Docker containers.
18518PythonApache-2.0Updated 18 days ago