RoboManipBaselines

Build, train, and deploy robot manipulation policies across real and simulated environments

💻 Code 📄 Paper 🗂️ Datasets
🎉 Version 3 Released
🕰️ Old Version Videos

Core Capabilities

RoboManipBaselines provides a unified software foundation for imitation learning in robotic manipulation, covering the full pipeline from data collection to deployment.

⚙️ End-to-end workflow
From demonstration data collection to policy training and rollout, all components are integrated into a single consistent pipeline.
🌍 Works across environments
Supports diverse environments including multiple simulators and real robots under a unified interface.
🧩 Easily extensible
New robots, tasks, and policies can be added with minimal implementation by extending base classes.
📊 Reproducible benchmarking
Enables systematic comparison using shared datasets and standardized evaluation procedures.

Policy Families

A unified interface supports a wide range of baseline and recent policy models.

MLP ACT Diffusion SARNN 3D Diffusion ManiFlow Flow Policy MT-ACT π0 GR00T

Environments

Supports diverse environments across both simulation and real-world settings.

Benchmarks

Enables systematic benchmarking with publicly available datasets.

Research Applications

Demonstrates how the framework can be extended for diverse research applications.

Resources

Access datasets, trained models, and evaluation results.

🌍 Environment catalog
Manipulation task environments across diverse robots
📂 Demonstration datasets
Datasets used for training and evaluation
🧠 Learned parameters
Pretrained policy checkpoints
📊 Evaluation results
Summary of evaluation experiments
RoboManipBaselines
An open-source framework for imitation learning in robotic manipulation.
GitHub | Paper