EmbodiChain is an end-to-end, GPU-accelerated framework for Embodied AI. It streamlines research and development by unifying high-performance simulation, automated generative data pipelines, modular model architectures, and efficient training workflows. This integration enables rapid experimentation, seamless deployment of intelligent agents, and effective Sim2Real transfer for real-world robotic systems.
Note
EmbodiChain is in Alpha and under active development:
- More features will be continually added in the coming months. You can find more details in the roadmap.
- Since this is an early release, we welcome feedback (bug reports, feature requests, etc.) via GitHub Issues.
- 🚀 High-Fidelity GPU Simulation: Realistic physics for rigid & deformable objects, advanced ray-traced sensors, all GPU-accelerated for high-throughput batch simulation.
- 🤖 Unified Robot Learning Environment: Standardized interfaces for Imitation Learning, Reinforcement Learning, and more.
- 📊 Scalable Data Pipeline: Automated data collection, efficient processing, and large-scale generation for model training.
- ⚡ Efficient Training & Evaluation: Online data streaming, parallel environment rollouts, and modern training paradigms.
- 🧩 Modular & Extensible: Easily integrate new robots, environments, and learning algorithms.
The figure below illustrates the overall architecture of EmbodiChain:
Explore EmbodiChain's module relationships with the interactive architecture visualization.
To get started with EmbodiChain, follow these steps:
EmbodiChain's task environments are decoupled from the core framework at the
import-package level. The official classic-control, manipulation, and special
task families live in the
embodichain_tasks source tree
and are included in the main embodichain wheel. Both pip install embodichain
and editable pip install -e . therefore install
embodichain and embodichain_tasks; no second installation command or
independent task version is required. Third-party packages declaring an
embodichain.tasks entry point are also auto-discovered. Launch any registered
task with the unified CLI:
embodichain run-env --gym_config path/to/gym_config.jsonTo build your own task environments, start from the
embodichain_task_template
repository. Tasks with custom Python behavior implement an EmbodiedEnv
subclass with @register_env; supported configuration-defined Task Programs
may instead compose the common environment from reusable env.yaml,
embodiment, integration, and policy components. In either case, add a runnable
gym config with id, install with pip install -e ., and launch it with
embodichain run-env. See the
embodichain_tasks README
for details.
EmbodiChain environments can also be used with the optional
RLinf integration;
the integration guide documents the separate runtime, task registration, and
current compatibility boundary. The
official RLinf EmbodiChain recipe
provides a concrete end-to-end reference.
We welcome contributions! Please see the CONTRIBUTING.md file in this repository for guidelines on how to get started.
See Academic Publications for a complete list of academic papers related to EmbodiChain.
If you find EmbodiChain helpful for your research, please consider citing our work:
@misc{EmbodiChain,
author = {EmbodiChain Developers},
title = {EmbodiChain: An end-to-end, GPU-accelerated, and modular platform for building generalized Embodied Intelligence},
month = {November},
year = {2025},
url = {https://github.com/DexForce/EmbodiChain}
}@misc{GS-World,
author = {Guiliang Liu and Yueci Deng and Zhen Liu and Kui Jia},
title = {GS-World: An Efficient, Engine-driven Learning Paradigm for Pursuing Embodied Intelligence using World
Models of Generative Simulation},
month = {October},
year = {2025},
journal = {TechRxiv}
}