In the future, robots will perceive the environment as we do, communicate with us through natural language and help us with our tasks. LEGENT is dedicated to developing robots that can chat, see, and act from virtual worlds to the real world. Designed to integrate large models with embodied agents, this platform prioritizes ease of use and scalability, focusing on developing:
-
An easy-to-use environment that simulates a physical world, where an agent can interact with humans through language, receive egocentric vision, and perform physical actions.
-
Automated generation of training data, including the generation of scenes, tasks, and agent trajectories. The platform is tailored to train large multimodal models as embodied models, using generated data from simulated worlds at scale. LEGENT serves as the data engine for embodied models in robotics and games, as well as for world models.
Interact with the embodied agent in realistic scenes.
robotics.mp4
Interact with the embodied agent in stylized scenes.
game.mp4
-
Language Interaction. Use natural language as the human-robot interaction interface.
-
Fundamental Physics. The simulation incorporates gravity, friction, and collision dynamics.
-
Diverse Rendering. By adjusting assets and rendering features, LEGENT can achieve photorealistic rendering and stylized rendering. Instructions for trying out these scenes can be found here.
photorealistic.mp4
-
Interactable Objects. Agents and humans can manipulate various 3D objects.
interactable_objects.mp4
-
Scalable Assets. LEGENT supports importing (1) your own 3D objects, (2) objects from academic datasets, and (3) objects created by generative models. Learn more here. Note that the available adequately annotated 3D objects are very limited and vary a lot in format and quality. We are compiling a unified, open object assets library that can be freely used for embodied agent research.
assets_generated.mp4
assets_minecraft.mp4
-
Humanoid Animation. Body movement and nonverbal expression are also important for embodied agents. LEGENT will continue to enhance support in this aspect.
-
Scene Generation. LEGENT integrates advanced scene generation algorithms to support scalable training.
scene.generation.mp4
-
Trajectory Generation. Automatic generation of training data for training multimodal models into language-grounded embodied models. A minimal example of a trajectory:
{ "id": "20240509-223825-320898", "interactions": [ { "from": "human", "text": "Where is the orange?" }, { "from": "agent", "trajectory": [ { "image": "20240509-223825-320898/0000.png", "action": "rotate_right(18)" }, { "image": "20240509-223825-320898/0001.png", "action": "move_forward(2.0)" }, { "image": "20240509-223825-320898/0002.png", "action": "move_forward(1.8), rotate_right(30)" }, { "image": "20240509-223825-320898/0003.png", "action": "speak(\"It's on the sofa.\")" } ] } ] } -
User-friendly. LEGENT requires no complex installation and can run cross-platform on both PCs and servers. It is as intuitive as a game while also supporting complex research needs.
LEGENT is currently organizing code and documents and improving existing features. It will be more convenient to use once this process is complete. If you want a more stable version, please stay tuned!
- Polish APIs and write complete documentations.
- Release the first stable version.
- Develop a more powerful data generation system for training LMM-based embodied agents.
- Add planning-level action APIs to support text-only research.
- Add humanoid animation action APIs to support text-to-motion research.
- Add physics-based character/body control by integrating more dedicated tools such as MuJoCo.
- Add multi-agent support.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent