Quick links
Definition
Embodied vs disembodied AI; the body–perception–decision–action loop.
Key technologies
VLA, world models, IL/RL, Diffusion Policy, Sim2Real.
Robot morphologies
Fixed-base, wheeled, quadruped, humanoid, biomimetic.
Key papers
RT series, OpenVLA, ACT, Diffusion Policy, and more.
Data & simulation
Open X-Embodiment, MuJoCo, Isaac, Habitat, RoboTwin.
Learning path
Embodied-AI-Guide, Lumina community, courses & newsletters.
In one sentence
Large models already handle text and images well because the web is full of that data. Getting a robot to pick up a cup requires continuous actions with physical consequences. Where data comes from, how actions are represented, and how success is measured—these are Embodied AI’s core tensions. (Adapted from Embodied-AI-Guide.)
Four research pillars
The survey Aligning Cyber Space with Physical World (arXiv:2407.06886) organizes the field into:
- Embodied perception — active vision, VLN, 3D scene understanding
- Embodied interaction — EQA, language-guided grasping, and more
- Embodied agents — multimodal foundation models; task & action planning
- Sim-to-real adaptation — world models, data collection, domain randomization
What this site covers
- Definitions, history, and core elements (body / perception / decision / action)
- Key technology stack and representative systems
- Humanoid, mobile, and manipulation robots; embodied devices
- Applications, companies, papers, datasets, and learning paths