Overview

Embodied AI asks how agents can connect perception and reasoning to actions that change a physical or simulated world. Common approaches include vision-language-action models, language-model agents with tools and hierarchical plans, reinforcement learning, persistent spatial or episodic memory, and world models trained with robot demonstrations or generated simulation data. Researchers evaluate these systems on navigation, manipulation, household tasks, and multi-agent coordination, using benchmarks as well as real robots. A central challenge is that strong perception or short-task scores do not guarantee reliable long-horizon behavior: agents still struggle with physical constraints, changing conditions, recovery, and safety, while simulated success may not transfer to deployment.

Latest papers

CardsList

272 papers21 papers in the last four weeks, down 42% on the four weeks before. 0.2% of all new papers.

  1. Gibson Env: Real-World Perception for Embodied Agents

    Aug 31, 2018Fei Xia, Amir Zamir, Zhi-Yang He +3Embodied AIRobotics Simulation