🤖 Latest Embodied AI Research

Introduces Voyager, the first LLM-powered embodied lifelong learning agent in Minecraft. The agent writes, refines, commits, and retrieves code to continually explore the world, acquiring diverse skills without human intervention.

RT-2 demonstrates how vision-language models trained on web-scale data can be adapted for robotic control. The model transfers semantic and visual knowledge from internet-scale datasets to enable better generalization in robotic manipulation tasks.

PaLM-E is a multimodal embodied language model that integrates vision and language for robotic control. With 562B parameters, it demonstrates how large language models can be grounded in real-world sensor data for embodied reasoning and planning.

Explores the concept of distributed human presence through AI-enabled cognitive replication and robotic embodiments. This work challenges traditional constraints of physical embodiment, proposing frameworks for multi-location human existence through advanced AI systems.

Addresses the challenge of embodied question answering where agents must navigate and explore 3D environments to answer questions. Uses point cloud representations for more efficient perception and demonstrates strong performance in photorealistic simulated environments.