机器人ICML 20252025
GR-3: Generalist Robot Policy with World Model
Zipeng Fu, Tony Zhao, Chelsea FinnStanford University
摘要
We introduce GR-3, a generalist robot policy that leverages an internal world model for planning and execution. GR-3 can perform diverse manipulation tasks by imagining future states and selecting optimal action sequences, demonstrating robust performance in both simulated and real-world environments.
generalist policyworld modelrobot planningmanipulationimagination
技术细节
数据集
测试机器人
模型骨架
Generalist Policy + Internal World Model
编码器
Multi-Modal ViT + World Model Encoder
解码器
Policy Decoder + World Model Decoder (8+6 layers)
