WorldToken: Time-First Sequence Modeling for Robotic Imitation Learning
The episode discusses WorldToken, a time-first sequence modeling approach for robotic imitation learning by Chunkai Yang and colleagues. Hosts analyze its architecture, focusing on how it organizes physical time as a primary sequence unit and its empirical findings regarding context utilization and scaling. Improvements suggested include better multimodal encoding, causal transformers, diffusion action heads for controlled exploration, and efficient sliding window context management.