WorldToken: Time-First Sequence Modeling for Robotic Imitation Learning
Chunkai Yang · hf · 2026-08-25
WorldToken introduces a time-first sequence modeling approach for robotic imitation learning. It fuses heterogeneous robot observations into per-timestep world tokens processed by a causal Transformer and diffusion action head. The paper includes scaling and temporal-context analyses on RoboCasa and RMBench benchmarks.
More from Embodied
- BrainCo's Dexterous Data Matrix combines human, robot, and virtual data for embodied AI — eyishazyer · 2026-08-25
- Wutong Sensing: Building the tactile data layer for physical AI — 创业邦 · 2026-08-25
- Europe is a lot less behind in robotics than people think — StephanSturges · 2026-08-25
- Robotic demos stuttering? FlashRT engine offers Hz-level latency benchmarks — Xianbao_QIAN · 2026-08-25
- X-Humanoid's Tiangong robot hits a 3.4m standing long jump on raw torque — CyberRobooo · 2026-08-25
- China's Galileo robotic dog transforms like a crab — basedjensen · 2026-08-25