Chronos Policy gives robots the actual time of observations and actions

Sensors and inference don't always follow a fixed rhythm. Chronos encodes physical time to better withstand delays and asynchronous data flows.

Chronos Policy conditions robotic decisions on the time of capture of observations and the planned time to execute actions. The authors compare the addition of timestamps to token characteristics with a temporal encoding of positions in the attention mechanism.

Physical time-based positional encoding is more robust in their trials. Training varies the timing of observation and start of actions to expose policy to different time frames. Simulated and real-world evaluations indicate gains on dynamic and multisensory tasks, while preserving nominal performance on quasi-static tasks.

Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.