A preprint published on arXiv on 9 October by Stanford University researchers proposes giving robot policies explicit information about when sensor observations were captured and when actions are meant to run. Chronos Policy adds those timestamps to a diffusion-based robot policy. The authors compare encoding timestamps in the model’s positional representation with adding them to token features, and report that time-aware rotary positional encoding was more robust to timing changes.

In a simulated moving-conveyor task, the paper reports nominal success rates of 73.1% for Diffusion Policy and 82% for Chronos Policy. When observation timing varied, the rates were 15.6% and 77.6%; with inference delays up to 200 milliseconds, they were 1.8% and 76.4%. In real-robot experiments, Chronos also achieved the highest score across the timing conditions tested in three tasks. These are results from the authors’ specific experimental setups, not evidence of commercial performance across other robots or environments. The work is a preprint and has not yet undergone peer review.