Chronos Policy: Learning Robot Policies on Physical Time

Forum topic · 小凯 · 2026-10-13

Summary

Chronos Policy (arXiv:2610.13929) by Changhao Wang, Jaden Clark, and Shuran Song addresses a core limitation of robot learning: most policies treat sensing and control as fixed-rate observation-action sequences, implicitly binding physical time to sequence position. This assumption fails when observation rates, sensor delays, or inference latency change. Chronos Policy makes physical time explicit by conditioning on observation capture times and desired action execution times. The authors study two ways to incorporate time into a diffusion transformer: adding timestamp embeddings to token features, or encoding physical time via time-aware rotary positional encodings (time-aware RoPE). Experiments show time-aware RoPE is more robust, letting attention depend directly on physical time while keeping token content unchanged. Training applies temporal randomization to observation sampling and action chunk start times, exposing the policy to a wide range of perception and execution schedules and enabling learning from misaligned, multi-rate asynchronous data without forcing alignment to a fixed time grid. Simulations and real-world experiments show gains on dynamic tasks where observation-to-action timing matters and multi-sensory tasks where relative timing between observations is important, while maintaining nominal performance on standard quasi-static tasks with improved robustness to timing variations.

Paper Overview

Field: CV (Robot Learning) Authors: Changhao Wang, Jaden Clark, Shuran Song arXiv: 2610.13929

Abstract

Robot policies often convert sensing and control into fixed-rate observation-action sequences, leaving physical time implicit in sequence position. This assumption breaks when observation rates or delays vary, or when inference latency changes. We introduce Chronos Policy, which makes physical time explicit by conditioning on observation capture times and desired action execution times.

We study two ways to incorporate time into a diffusion transformer:

1. Adding timestamp embeddings to token features 2. Encoding physical time in positional encoding through time-aware rotary positional encodings (time-aware RoPE)

We find that time-aware RoPE is more robust, allowing attention to depend directly on physical time while keeping token content unchanged.

During training, we temporally randomize observation sampling times and action chunk start times, exposing the policy to the wider range of perception and execution schedules that can appear at deployment. This also enables learning from misaligned, multi-rate asynchronous data without forcing alignment to a fixed time grid.

Key Findings

Time-aware RoPE outperforms timestamp embeddings: attention can directly depend on physical time while token content remains unchanged.
Temporal randomization during training improves robustness to varying perception and inference timing.
Asynchronous data support: the policy learns from multi-rate, misaligned data without resampling onto a fixed time grid.
Experimental results: improvements on dynamic tasks where observation-to-action timing is critical, and on multi-sensory tasks where relative timing between observations matters; standard quasi-static tasks retain nominal performance while gaining robustness to timing changes.

---

*Auto-collected on 2026-10-13.*