Chronos Policy: Learning Robot Policies on Physical Time

Changhao Wang; Jaden Clark; Shuran Song

cs.ROOct 9, 2026

Chronos Policy: Learning Robot Policies on Physical Time

Authors: Changhao Wang, Jaden Clark, Shuran Song

Organizations: Stanford University

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: adding timestamp embeddings to token features, or 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 which observations are sampled and when the action chunk begins, exposing the policy to a wider range of sensing and execution schedules that may occur during deployments. This also enables learning from asynchronous, multi-rate data without forcing all measurements onto a fixed temporal grid. Across simulation and real-world experiments, Chronos Policy improves performance on dynamic tasks, where observation-to-action timing is critical, and on multisensory tasks, where the relative timing between observations matters. On standard quasi-static tasks, it preserves nominal performance while improving robustness to changes in sensing and inference timing.

Figures & tables

Figure 1: Timing-related issues in robot policy learning. (a) Real robot systems exhibit multiple sources of temporal variation, including variable sensor sampling rate, observation latency and dropout, inference latency, control frequency, and execution latency. These timing effects matter differently across tasks. (b) In dynamic tasks, the key temporal relationship is between when an observation is captured and when an action takes effect . (c) In multisensory tasks, the key temporal relationship is between when different observations are captured . (d) Even in quasi-static tasks, differences between data-collection and deployment timing can create temporal mismatch. Standard fixed-rate policy representations obscure these distinctions by treating sequence position as a proxy for time.

Figure 2: Overview of Chronos Policy. Tokens are paired with their physical times. Time-aware RoPE uses these physical timestamps as positional coordinates, so attention depends on elapsed time rather than sequence index. The diffusion transformer then predicts an action chunk at the requested execution times.

Figure 3: Simulation ablation of Chronos Policy design. We compare temporal randomization, additive time encoding, and time-aware RoPE to identify an effective recipe for time awareness. Temporal randomization is most effective when paired with explicit timestamps, while time-aware RoPE provides the strongest robustness across Conveyor and Pivoting.

Figure 4: Simulation evaluation across manipulation settings. We compare Diffusion Policy and Chronos Policy on (a) dynamic Conveyor, (b) multisensory Pivoting, and (c) six standard DexMimicGen tasks under nominal and perturbed sensing and inference timing.

Figure 5: Real-world evaluation under timing variation. We compare Diffusion Policy, Diffusion Policy with RTC, and Chronos Policy on Dynamic Cup, Peg Insertion, and Pivoting under nominal, observation, inference, and combined timing conditions. Lighter bar segments indicate partial success.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Architecture of Chronos Policy. Proprioceptive, optional force/torque, and noisy action tokens interact through self-attention and attend to image features through cross-attention. Time-aware RoPE encodes observation capture times and queried action times in attention. The transformer iteratively denoises the action chunk.

Setting	Standard tasks	Conveyor	Pivoting
Training epochs	50	50	50
Crop resolution	$224\times 224$	$196\times 196$	$196\times 196$
DDIM steps: training / inference	24 / 8	24 / 8	24 / 8
Action excution horizon	12	8	8

Appendix

Table 1: Simulation training and execution settings. The action prediction horizon is 16 and action frequency is 20 Hz for every task.

Condition	Image	Pose	Force/torque	Delay
Nominal	20	20	100	0
Small	12.5–33.33	12.5–33.33	75–100	0–50
Large	1–50	1–50	1–100	0–100

Appendix

Table 2: Simulation observation-timing conditions. Sensor frequencies are in Hz and delivery delays are in ms. Small and Large correspond to the observation perturbations used in the main figures; sensor streams are sampled independently.

Conveyor
Method	Nominal	Small obs.	Wide sensors	Delay 0–100 ms	Delay 0–200 ms	Combined
DP	$73.1\pm 7.2$	$15.6\pm 1.0$	$0.0\pm 0.0$	$14.2\pm 2.8$	$1.8\pm 1.5$	$0.0\pm 0.0$
DP + rand.	$53.6\pm 2.3$	$60.9\pm 3.9$	$6.4\pm 2.1$	$47.3\pm 2.4$	$19.6\pm 2.7$	$0.2\pm 0.4$
RoPE w/o rand.	$79.8\pm 4.2$	$67.1\pm 3.4$	$20.2\pm 0.8$	$76.4\pm 8.6$	$69.3\pm 3.1$	$10.7\pm 0.7$
AddTime	$\mathbf{85.8\pm 7.6}$	$74.9\pm 4.5$	$24.0\pm 1.3$	$78.9\pm 4.4$	$71.8\pm 3.4$	$7.8\pm 2.0$
RoPE (ours)	$82.0\pm 3.5$	$\mathbf{77.6\pm 3.0}$	$\mathbf{45.6\pm 10.2}$	$\mathbf{83.1\pm 2.8}$	$\mathbf{76.4\pm 3.4}$	$\mathbf{32.7\pm 5.7}$

Appendix

Table 3: Full Conveyor and Pivoting results and design ablations. Success rate (%), mean

\pm

sample SD across three training seeds after averaging three paired evaluation batches within each seed. Each condition contains 450 episodes per method. RoPE is the full Chronos Policy; AddTime changes only the time encoding. Rand. denotes temporal randomization. Best mean performance within each task and condition is shown in bold .

Threading
Method	Nominal	Small obs.	Wide sensors	Delay 0–100 ms	Delay 0–200 ms	Combined
DP	$\mathbf{89.3\pm 8.3}$	$88.7\pm 5.0$	$\mathbf{78.7\pm 11.0}$	$\mathbf{90.0\pm 7.2}$	$79.3\pm 5.0$	$\mathbf{77.3\pm 8.1}$
Chronos Policy	$88.7\pm 3.1$	$\mathbf{90.0\pm 2.0}$	$77.3\pm 3.1$	$84.7\pm 5.0$	$\mathbf{86.7\pm 3.1}$	$63.3\pm 3.1$
Assembly
DP	$\mathbf{75.3\pm 7.0}$	$\mathbf{73.3\pm 5.0}$	$13.3\pm 6.4$	$\mathbf{72.7\pm 7.6}$	$\mathbf{72.7\pm 1.2}$	$0.0\pm 0.0$
Chronos Policy	$68.7\pm 1.2$	$\mathbf{73.3\pm 2.3}$	$\mathbf{61.3\pm 3.1}$	$71.3\pm 6.1$	$71.3\pm 6.4$	$\mathbf{57.3\pm 7.6}$

Appendix

Table 4: Full results on the six standard manipulation tasks. Success rate (%), mean

\pm

sample SD over three training seeds. Each seed is evaluated for 50 episodes per condition. DP denotes the transformer-based Diffusion Policy baseline. Best mean success within each task and condition is shown in bold .

Task	Demos	Full success	Partial success
Dynamic Cup	252 / 13	Cup upright on the moving saucer	Cup placed on its side on the saucer
Peg Insertion	220 / 11	Peg inserted within 30 s	Peg inserted after 30 s
Pivoting	322 / 17	Object brought upright and stable	Object reaches upright, then falls

Appendix

Table 5: Real-world datasets and scoring. Full success receives one credit, partial success half a credit, and failure zero. Demonstration counts are training / validation.

Stream	Nominal rate (Hz)	Perturbed rate (Hz)	Added delay (ms)
RGB	20	12.5–33.33	0–100
Pose / gripper	20	8.33–33.33	0–100
Force/torque	90	37.5–150	0–100

Appendix

Table 6: Real-world observation-timing settings. Each stream has an independent delivery delay. Force/torque applies to Peg Insertion and Pivoting.

Policy	Scenario	Nominal	Observation	Inference	Combined
DP	Clockwise	40.0 (2/0/3)	50.0 (2/1/2)	20.0 (1/0/4)	40.0 (2/0/3)
	Counterclockwise	60.0 (3/0/2)	20.0 (1/0/4)	60.0 (3/0/2)	20.0 (1/0/4)
	Overall	50.0 (5/0/5)	35.0 (3/1/6)	40.0 (4/0/6)	30.0 (3/0/7)
DP + RTC	Clockwise	60.0 (3/0/2)	30.0 (1/1/3)	60.0 (2/2/1)	20.0 (1/0/4)
	Counterclockwise	60.0 (3/0/2)	50.0 (2/1/2)	40.0 (2/0/3)	30.0 (1/1/3)
	Overall	60.0 (6/0/4)	40.0 (3/2/5)	50.0 (4/2/4)	25.0 (2/1/7)

Appendix

Table 7: Dynamic Cup results by motion direction. Each direction contains five trials; Overall aggregates all 10 trials per policy and timing condition. Each cell reports task score (%) followed by full/partial/failure counts.

Policy	Scenario	Nominal	Observation	Inference	Combined
DP	Forward	90.0 (4/1/0)	50.0 (2/1/2)	50.0 (2/1/2)	50.0 (2/1/2)
	Backward	100.0 (5/0/0)	80.0 (3/2/0)	40.0 (2/0/3)	40.0 (1/2/2)
	Overall	95.0 (9/1/0)	65.0 (5/3/2)	45.0 (4/1/5)	45.0 (3/3/4)
DP + RTC	Forward	80.0 (4/0/1)	20.0 (1/0/4)	50.0 (2/1/2)	70.0 (3/1/1)
	Backward	90.0 (4/1/0)	60.0 (3/0/2)	60.0 (3/0/2)	0.0 (0/0/5)
	Overall	85.0 (8/1/1)	40.0 (4/0/6)	55.0 (5/1/4)	35.0 (3/1/6)

Appendix

Table 8: Peg Insertion results by hole orientation. Forward and backward orientations each contain five trials; Overall aggregates all 10 trials. Each cell reports task score (%) followed by full/partial/failure counts.

Policy	Scenario	Nominal	Observation	Inference	Combined
DP	Large eraser (ID)	100.0 (5/0/0)	100.0 (5/0/0)	80.0 (4/0/1)	100.0 (5/0/0)
	Small eraser (ID)	60.0 (3/0/2)	60.0 (3/0/2)	60.0 (3/0/2)	60.0 (3/0/2)
	Box (OOD)	40.0 (2/0/3)	20.0 (0/2/3)	20.0 (1/0/4)	40.0 (2/0/3)
	Sideways (OOD)	10.0 (0/1/4)	0.0 (0/0/5)	0.0 (0/0/5)	10.0 (0/1/4)
	Overall	52.5 (10/1/9)	45.0 (8/2/10)	40.0 (8/0/12)	52.5 (10/1/9)
DP + RTC	Large eraser (ID)	100.0 (5/0/0)	100.0 (5/0/0)	100.0 (5/0/0)	80.0 (4/0/1)

Appendix

Table 9: Pivoting results by object and generalization setting. Large and small erasers are in-distribution (ID); the box and sideways setting are out-of-distribution (OOD). Each individual scenario contains five trials, and Overall aggregates all 20 trials. Each cell reports task score (%) followed by full/partial/failure counts.

Explore similar work

Shared Execution-Clock Drifting Policy for Dynamic Precision Manipulation
Sep 20, 2026Zhenchen Dong, Qingran Wu, Jinna Fu +4Dynamic Robotic Manipulation Robot Imitation Learning
The Hong Kong Polytechnic University. · University of California San Diego. · Zhejiang University. +2
Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity
Jul 6, 2026Che-Sang Park, Junsu Ha, Jianlong Fu +1Robot Policy Learning Diffusion-Based Planning
Robotics Laboratory, Seoul National University, Seoul, South Korea · Microsoft Research Asia, Beijing, China
Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation
Jun 29, 2026Yulin Zhou, Yimeng Wang, Nengyu Wang +9Robot Policy Learning Long-Horizon Robotic Manipulation
School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China

Sep 20, 2026, cs.ROJ/K move · Enter open · S save

Shared Execution-Clock Drifting Policy for Dynamic Precision Manipulation

Zhenchen Dong, Qingran Wu, Jinna Fu, Jiaming Wu+3

The Hong Kong Polytechnic University. · University of California San Diego. · Zhejiang University.+2

Dynamic Robotic Manipulation Robot Imitation Learning Trajectory Representation Learning

Open paper AskPDF arXiv

Manipulation under time constraints requires both accurate actions and an execution rhythm that matches the evolving scene. This becomes critical when a robot must intercept moving objects or complete a sequence of adjustments before a deadline. Although one-step policies reduce generation cost, their directly predicted action sequences leave temporal allocation implicit. We propose Shared Execution-Clock Drifting (SECD), which makes execution rhythm an explicit part of one-step action generation. Conditioned on an observation and a latent sample, the policy jointly predicts a progress-indexed action curve and a shared monotone clock that maps fixed control times to locations on the curve. Demonstration-derived alignment anchors this decomposition, which is trained jointly through drifting on the decoded actions. The resulting policy retains a fixed-rate control interface and requires one network evaluation. We evaluate SECD across four real-robot tasks with inference on NVIDIA Thor. Across 300 trials, it achieves 77.00% task-averaged success and outperforms the evaluated one-step baselines on every task, including 91% success in cup retrieval from a 16 m/min conveyor and 54% in restoring and folding a crumpled shirt within 90 s. A fixed-clock variant reaches 79% on the same conveyor protocol. Complementary state-based RoboMimic experiments, including cross-seed ablations on Transport and Square, further support the joint design of the temporal representation and demonstration alignment. Project page: https://secd-anonymous-ewn.pages.dev/

Abstract

Explore similar work

Shared Execution-Clock Drifting Policy for Dynamic Precision Manipulation

Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity

Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation