acceptodds
Under review as a conference paper at ICLR 2027

Chronos Policy: Learning Robot Policies on Physical Time

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 through 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.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, Imitation Learning, Diffusion Policy, Robotic Manipulation, Contact-Rich Manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NSQBOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes a method to train policies that are sensitive to training-inference mismatches in physical time w.r.t actions, observations/other sensors.

Critical strengths

The paper is easy to read and follow for the most part. The experiments clearly demonstrate the effectiveness of the method against a baseline that does not use it.

Critical weaknesses

The set of baselines used for comparison are insufficient, as the main comparison is with Diffusion Policy. The main novelty is temporal awareness with RoPE embeddings, which will be better substantiated by thorough empirical comparison with more baselines, even if the tasks/environments for such comparison are much fewer in number.

Questions for authors
  1. It's not clear why some experiments include comparison with RTC and some do not. There are also other approaches eg. DiscreteRTC (which you've cited), "Delay-Aware Diffusion Policy: Bridging the Observation–Execution Gap in Dynamic Tasks", "FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking" etc. that claim to work in similar time-mismatch scenarios. Please add additional comparisons.

  2. Fig. 3 was hard to follow. Please add details about what each baseline does where relevant.

  3. Does the Diffusion policy share the same architecture as the Chronos policy (other than the time-aware embeddings)

3 comments
AuthorsOpenReview ↗

Response to Reviewer NSQB

Thank you for your kind words and insightful suggestions on our paper.

Q1. “It's not clear why some experiments include comparison with RTC and some do not. There are also other approaches eg. DiscreteRTC (which you've cited), "Delay-Aware Diffusion Policy: Bridging the Observation–Execution Gap in Dynamic Tasks", "FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking" etc. that claim to work in similar time-mismatch scenarios. Please add additional comparisons.”

We have expanded the simulation comparisons to include DP with latency compensation (DP + LC), RTC, DiscreteRTC, Delay-Aware Diffusion Policy (DA-DP), FutureRTC, and a timestamp resampled baseline.

Q2. “Fig. 3 was hard to follow. Please add details about what each baseline does where relevant.”

We thank the reviewer for pointing this out. Figure 3 compares five variants to study the effects of time representation and temporal randomization:

  • DP: Standard Diffusion Policy without explicit timing or temporal randomization.

  • DP + Randomization: DP trained with temporal randomization but without explicit timing.

  • RoPE without Randomization: Uses physical-time RoPE but is trained without temporal randomization.

  • AddTime: Uses additive timestamp embeddings together with temporal randomization.

  • Chronos: Uses physical-time RoPE together with temporal randomization.

Temporal randomization varies the observation sampling times and action execution times during training. The results are averaged across three training seeds, with 150 evaluation episodes per seed for each condition. The average performance under timing perturbations is computed across the five non-nominal conditions, and error bars show the standard deviation across seeds.

We will add these details to the figure caption to make the comparisons clearer.

Q3. “Does the Diffusion policy share the same architecture as the Chronos policy (other than the time-aware embeddings)”

Yes, DP and Chronos Policy share the same transformer backbone.

I thank the authors for their detailed response. The new comparisons show Chronos as a simple yet effective method. I think this is a meaningful contribution and therefore, increase my score to 4.

Thank you so much for your thoughtful feedback and encouraging comments! We sincerely appreciate your recognition of our work and your willingness to reconsider the score. We are very grateful for your time and consideration!

Reviewer_iDF5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the sensitivity of robot diffusion policies to observation timing and inference delays and proposes Chronos Policy. Specifically, it encodes sensor capture times and requested action execution times using time-aware rotary positional embeddings and trains with temporal randomization. The formulation represents relative physical time in attention. Experiments on dynamic and quasi-static simulation tasks and three real-robot tasks demonstrate improved robustness to changes in sensing and inference timing.

Critical strengths
  • AddTime and RoPE share timestamps, data, token layout, and randomization, while separate controls remove explicit timing or augmentation. These comparisons address plausible alternatives involving more data diversity or a different backbone. The custom simulation ablations include three training seeds and 450 episodes per condition.
  • The study spans moving-target and contact-rich tasks, with RTC included in hardware comparisons. RTC and Chronos both estimate latency from past requests, so an unfair future-latency oracle advantage in hardware is not supported. Reported full/partial/failure counts also make the hardware scoring inspectable.
Critical weaknesses
  • The novelty claim needs tighter attribution and a decisive prior-method comparison. Previous works [1, 2] have already introduced continuous-time RoPE and physical-time action-query encoding, respectively. BulletTime [1] is correctly cited, while LaWAM [2] is omitted. LaWAM uses additive embeddings and lacks Chronos's joint irregular sensor/action RoPE, so the full interface is not established as duplicated. The manuscript identifies DA-DP [3] as its closest delay-aware predecessor but does not evaluate it. AddTime already provides a useful timestamp-conditioned control, isolating additive versus rotary encoding within the proposed interface. A DA-DP comparison would clarify the increment over its different delay-conditioned training formulation, especially in inference-delay conditions. Missing original Diffusion Policy [4] and DexMimicGen [5] bibliography entries weaken attribution to the original policy formulation and named benchmark.
  • The decisive benchmarks and deployment protocol are incompletely reproducible. Section 4.1 and Appendices A–B omit a complete Conveyor/Pivoting data/task specification: demonstration counts/sources, numerical dynamics distributions, success/timeout rules, and split construction. Architecture and timing ranges do not reconstruct this evidence. Hardware testing has 10–20 trials per condition and unclear training-seed replication and pairing/order. The latency estimator and overlap/gap scheduler are also underspecified, preventing assessment of sensitivity to prediction error. Stochastic delays are tested and simulation has three seeds; neither resolves these gaps. Nominal Insertion's 100 versus 95 score differs by one partial outcome, so small hardware advantages should remain descriptive.

References

[1] BulletTime: Decoupled Control of Time and Camera Pose for Video Generation

[2] LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

[3] Delay-Aware Diffusion Policy: Bridging the Observation-Execution Gap in Dynamic Tasks

[4] Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

[5] DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning

Questions for authors
  • Clarification: Please provide the custom task definitions, demonstration-generation details, and independent split rules.
  • Requested comparison: Does the unified interface improve over a matched DA-DP-style delay-conditioned policy, particularly on inference-delay conditions, and what precisely extends LaWAM's physical-time action queries?
  • Clarification: Which latency estimator and overlap/gap scheduler were used, and what errors occurred during hardware evaluation?
  • Requested analysis: Report hardware uncertainty and pairing/training-seed details; relate performance to measured latency-prediction errors.
  • Deployment context: How does the approach relate to the uncited A2C2 [1], which corrects delayed action chunks using current observations and chunk-position features? Its feedback head differs from joint physical-time conditioning, but provides relevant context beyond RTC alone.

References

[1] Leave No Observation Behind: Real-time Correction for VLA Action Chunks

4 comments
AuthorsOpenReview ↗

Response to Reviewer iDF5 [1/2]

Thank you for the detailed suggestions, which have helped us significantly improve the paper. We address each question below.

Q1. “Clarification: Please provide the custom task definitions, demonstration-generation details, and independent split rules.”

SpecificationConveyorPivoting
Demonstrations1,742 trajectories at 100 Hz242 trajectories at 100 Hz
Train/validation split1,707 / 35216 / 26
Evaluation variationsBelt speed: 0.01–0.30 m/s; object position: ±2 cm (x), ±1 cm (y); cube yaw: ±0.25 radMass: 1.10–1.65 kg; length: 12–18 cm; width: 8–12 cm; thickness: 2.5 cm; sliding friction: 0–0.12
Success criterionLift the object more than 12 cm above the belt within 8 cm of the TCP, then place it in the target binRotate the object more than 75° from horizontal
Episode timeout30 s30 s

Data collection. Both datasets are collected through human teleoperation using the Force Dimension Omega.7 haptic device. The device enables precise control of the robot's 6-DoF end-effector pose and gripper, while providing force feedback from the simulation. This allows the operator to feel contact forces during demonstrations, which is particularly important for contact-rich tasks such as Pivoting. We plan to release both datasets and benchmarks to support further research on timing-aware and contact-rich robot learning.

For each RoboDyna task, we use 50 expert demonstrations from the original benchmark, with 45 for training and 5 for validation.

Q2. “Requested comparison: Does the unified interface improve over a matched DA-DP-style delay-conditioned policy, particularly on inference-delay conditions, and what precisely extends LaWAM's physical-time action queries?”

We thank the reviewer for suggesting these important baselines. We have added Delay-Aware Diffusion Policy (DA-DP) to Tables 1–4. DA-DP improves robustness to inference latency in most settings, but Chronos Policy achieves higher average success overall. Similar to latency compensation, DA-DP primarily addresses action delays, while Chronos Policy jointly models observation and action timing, allowing it to handle varying sensor frequencies, observation delays, and inference latency.

We also agree that LaWAM is relevant to our work. LaWAM uses additive sinusoidal embeddings of physical action time, which is closely related to our AddTime baseline. Specifically, LaWAM uses these embeddings to align action tokens from datasets with different control frequencies, demonstrating the importance of physical time in mixed-frequency training.

Chronos Policy goes one step further by jointly modeling when observations are captured and when actions will be executed. This is particularly important for real-world deployment, where sensor frequencies and delays can change unexpectedly during execution. In addition, our controlled comparisons between AddTime and time-aware RoPE show that rotary time encoding provides better robustness under these timing variations. We will clarify the relationship to LaWAM and emphasize this distinction in the revised paper.

Q3. “Clarification: Which latency estimator and overlap/gap scheduler were used, and what errors occurred during hardware evaluation?”

We estimate inference latency based on previous predictions and update the estimate when inference takes longer than expected. On the real robot, inference typically takes around 70–100 ms, with an estimation error of 4–7 ms between the predicted and actual latency. During asynchronous execution, the robot continues executing the previous action chunk while waiting for the next prediction. Once the new prediction arrives, expired actions are skipped, following the same latency compensation mechanism used in UMI.

AuthorsOpenReview ↗

Response to Reviewer iDF5 [2/2]

Q4. “Requested analysis: Report hardware uncertainty and pairing/training-seed details; relate performance to measured latency-prediction errors.”

We thank the reviewer for this suggestion. We now report 95% confidence intervals using stratified nonparametric bootstrap for all real-world experiments (Tables 5–7). Specifically, we resample the original trials with replacement within each evaluation scenario, keeping the same number of trials per scenario. We then calculate the task score for each resample and use the 2.5th and 97.5th percentiles of the resulting score distribution as the confidence interval. We use 10 trials per condition for Dynamic Cup and Insertion, and 20 for Pivoting.

Table 5. Dynamic Cup (10 trials per condition)

ConditionDPDP + RTCChronos
Nominal50.0 [20.0, 80.0]60.0 [30.0, 90.0]80.0 [50.0, 100.0]
Observation35.0 [10.0, 60.0]40.0 [15.0, 65.0]65.0 [40.0, 90.0]
Inference40.0 [10.0, 70.0]50.0 [25.0, 75.0]65.0 [35.0, 90.0]
Combined30.0 [0.0, 60.0]25.0 [0.0, 50.0]55.0 [25.0, 85.0]

Table 6. Peg Insertion (10 trials per condition)

ConditionDPDP + RTCChronos
Nominal95.0 [85.0, 100.0]85.0 [65.0, 100.0]100.0†
Observation65.0 [45.0, 85.0]40.0 [10.0, 70.0]75.0 [50.0, 100.0]
Inference45.0 [15.0, 75.0]55.0 [25.0, 85.0]75.0 [50.0, 100.0]
Combined45.0 [20.0, 70.0]35.0 [15.0, 50.0]65.0 [35.0, 90.0]

Table 7. Pivoting (20 trials per condition)

ConditionDPDP + RTCChronos
Nominal52.5 [37.5, 67.5]57.5 [50.0, 67.5]75.0 [57.5, 90.0]
Observation45.0 [32.5, 57.5]52.5 [37.5, 67.5]75.0 [60.0, 90.0]
Inference40.0 [25.0, 55.0]65.0 [50.0, 80.0]77.5 [67.5, 87.5]
Combined52.5 [37.5, 67.5]47.5 [30.0, 65.0]75.0 [62.5, 87.5]

Each cell reports task score (%) [approximate 95% confidence interval]. Full successes receive a score of 1, partial successes 0.5, and failures 0. † Chronos achieved 10/10 full successes under nominal Insertion. Since bootstrap resampling produces no variation in this case, we report a separate 95% Wilson confidence interval of [72.2%, 100%] for the full-success probability.

Overall, Chronos Policy achieves higher scores across the tested real-world conditions, although the confidence intervals are relatively wide due to the limited number of trials. These intervals reflect uncertainty across evaluation trials rather than training seeds, and we cannot confirm that trials were paired across methods.

Regarding latency prediction, we measured a mean absolute error (MAE) of 4–7 ms in separate hardware tests, as discussed in Q3. This error is relatively small compared to our 20 Hz action frequency (50 ms per action step), and we did not observe a noticeable impact on task success.

Q5. “Deployment context: How does the approach relate to the uncited A2C2 [1], which corrects delayed action chunks using current observations and chunk-position features? Its feedback head differs from joint physical-time conditioning, but provides relevant context beyond RTC alone.”

We thank the reviewer for pointing out A2C2. A2C2 improves responsiveness under inference latency by adding a lightweight correction head that adjusts actions using the latest observations. Its temporal encoding represents the relative position of an action within a chunk, rather than its actual execution time.

Our focus is different. Chronos Policy explicitly conditions on the physical capture times of observations and the requested execution times of actions, allowing it to handle varying sensor frequencies, observation delays, and inference latency within a single policy. In contrast, A2C2 focuses on correcting previously generated actions during execution and does not explicitly model these sensing timing variations.

We agree that the two approaches are complementary, as A2C2's correction head could potentially be added to Chronos for faster feedback during execution. We will include A2C2 and the other missing references in the revised paper.

Dear Reviewer iDF5,

Thank you again for your thoughtful and constructive feedback. As the rebuttal period will end soon, we would like to briefly follow up on our responses to your questions. If any points remain unclear or require further clarification, we would be happy to address them before the discussion period ends.

Best regards,

The Authors

Dear Reviewer iDF5,

We would like to follow up and check again whether our responses have addressed your concerns. We would greatly appreciate any additional feedback.

Thank you again for your time and constructive feedback!

Reviewer_sdQNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Chronos Policy conditions a diffusion-transformer policy on the capture time of each observation and the intended execution time of each action, by replacing index-based positional encoding with RoPE driven by physical timestamps. During training, per-stream sampling rates, delivery delays and the action-chunk start delay are randomized while the true timestamps are kept. It is evaluated on a simulated conveyor task, simulated force-aware pivoting, six DexMimicGen tasks and three real-robot tasks under nominal timing and injected observation and inference perturbations, with a 2x2 ablation (timestamps x randomization) and an additive-time variant.

Critical strengths
  1. The 2x2 ablation gives a clean, useful result: timing augmentation without timestamps hurts (Pivoting nominal 61.6 → 29.8 for DP + rand.), while with timestamps it helps (perturbed average 33.3 → 55.3 on Pivoting, 48.8 → 63.1 on Conveyor).

  2. The change is small and drop-in, preserves nominal performance on quasi-static tasks (76.4 vs. 77.1), and gives large gains on the conveyor task, where knowing frame spacing is physically necessary.

Critical weaknesses
  1. Novelty is narrow and the obvious alternatives are missing. Continuous-coordinate RoPE comes from prior video work (BulletTime, VideoRoPE). On the action side, the action interval is fixed at 20 Hz throughout, so the action-side timing reduces to a single start delay d_t, which is what Delay-Aware Diffusion Policy already conditions on; it is cited as most closely related but not compared. Two standard baselines are also missing: latency matching (dropping stale actions from DP's chunk, as in UMI), and resampling the low-dimensional streams onto the nominal grid using timestamps, trained with the same randomization. DP's drop under delay (Conveyor 73.1 → 1.8 at ≤200 ms) suggests it executed stale actions, but the paper does not say how DP handles d_t. The premise that real systems see such timing variation is also not measured: under the milder Small and ≤100 ms conditions, DP on the standard tasks loses only 1–5 points.

  2. The robustness gains mix timestamps with training-distribution coverage. The main comparisons pit DP without randomization against Chronos with it, and the Small condition and the ≤100 ms delay lie inside the training randomization range. In the controlled ablation, RoPE without randomization is worse than plain DP on Pivoting (33.3 vs. 48.4 perturbed average), so timestamps alone do not generalize to unseen timing. The real-robot results use 10–20 trials per cell with no confidence intervals. On real Pivoting, DP is flat across timing conditions (52.5/45/40/52.5), so the gap there is in nominal performance and comes mostly from the OOD Sideways setting (10 vs. 60), not from timing robustness.

Questions for authors
  1. In simulation, does DP drop actions whose execution time has passed? If not, please add a latency-matched DP and a timestamp-resampled DP trained with the same randomization.
  2. Why are Delay-Aware Diffusion Policy and ManipForce not included as baselines?
  3. Only 16 of the 32 coordinate pairs per head are rotated, and the longest period is 2 s while 1 Hz F/T windows span about 32 s. Is the other half intentionally position-free, and how is phase wrap-around handled?
  4. Can you report results under timing strictly outside the training range, and the measured timing statistics of the real system in the Nominal condition?
4 comments
AuthorsOpenReview ↗

Response to Reviewer sdQN [1/2]

We thank the reviewer for the detailed feedback, we have further clarified our main contributions and implemented key baselines.

W1. “Novelty is narrow and the obvious alternatives are missing.”

We agree that time-aware RoPE has been explored in prior work on video generation. Our main contribution is a unified robot policy formulation that explicitly incorporates both observation and action timing. We further study different ways to encode time into the policies, including additive embeddings and continuous-time RoPE, together with temporal randomization. In addition, we believe existing simulation benchmarks do not systematically evaluate timing effects. We hope our benchmark will encourage further research on timing-aware robot learning.

Q1. “In simulation, does DP drop actions whose execution time has passed? If not, please add a latency-matched DP and a timestamp-resampled DP trained with the same randomization.”

We thank the reviewer for pointing this out. Our original DP baseline in simulation did not discard stale actions. We have now added DP + LC, which skips expired actions. For real-world experiments, our DP baseline already includes this latency compensation mechanism, following the original UMI implementation.

The results show that latency compensation helps reduce the impact of inference delays in both dynamic and contact-rich tasks. However, it still falls short of Chronos Policy, suggesting that explicitly learning the timing of observations and actions provides additional benefits beyond simply discarding stale actions.

We also evaluate Timestamp-Resampled DP on all four tasks. This method uses linear interpolation to resample low-dimensional observations, including positions, joints, gripper states, force/torque, onto their nominal sampling grids. Image frames remain unchanged.

Our results show that resampling low-dimensional observations helps mitigate the effects of sensor frequency changes and delays to some extent. However, image observations are often a major source of sensing delays and cannot be directly interpolated in the same way. As a result, Timestamp-Resampled DP still not able to match the performance of Chronos Policy, which explicitly conditions on the actual timing of all observations.

Q2. “Why are Delay-Aware Diffusion Policy and ManipForce not included as baselines?”

Thanks for pointing out the important baselines. We now include a Delay aware diffusion policy (DA-DP) baseline in Tables 1–4. DA-DP is able to improve the robustness under inference latency in most of the settings. However, Chronos Policy has overall higher mean success in all the settings. Also similar to latency compensation, delay aware diffusion policy only handles the inference latency, and it does not have a good way to handle the observation frequency and delay.

We thank the reviewer for pointing out ManipForce, which is highly relevant to contact-rich tasks such as Pivoting. Its Frequency-Aware Multimodal Transformer (FMT) handles visual and force/torque observations at different sampling rates through frequency-aware embeddings and cross-attention, preserving high-frequency information rather than simply downsampling it. We will discuss this work more explicitly.

However, our focus is different and complementary. ManipForce primarily addresses how to fuse modalities with different sampling rates, while Chronos Policy focuses on unexpected changes in sensor timing and delays during execution. Although ManipForce studies different sampling frequencies, it does not explicitly address these runtime timing variations. We also believe Chronos Policy could be incorporated into ManipForce by using actual observation timestamps in its temporal embeddings, potentially combining the benefits of both approaches.

AuthorsOpenReview ↗

Response to Reviewer sdQN [2/2]

Q3. “Only 16 of the 32 coordinate pairs per head are rotated, and the longest period is 2 s while 1 Hz F/T windows span about 32 s. Is the other half intentionally position-free, and how is phase wrap-around handled?”

Yes. We intentionally apply time-aware RoPE to half of the attention dimensions, leaving the other half unrotated. This allows the policy to combine time-dependent attention with content-based interactions. Similar designs have been explored in existing architectures, such as GPT-NeoX, while Qwen3-VL also allocates different rotary dimensions to temporal and spatial information.

Regarding phase wrap-around, this is an inherent property of RoPE, which uses multiple rotation frequencies rather than a single periodic encoding. Although each frequency wraps around individually, their combination helps distinguish timestamps beyond any single period, so we do not explicitly unwrap the phases.

The rotary frequency range is also a hyperparameter. We have already adjusted it from the default RoPE frequency schedule to better match the physical timescales of our robotics tasks. We agree that the current range does not guarantee unique representations over arbitrarily long histories, but the combination of multiple frequencies helps reduce ambiguity. We will clarify this design choice in the paper, and exploring longer periods would be a useful additional ablation..

Q4. “Can you report results under timing strictly outside the training range, and the measured timing statistics of the real system in the Nominal condition?”

We thank the reviewer for this suggestion. We have evaluated two levels of timing perturbations for each task: a smaller range that is closer to the training distribution and a larger range that extends beyond the training distribution. The larger-range evaluations already exactly out-of-distribution (OOD) timing conditions. As shown in Tables 1–2, Chronos Policy maintains better robustness under these broader timing variations, particularly on the dynamic Conveyor task.

For the real-world experiments, even the Nominal condition includes natural sensing and inference delays. Based on our system measurements, the inference latency is typically around 70 ms, with the 95th percentile below 80 ms across the measured tasks. These measurements reflect the actual system latency without artificially injected delays. We will include these statistics to better clarify the Nominal setting.

W2. “The robustness gains mix timestamps with training-distribution coverage. […].”

We agree that the improvement comes from both explicit timestamp conditioning and temporal randomization. However, we want to emphasize an important finding from our ablation study (Fig. 3): temporal randomization is only effective when the policy explicitly takes time as an input. Otherwise, varying the timing during training introduces ambiguity, as the policy cannot distinguish between observations captured at different times, which can hurt performance.

Dear Reviewer sdQN,

Thank you again for your thoughtful and constructive feedback. As the rebuttal period will end soon, we would like to briefly follow up on our responses to your questions. If any points remain unclear or require further clarification, we would be happy to address them before the discussion period ends.

Best regards, The Authors

Dear Reviewer sdQN,

We would like to follow up and check again whether our responses have addressed your concerns. We would greatly appreciate any additional feedback.

Thank you again for your time and constructive feedback!

Comments on the paper

7 comments
AuthorsOpenReview ↗

General response: contributions and new experiments

We thank all reviewers for the thoughtful feedback. Those feedback help us to further clarify the contribution and expand the baselines and make the paper results more complete. Below, we summarize our key responses before addressing each reviewer individually. Additional videos and interactive comparisons are available on our rebuttal webpage (https://chronospolicy.github.io/rebuttal/).

1) Contributions and Relation to Prior Work

We thank the reviewers for pointing out that time-aware RoPE has been explored in prior work, including BulletTime and VideoRoPE. Our main contribution is not time-aware RoPE itself, but a unified policy formulation that explicitly models both observation and action timing in robot learning. Through our systematic simulation studies, we find that time-aware RoPE is an effective way to incorporate timing information into the policy. Specifically, our contributions are:

1) A unified formulation of observation and action timing. Unlike most existing policies that do not explicitly model time, Chronos Policy conditions on when observations were captured and when actions will be executed. This allows a single policy to handle varying sensor rates, measurement delays, and inference latency. We believe this is our most important contribution: physical time should be an explicit input to robot policies, rather than being implicitly represented by sequence positions.

2) A systematic study of time representations. We compare additive timestamp embeddings and continuous-time RoPE under the same training settings. Our results show that RoPE provides better robustness to timing variations.

3) Benchmarks for real-world timing variations. We introduce a simulation benchmark with dynamic and contact-rich tasks that captures realistic timing variations in sensing and execution. To our knowledge, existing simulation benchmarks do not systematically evaluate these timing effects. We hope this benchmark will encourage further research on timing-aware robot learning.

AuthorsOpenReview ↗

Additional baseline comparisons

2) Additional Baseline Comparisons

We have expanded our comparisons to include DP with latency compensation (DP + LC), RTC, DiscreteRTC, Delay-Aware Diffusion Policy (DA-DP), FutureRTC, and a timestamp resampled DP on Conveyor and Pivoting (Tables 1–2). We also evaluate representative baselines on two dynamic RoboDyna [1] tasks (Tables 3–4).

[1] Zhiyuan Li, Ruiheng Yang, Xuan Zhao, and Amir Rasouli. “Benchmarking Robotic Manipulation under Physically Realistic Dynamics.” NeurIPS 2026 Workshop on Robot Learning with World Models, 2026.

AuthorsOpenReview ↗

Table 1 — Conveyor

Table 1. Conveyor

MethodNominalSensorBroad sensorDelay 0–100 msDelay 0–200 msCombined
DP73.11 ± 7.1915.56 ± 1.020.00 ± 0.0014.22 ± 2.781.78 ± 1.540.00 ± 0.00
DP + LC73.11 ± 6.5817.11 ± 1.680.67 ± 0.6774.89 ± 1.0268.22 ± 5.550.22 ± 0.38
RTC72.22 ± 7.9516.00 ± 2.310.00 ± 0.0075.11 ± 3.0168.22 ± 2.780.22 ± 0.38
Discrete RTC46.00 ± 9.828.67 ± 1.150.00 ± 0.0050.67 ± 15.0743.33 ± 14.890.00 ± 0.00
DA-DP76.89 ± 3.0818.89 ± 2.780.00 ± 0.0023.11 ± 7.345.11 ± 2.040.22 ± 0.38
FutureRTC74.00 ± 5.8116.44 ± 2.140.44 ± 0.7764.89 ± 4.0255.78 ± 4.735.33 ± 2.00
Timestamp Resampled DP74.00 ± 7.6944.67 ± 4.810.44 ± 0.3810.89 ± 1.392.67 ± 1.330.22 ± 0.38
Chronos82.00 ± 3.4677.56 ± 3.0145.56 ± 10.2583.11 ± 2.7876.44 ± 3.3632.67 ± 5.70

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.