acceptodds
Under review as a conference paper at ICLR 2027

Physical Time As A Missing Dimension of Robot Policy Memory Dynamics

Abstract

Existing robot policy memories typically evolve with observation sequences, implicitly treating each update as the same amount of physical time. Yet identical observation histories can span different durations and require different actions, exposing physical time as a missing dimension of policy memory. We introduce Temporal Flow Policy (TFP), which learns memory dynamics to retain past events and evolve the memory state over measured time intervals. Observations and prior memory determine the content and time scales of each transition; elapsed time determines how far the state advances. TFP conditions diffusion and flow-matching policies through this memory state and is trained with their native imitation objectives. We evaluate TFP with Diffusion Policy and through controlled comparisons of memory designs. These experiments show that successful timing depends on both measured intervals and retained history. TFP achieves higher average success than the recurrent and state-space memories we equip with the same intervals and than a variant that feeds the interval to a learned gate. Memory readouts recover the time since an event and motion information after occlusion. Changing the supplied intervals systematically delays or advances actions, linking time inputs to behavior. Real-robot tasks requiring timing or memory, together with LIBERO evaluations, demonstrate benefits in physical and general manipulation. These results suggest viewing robot memory not as stored observation history, but as a state that retains past events and evolves with physical time.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Learning, Temporal Memory, Continuous-Time Modeling, Memory-Augmented Policies, History-Dependent Decision Making

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BbDVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes to take physical time (the observation cadence) into account through exponential decay (a local ODE) for robot memory.

So at inference time, the robot’s memory adjusts its “retention” to the real physical cadence, even for cadences not seen in training, hence better performance.

Neither the exponential decay nor the use of irregular physical time is new (see GRU-D, LTC and CfC). The novelty lies in combining this with robot memory.

Experiments show that simply feeding the cadence as an input does not help when training uses a single fixed cadence. In simulation, TFP shows gains over the other memories under changed cadences, but on the real robot it is only compared with policies without memory.

Critical strengths
  1. The idea is interesting and natural: robot memory should evolve with physical time, not with the number of observations. TFP implements it as a small plug-in memory that works with two very different backbones, a Diffusion Policy trained from scratch and π0.5.

  2. The mechanism is tested directly. With the trained weights fixed, feeding a constant 0.1 s interval instead of the measured one sharply reduces success. Rescaling the supplied clock delays or advances the robot’s actions accordingly, and probes recover the time since an event from the memory.

Critical weaknesses
  1. The comparison does not show whether a learned memory could do the same. All models are trained at a single fixed cadence, so baselines that take the interval as an input (GRU+Δt, the learned Δt gate) never see it vary and cannot learn to use it for sure. Training the baselines with varied cadences would show how much of TFP’s advantage comes from its built-in decay (or it is more like a data problem, once we have data, the model learns itself).

  2. The real-robot experiments show that memory helps, not that physical time does. TFP is compared only with memoryless reactive policies, at the training cadence, where a step-counting memory such as a GRU could also time the tasks.

Questions for authors

Basically same as weakness:

  1. If GRU+Δt, the learned Δt gate, or a Mamba-style SSM were trained with varied observation intervals, would they match TFP? Cuz that parameterization should be sufficiently rich. Ideally the collected data should also just have irregular time cadence, so what I wonder is that how bad those baselines generalize, even with the right data.

  2. On the real robot, it is the same thing, the current experiments demonstrate the memory helps, which is kinda expected. Would be great to have a simple comparison to a step-counting memory (e.g., a GRU) and TFP with the interval fixed at 0.1.

1 comment
AuthorsOpenReview ↗

Response to Reviewer BbDV

We thank the reviewer for pointing out that our original real-robot comparison did not isolate the contribution of physical-time dynamics from persistent memory. We have conducted additional experiments to address both concerns.

1. Mixed-Cadence Training

We additionally evaluate mixed-cadence training on π0.5 Button Press, using resampled observation schedules and measured intervals during training.

Model10 Hz5 HzJitterDropoutMean
GRU+Δt1428322023.5
Learned Δt gate2238322629.5
S4D+Δt9492888289.0
TFP9296949093.0

TFP achieves the highest mean success (93.0%), compared with 89.0% for S4D+Δt. While S4D+Δt performs slightly better at 10 Hz, TFP achieves higher success under all three non-reference schedules. Thus, training with varied observation intervals alone does not fully reproduce TFP's performance.

2. Additional Real-Robot Comparisons

Following the reviewer's suggestion, we added a plain GRU baseline and a fixed-interval TFP variant to the five Galaxea A1 tasks. Together with the time-aware recurrent and state-space baselines reported in our additional experiments, the results are:

MethodObject SwappingCountingWater PouringTea SteepingStaged HeatingAvg.
Plain GRU505550454048
GRU+Δt556555555056
S4D+Δt606545154546
TFP fixed-0.1556550404050
TFP measured (ours)607560656064

All methods use the same Diffusion Policy backbone and real-robot task protocol, with 20 evaluation trials per task.

Plain GRU does not receive explicit time intervals and serves as an observation-count-based memory baseline. TFP fixed-0.1 uses a constant nominal interval, whereas TFP measured uses the actual elapsed time between consecutive policy queries.

TFP measured achieves 64% average success, outperforming Plain GRU by 16 percentage points and TFP fixed-0.1 by 14 points. It also achieves higher success than both variants on all five tasks.

The relatively small difference between Plain GRU and TFP fixed-0.1 (48% vs. 50%) suggests that the improvement is not explained simply by introducing a different recurrent memory architecture. The additional gain with measured intervals provides evidence that accounting for actual elapsed physical time benefits real-robot manipulation.

Together, these experiments address the role of both persistent history and measured physical time in TFP, beyond simply adding recurrent memory or providing Δt as an input feature.

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

This paper introduces Temporal Flow Policy (TFP), a time-aware recurrent memory module for robot policies. The authors argue that conventional memory mechanisms update with observation sequences without explicitly accounting for elapsed physical time, which can be problematic when observation intervals vary. TFP addresses this by predicting a candidate memory state and adaptive time constants, then updating the memory through an exponential rule based on the measured time interval. The resulting memory conditions action generation in Diffusion Policy and pi0.5 through adaptive normalization and is trained using standard imitation learning objectives. Experiments on three simulated timing tasks demonstrate improved robustness to changes in observation frequency compared to recurrent and state-space baselines. The authors also report improvements on five real-robot manipulation tasks and on LIBERO, alongside ablations and memory probing experiments.

Critical strengths
  1. Incorporating elapsed physical time directly into recurrent memory updates is a reasonable approach to handling irregular observation intervals.

  2. Comparisons with GRU, S4D, and LTC, including time-aware variants, demonstrate improved robustness across observation schedules.

  3. Clock-scaling interventions and memory probes provide convincing evidence that the policy uses temporal information to guide actions.

Critical weaknesses
  1. Unclear motivation and problem scope. The paper presents physical time as a missing dimension of robot policy memory, but this claim seems too broad. Under a fixed observation frequency, recurrent models can implicitly track elapsed time through observation count, while timestep embeddings can make this information explicit. The actual problem appears to be robustness to irregular observation intervals, which should be distinguished more clearly from the general problem of learning memory.

  2. Limited methodological novelty. The proposed memory update is closely related to existing continuous-time recurrent models, particularly LTC and CfC. The main architectural difference is the separate prediction of candidate states and adaptive time constants. While the experimental results are promising, the paper does not sufficiently establish why this particular parameterization constitutes a significant methodological advance over existing continuous-time memory mechanisms.

  3. Insufficient evidence that the proposed dynamics, rather than recurrent memory itself, explain the real-robot improvements. The controlled simulations include several time-aware baselines, which is a strength. However, the real-robot experiments compare TFP only against reactive policies. The large improvements on object swapping, counting, and other tasks could therefore result primarily from introducing persistent memory, rather than from modeling physical time. Evaluating GRU, S4D, and their time-aware variants on the same real-robot tasks would substantially strengthen the conclusions.

  4. Limited evaluation of general memory capabilities. The three controlled simulation tasks primarily require tracking elapsed time or retaining a small amount of motion information. These experiments do not establish whether TFP can preserve arbitrary historical information over long horizons, handle multiple independent events, or resist interference between memories. LIBERO is not specifically designed to evaluate these capabilities. Memory-intensive benchmarks such as RoboMME (Dai et al., 2026) and MIKASA-Robo (Cherepanov et al., 2026) would provide more informative evidence about the generality of TFP.

  5. Long-horizon training and computational costs are insufficiently analyzed. Appendix A.1 describes truncated BPTT for LIBERO, while Appendix D.8 reports eight-chunk unrolls and substantial GPU requirements. Although hidden states persist across training segments, gradients cannot propagate through their boundaries. It remains unclear how effectively the memory learns to retain information that becomes relevant hundreds of steps later, or how the method's training cost compares with alternative recurrent memories under matched computational budgets.

  6. Insufficient implementation details for pi0.5. Several important architectural details remain unclear, particularly the source of the memory input features, what exactly is cached from the frozen visual-language backbone, and how memory conditioning interacts with the action expert's attention and adaptive RMS normalization.

  7. Limited causal evidence for memory use in general manipulation. The clock interventions convincingly demonstrate sensitivity to elapsed time on the controlled tasks. However, the LIBERO analysis is less conclusive. Linear probing establishes that information is recoverable from memory, while the memory-replacement experiment measures action differences that the authors acknowledge may also arise from action sampling. Stronger counterfactual interventions are needed to demonstrate that task-relevant historical information actually determines action selection.

  8. Incomplete evaluation of practical timing generalization. The main temporal experiments train at 10 Hz and evaluate under modified observation schedules, but keep the target task durations fixed. They therefore demonstrate generalization to new observation intervals, not necessarily to new temporal requirements or longer memory horizons. Moreover, the real-robot experiments are conducted under nominal timing conditions, leaving the practical benefits under variable inference latency and asynchronous execution insufficiently validated.

Questions for authors
  1. Motivation. Under a fixed observation frequency, recurrent models can implicitly track elapsed time through observation count, while temporal embeddings can explicitly encode timestep information. Could you clarify what is fundamentally missing from these approaches and under which conditions continuous-time memory dynamics become necessary?

  2. Novelty and theoretical justification. The proposed exponential update is closely related to LTC/CfC. What is the main methodological contribution beyond separately predicting candidate states and time constants? Moreover, the interval-composition property holds only for fixed coefficients, whereas TFP recomputes them at every query. How does this property justify robustness to changing observation frequencies in practice?

  3. pi0.5 implementation. What exactly are the memory inputs: SigLIP features or intermediate PaliGemma hidden states? What is cached during training, and how does memory conditioning interact with the action expert's attention and AdaRMS layers?

  4. BPTT and long-term credit assignment. Appendix D.8 reports eight-chunk unrolls for pi0.5. How can the memory learn to preserve information that becomes relevant hundreds of steps later when gradients are truncated between segments? An ablation on BPTT length and training memory consumption would help assess scalability.

  5. Real-robot comparisons. Why are GRU, S4D, and their time-aware variants not evaluated on the real-robot tasks? Comparing TFP only against reactive policies makes it difficult to determine whether the improvements come from the proposed temporal dynamics or simply from adding recurrent memory.

  6. General memory capabilities. The controlled tasks mainly require tracking elapsed time or retaining simple motion information. How well does TFP handle long-horizon tasks requiring arbitrary historical information, multiple cues, or resistance to memory interference? Evaluations on memory-intensive benchmarks such as MIKASA-Robo or LIBERO-Mem would help establish the generality of the approach.

  7. First-observation initialization. Appendix A.2 states that LIBERO uses zero initial memory and dt_0 = 0. According to Equation 3, this implies that the first observation cannot update the memory. How is this handled when task-relevant information is visible only at the beginning of an episode?

  8. Causal use of memory. The memory probes demonstrate that historical information is recoverable, but this does not establish that the policy uses it for action prediction. Could you provide controlled counterfactual memory interventions, particularly on LIBERO, where the current observation is fixed and only task-relevant memory information is changed? This would help distinguish actual memory conditioning from merely storing decodable information.

2 comments
AuthorsOpenReview ↗

Response to Reviewer FedN (1/2): Evidence and Clarifications

We thank the reviewer. Each central TFP claim has a targeted behavioral test. Button/Passage require vanished cues; Interception requires tracking through occlusion. Memory/clock interventions test functional roles. Hardware tests sequences and durations; LIBERO tests general manipulation.

1. Motivation and contribution

At fixed frequency, update count approximates time; with irregular observations, equal counts span different durations. TFP uses Δt\Delta t as a memory integration interval, not just an input. Candidate μ\mu and rate λ\lambda depend on observation/history; memory conditions diffusion/flow-matching actions.

For fixed μ,λ\mu,\lambda, Φδ(h)=μ+e−δλ⊙(h−μ)\Phi_\delta(h)=\mu+e^{-\delta\lambda}\odot(h-\mu) obeys Φδ2∘Φδ1=Φδ1+δ2\Phi_{\delta_2}\circ\Phi_{\delta_1}=\Phi_{\delta_1+\delta_2}. This is local composition; μ,λ\mu,\lambda change with observations. Compared with LTC/CfC (Appendix A.3), TFP predicts content/rates independently and integrates them into generative policies. Removing history-dependent content drops Button-Press success 94%→16%. Static rates give 60/54.7/48.7% at 5 Hz/jitter/dropout versus 95.3/94.7/90% for TFP; both use actual Δt\Delta t.

2. Functional memory and clock interventions

Memory use is tested by frozen-weight interventions, beyond probes. On Button/Passage/Interception, normal success is 88/97.3/55.3% (Diffusion Policy) and 94/90/47.3% (π0.5\pi_{0.5}). Resetting memory every query yields 0% across all six cases (Table 8c,d). Each task requires information no longer visible. Appendix D.6 provides probes.

Clock intervention (Table 8a,b; Fig. 4) changes only Δt\Delta t with weights and execution fixed. A constant 0.1-s clock harms success under changed cadence; scaling Δt\Delta t by 0.5 delays actions and by 2 advances them. Clock scaling shifts the timing of observed Button Press releases and Passage entries in the expected direction (Fig. 4).

3. Memory-dependent manipulation and LIBERO

On Galaxea A1, Object Swapping, Counting Pick-and-Place, and Water Pouring test sequence memory; pouring also requires sustained control. Tea Steeping requires 3.0±0.53.0\pm0.5-s immersion, and Staged Heating two ordered intervals. Across five tasks (50 demos/task; 20 trials/method), TFP raises mean success 17%→64% (Diffusion Policy) and 26%→76% (π0.5\pi_{0.5}; Table 3).

Matched GRU/S4D/TFP achieve 95.6/96.6/98.7% on LIBERO, 91.1/90.0/93.7% on LIBERO-plus, and 89.1/92.5/96.9% on Long-10 (Table 4; three seeds). Together, they validate the functional role of persistent memory and its benefits across general manipulation settings.

4. New training analyses and cost

Mixed-cadence training (matched resampling, actual intervals, 2,000 steps; 50 trials/schedule): 10-Hz/5-Hz/jitter/dropout mean success is 23.5% GRU+Δt\Delta t, 29.5% learned gate, 89% S4D+Δt\Delta t, 93% TFP; at 5 Hz/jitter/dropout, TFP 96/94/90% vs. S4D 92/88/82%.

BPTT ablation (π0.5\pi_{0.5} Button Press; 200 trials/setting): full-prefix 93%, K=16K=16 83.5%, K=8K=8 19.5%, K=4K=4 30.5%. All retain full forward history; KK limits gradients. Controlled tasks use full prefixes; LIBERO carries state across segments with eight-chunk truncated BPTT (Appendix A.1). Persistent state costs O(dh)O(d_h); activation memory scales with unroll. Largest π0.5\pi_{0.5} runs use >100 GB and ~80 h on four H200s (Appendix D.8).

5. π0.5\pi_{0.5} implementation and first frame

For π0.5\pi_{0.5} Button Press, memory receives pooled SigLIP features projected into PaliGemma input space before its Transformer, plus input-token text means and robot state. Three 256-d projections form a 768-d input to 256-d memory. Training caches visual tokens/summaries, instruction summaries, and states, not PaliGemma activations/KV states; its Transformer runs normally. Memory projects to 1,024 d, is added after the flow-time MLP, and conditions action-expert AdaRMS without new attention tokens/masks. It updates per policy query, not solver step. LIBERO uses visual/state memory and native language context (Sections 3.3, A.1–A.2).

For LIBERO, h0=0,Δ0=0h_0=0,\Delta_0=0 define the time origin. The backbone still sees the first frame; relevant cues persist over later queries and enter memory at positive intervals. A cue appearing only at the first query and disappearing before the next is a boundary case unlike the tested tasks, and doesn't commonly happen in real-world scenarios. Controlled runs also support a 0.1-s bootstrap interval (Appendix A.2).

6. Practical timing and follow-up

Cadence tests preserve physical target durations; clock interventions isolate elapsed-time dependence. Inference measures wall-clock intervals; hardware Steeping/Heating tests timed behavior. Matched real-robot GRU/S4D results will follow in response 2/2.

AuthorsOpenReview ↗

Response to Reviewer FedN (2/2): Additional Real-Robot Baselines

We thank the reviewer for suggesting additional recurrent baseline comparisons on real robots. We evaluated Plain GRU alongside two time-aware baselines, GRU+Δt\Delta t and S4D+Δt\Delta t, across all five Galaxea A1 tasks. All methods use the same Diffusion Policy backbone, with 20 evaluation trials per task and method.

Real-Robot TaskPlain GRUGRU+Δt\Delta tS4D+Δt\Delta tTFP (Ours)
Object Swapping50%55%60%60%
Counting Pick-and-Place55%65%65%75%
Water Pouring50%55%45%60%
Tea Steeping45%55%15%65%
Staged Heating40%50%45%60%
Average48%56%46%64%

TFP achieves the highest or tied-highest success rate across all five tasks, outperforming Plain GRU by 16 percentage points, GRU+Δt\Delta t by 8 percentage points, and S4D+Δt\Delta t by 18 percentage points on average.

Plain GRU underperforms TFP on all five real-robot tasks, achieving only 48% average success compared with TFP's 64%. Providing elapsed-time information as an additional input does not close this gap: GRU+Δt\Delta t and S4D+Δt\Delta t achieve only 56% and 46%, respectively. These results directly counter the explanation that TFP's improvements arise merely from adding recurrent memory or supplying Δt\Delta t as an input feature, and provide further evidence for the effectiveness of TFP's adaptive physical-time memory dynamics.

Reviewer_jL3HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TFP is a recurrent memory for diffusion/flow policies. Its state relaxes toward a candidate, with per-channel time constants predicted from the current features and the previous memory, integrated in closed form over the measured interval Δn (Eqs. 2–3). The memory conditions Diffusion Policy or π0.5 through adaptive normalization (Eq. 5). Models trained at 10 Hz are tested at 10 Hz, 5 Hz, jitter and dropout on three simulated timing tasks, against GRU, S4D and LTC memories with and without Δt (Table 1). The paper adds clock-scaling, reset and probe analyses (Fig. 4, Tables 8–10), ablations (Table 2), five real-robot tasks against reactive policies (Table 3), and LIBERO (Table 4).

Critical strengths
  • A clean diagnostic with causal evidence. Training at one cadence and testing at several (L260–265) exposes the failure of step-indexed memory: GRU on DP Passage scores 138/150 at 10 Hz and 6/150 at 5 Hz (Table 1). Scaling the supplied clock shifts action timing in the predicted direction in all 50 paired trials (π0.5 Button; both backbones on Passage; §4.2, Fig. 4). This gives a reusable test for VLA memory.

  • A simple drop-in module. It is best in 22 of 24 Table 1 columns (ties included), across both a from-scratch DP and the π0.5 action expert. This is useful regardless of how the gains are attributed.

Critical weaknesses

The evidence does not show that the gains come from physical-time dynamics

  • The interval-vs.-input-feature question is untested. All models train only at Δ = 0.1 s (L260–262), so Δ-as-input models never see Δ vary. GRU+Δt tracks GRU throughout Table 1 (e.g., DP Button at 10 Hz: 66/150 vs. 64/150). The learned Δt gate fails even at the training cadence: its 15.0% on DP Passage over four schedules (Table 2) is 90/600, so at most 90/150 at 10 Hz, against 138/150 for plain GRU.
  • TFP's lead over S4D+Δt is not a cadence effect. Task-averaged from Table 1 (150 trials per Button/Passage cell, 900 per Interception cell), the lead is 15.7 pp at 10 Hz vs. 15.3 pp at the other schedules for DP, and 19.4 vs. 2.2 pp for π0.5. Off the training cadence, both reach 420/450 on π0.5 Button. At 10 Hz, Δ is constant: TFP and TFP fixed-Δt are identical there.
  • Consequence. The headline averages of 77.5 vs. 62.1 and 73.3 vs. 66.8 (L309–311) reflect an in-distribution advantage of unidentified origin, while cadence robustness is shared with S4D+Δt. The contribution narrows from "physical time as a missing dimension" to a stronger memory cell whose source of improvement is not established.
Questions for authors
  • If GRU+Δt, the learned Δt gate and TFP are retrained with mixed-cadence sampling on one task, does TFP's lead persist?
  • What drives the 10 Hz lead? Please report the Table 2 ablations per schedule, and add a GRU with matched multi-timescale initialization or CfC with measured intervals. LTC+Δt trails GRU at 10 Hz (DP Passage: 37/150 vs. 138/150).
  • What is the controller rate, and how many actions run per query under each schedule? π0.5 S4D+Δt scores 126/150 on Passage at 5 Hz but 78/150 at 10 Hz, its training cadence.
2 comments
AuthorsOpenReview ↗

Response to Reviewer jL3H (1/2): Mixed-Cadence Training and Ablations

Thank you for the thoughtful review.

The additional experiments clarify the contributions of history-dependent candidate content and adaptive timescales within TFP. Its advantage persists under mixed-cadence training, and matching a GRU's initial retention does not reproduce its performance. The frozen-weight clock interventions separately link measured intervals to action timing.

1. Mixed-Cadence Training

We train TFP, GRU+Δt\Delta t, a learned Δt\Delta t gate, and S4D+Δt\Delta t on π0.5\pi_{0.5} Button Press with mixed observation schedules. All models use 2,000 optimization steps and 50 evaluation trials per schedule.

Table 1. Mixed-cadence training on π0.5\pi_{0.5} Button Press. Success rates (%).

Model10 Hz5 HzJitterDropoutMean
GRU+Δt\Delta t1428322023.5
Learned Δt\Delta t gate2238322629.5
S4D+Δt\Delta t9492888289.0
TFP9296949093.0

TFP achieves 93% mean success versus 89% for S4D+Δt\Delta t, with gains of 4, 6, and 8 percentage points under 5 Hz, jitter, and dropout, respectively. Its mean success also exceeds GRU+Δt\Delta t and the learned gate by 69.5 and 63.5 percentage points. Thus, on this task and under the same training-step budget, the advantage over interval-as-input baselines persists when their training intervals vary. Means weight the four schedules equally throughout.

2. Schedule-Wise Ablation

We report the Table 2 ablations by observation schedule for both backbones and all three controlled tasks. These retain the original fixed-10-Hz training protocol.

Table 2. Schedule-wise ablations across three tasks. Success rates (%). Each cell lists 10 Hz / 5 Hz / Jitter / Dropout / Mean.

VariantDiffusion Policyπ0.5\pi_{0.5}
Button Press
Full TFP88.0 / 94.7 / 76.0 / 66.0 / 81.294.0 / 95.3 / 94.7 / 90.0 / 93.5
Candidate w/o history46.0 / 44.0 / 40.0 / 40.7 / 42.716.0 / 12.0 / 10.0 / 10.0 / 12.0
Static time constants88.0 / 54.7 / 50.0 / 41.3 / 58.594.0 / 60.0 / 54.7 / 48.7 / 64.3
Learned Δt\Delta t gate30.0 / 14.7 / 8.0 / 5.3 / 14.594.0 / 8.0 / 4.0 / 2.0 / 27.0
Delayed Passage
Full TFP97.3 / 96.7 / 98.7 / 96.7 / 97.390.0 / 84.7 / 88.0 / 84.0 / 86.7
Candidate w/o history12.0 / 8.0 / 6.0 / 6.0 / 8.072.0 / 68.0 / 62.0 / 60.7 / 65.7
Static time constants90.0 / 85.3 / 80.0 / 75.3 / 82.720.0 / 14.7 / 10.0 / 8.7 / 13.3
Learned Δt\Delta t gate22.0 / 16.0 / 12.0 / 10.0 / 15.054.7 / 40.0 / 30.0 / 23.3 / 37.0
Interception
Full TFP55.3 / 52.6 / 56.3 / 51.3 / 53.947.3 / 33.2 / 37.1 / 41.0 / 39.7
Candidate w/o history46.0 / 45.3 / 42.7 / 42.7 / 44.234.0 / 32.0 / 30.0 / 30.8 / 31.7
Static time constants46.7 / 45.3 / 44.0 / 43.2 / 44.834.0 / 32.0 / 30.0 / 29.6 / 31.4
Learned Δt\Delta t gate32.0 / 30.7 / 30.0 / 29.3 / 30.533.0 / 31.0 / 29.0 / 27.4 / 30.1

The static-τ\tau variant retains measured Δt\Delta t in the exponential update; the learned-gate variant receives Δt\Delta t as an input feature.

At 10 Hz on Button Press, removing history-dependent candidate content reduces success from 88.0% to 46.0% for Diffusion Policy and from 94.0% to 16.0% for π0.5\pi_{0.5}, while static time constants match Full TFP. This identifies history-dependent candidate content as a key contributor at the reference cadence.

Under 5 Hz, jitter, and dropout, Full TFP exceeds static time constants on Button Press by 40.0, 26.0, and 24.7 percentage points for Diffusion Policy, and by 35.3, 40.0, and 41.3 points for π0.5\pi_{0.5}. Since both variants use measured Δt\Delta t in the exponential update, these gaps demonstrate the additional contribution of adaptive timescales under changed observation schedules.

Adaptive timescales also contribute at a fixed cadence on Delayed Passage: at 10 Hz, replacing them with static constants reduces success from 97.3% to 90.0% for Diffusion Policy and from 90.0% to 20.0% for π0.5\pi_{0.5}.

At the reference cadence, these ablations identify contributions from history-dependent candidate content and, depending on the task, adaptive timescales within TFP. Under changed schedules, measured intervals additionally determine how far those learned dynamics advance in physical time. This distinguishes the memory components supporting 10-Hz performance from the role of elapsed time when observation intervals vary.

AuthorsOpenReview ↗

Response to Reviewer jL3H (2/2): Matched GRU and Controller Execution Protocol

3. Matched Multi-Timescale GRU

We evaluate GRU+Δt\Delta t with matched multi-timescale initialization on π0.5\pi_{0.5} Button Press. The reset gate, update gate, and candidate receive ϕ(Δt)=log⁡(1+Δt/(0.1,s))\phi(\Delta t)=\log(1+\Delta t/(0.1\\,\mathrm{s})) as an additional input.

We use TFP's geometric timescale range and match its initial retention at the 0.1-second reference interval:

τi=geomspace⁡(1.01τmin⁡,,τmax⁡/1.01)i,ui(0)=1−exp⁡(−0.1/τi).\tau_i=\operatorname{geomspace}\left(1.01\tau_{\min},\\,\tau_{\max}/1.01\right)_i,\qquad u_i^{(0)}=1-\exp(-0.1/\tau_i).

Here uiu_i weights the candidate in the GRU update. The update-gate input weights are initialized to zero and its bias to logit⁡(ui(0))\operatorname{logit}(u_i^{(0)}). All parameters remain trainable.

Table 3. Matched multi-timescale GRU initialization on π0.5\pi_{0.5} Button Press. Success rates (%). Ordinary GRU+Δt\Delta t results are taken from Table 1 of the manuscript.

Model10 Hz5 HzJitterDropoutMean
Ordinary GRU+Δt\Delta t7.32.730.06.711.7
Multi-timescale GRU+Δt\Delta t4048414.0

Both models were trained for 2,000 steps; the multi-timescale GRU was evaluated over 50 trials per schedule. The matched-initialization GRU achieves 14.0% mean success, compared with the reported 11.7% for ordinary GRU+Δt\Delta t, and 4.0% at 10 Hz, compared with 7.3%. Matching initial retention therefore does not reproduce TFP's performance on this task.

The ablations examine which memory components support performance, and the mixed-cadence and matched-GRU controls test the training-interval and initialization explanations. The clock interventions then test whether the policy actually uses elapsed time: with policy weights and the physical execution schedule fixed, changing only the supplied clock shifts action timing in the predicted direction. This directly links the time input to the policy's temporal behavior.

4. Controller Execution Protocol

The action controller operates at 20 Hz and MuJoCo at 500 Hz (25 physics steps per action). Each π0.5\pi_{0.5} query predicts a 10-step action chunk spanning 0.5 seconds.

When a new observation arrives, the policy updates its memory and replans, replacing the remaining unexecuted actions in the queue. Otherwise, the controller continues executing the existing action queue without updating memory.

Table 4. Action execution under different observation schedules.

ScheduleObservation intervalActions executed before next query
10 Hz0.10 s2
5 Hz0.20 s4
Jitter0.05–0.20 s1–4
Dropout0.10/0.20/0.40 s2/4/8

For jitter, intervals are sampled from 0.05,0.10,0.15,0.20\\{0.05,0.10,0.15,0.20\\} seconds. For dropout, the corresponding interval probabilities are 0.6, 0.2, and 0.2, respectively. Inference is synchronous in simulation, and GPU inference wall-clock time does not advance simulated time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

We thank the reviewers for their constructive feedback. In response, we conducted additional controlled experiments, expanded real-robot comparisons, and provided detailed methodological clarifications. These results directly address the central questions concerning the contribution of physical-time dynamics beyond recurrent memory and explicit time inputs.

First, the new experiments further establish the contributions of TFP's memory dynamics. Under matched mixed-cadence training, TFP achieves 93.0% mean success on π0.5 Button Press, exceeding S4D+Δt (89.0%) and the other time-aware baselines, with higher success under all three changed observation schedules. Schedule-wise ablations under the original fixed-10-Hz training protocol identify the distinct roles of history-dependent candidate content and adaptive timescales. Removing history dependence reduces 10-Hz success from 94% to 16% on π0.5 Button Press. Matching a GRU's initial multi-timescale retention also fails to reproduce TFP's performance.

Second, controlled interventions establish the functional roles of memory and physical time. Resetting memory at every query reduces success to 0% across all six controlled task-backbone settings. With policy weights and physical execution fixed, changing only the supplied clock shifts the timing of observed Button Press releases and Passage entries in the expected direction.

Third, the expanded real-robot experiments directly distinguish TFP from recurrent and time-aware alternatives. Across five Galaxea A1 tasks, using the same Diffusion Policy backbone and 20 evaluation trials per method and task, TFP achieves 64% average success, compared with 48% for Plain GRU, 56% for GRU+Δt, 46% for S4D+Δt, and 50% for TFP with a fixed 0.1-s interval. TFP achieves the highest or tied-highest success on all five tasks. These matched comparisons demonstrate benefits beyond simply adding recurrence or supplying time as an input, while the fixed-interval comparison provides direct evidence for the value of measured physical time.

We also clarified TFP's relationship to LTC/CfC, its local interval-composition property, π0.5 implementation, BPTT behavior, first-frame handling, and controller execution protocol. The matched LIBERO, LIBERO-plus, and Long-10 evaluations further support its effectiveness across general manipulation tasks.

Together, the new experiments and clarifications substantially strengthen the methodological, causal, and real-robot evidence supporting TFP's adaptive physical-time memory formulation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.