Source-linked AI summary
World Models Under Asynchronous Sensor Observations
Akash Anand, Abhay Anand, Yash Vishe
TL;DR
Physical sensors provide heterogeneous, partly stale observations, challenging world models built around synchronized simulation states. The paper compares transformer models with progressively richer schedule information across open-loop, planning, and refresh-gated systems, finding that time-to-refresh helps chiefly when refresh events participate in system dynamics.
Problem
World models must predict from heterogeneous-rate sensors without interpolating unobserved measurements or discarding valid updates, but the value of forward-looking schedule information is unclear.
Method
The paper evaluates capacity-matched transformer world-model rungs using held observations, staleness, and time-to-refresh across open-loop rollouts, model-predictive control, and a refresh-gated actuator system.
Results
Time-to-refresh shows no measurable open-loop improvement, while refresh-gated actuation recovers 7.31 of 9.14 divergence versus 3.02 without it, or 80% versus 33%.
Takeaways & Limitations
Schedule information is useful according to the causal role of refresh events: staleness describes observed history, while time-to-refresh helps when refresh timing enters the dynamics.
Takeaways & Limitations
The evaluation is limited to state-based observations and a fixed family of refresh schedules.
Abstract
from arXiv · showhide
Learned world models typically assume that observations arrive synchronously, an abstraction inherited from simulators that return a complete state vector at each environment step. Physical sensing instead operates at heterogeneous rates, leaving most observation channels stale at any given instant. Interpolating stale channels introduces measurements that were never observed, while downsampling to the slowest sensor discards valid measurements. A natural alternative is to zero-order-hold the most recent reading and provide the known sampling schedule to the model through two features, staleness and time-to-refresh. We test this prediction using transformer world models across three regimes of increasing causal coupling: open-loop rollouts in continuous-control locomotion, closed-loop model-predictive planning in which each learned model serves as the planner dynamics, and a linear latched-actuator system in which refresh events apply a zero-order-held command to the plant. Our findings show that the effectiveness of time-to-refresh depends on the causal role of the sampling schedule, specifically when refresh events affect the system rather than merely report its state. These results establish when sampling schedules provide useful information for predictive world models operating under asynchronous physical observations.
1 Introduction
The paper addresses asynchronous physical observations by holding each channel’s latest reading and exposing schedule information to transformer world models. Across increasingly coupled settings, time-to-refresh helps when refresh events affect system dynamics, but not when they merely report state.
- Motivation: Physical sensors operate at heterogeneous rates, producing stale channels that cannot be reconciled without either inventing unobserved measurements or discarding valid updates.The paper motivates zero-order holding as an alternative representation for asynchronous sensing.
- Approach: The model ladder adds schedule information progressively: A-ZOH receives held observations, F-blind adds staleness, and F adds time-to-refresh.The primary comparison F − F-blind isolates the contribution of forward-looking refresh timing while matching architecture and input width.
- Evaluation: The evaluation spans open-loop locomotion rollouts, model-predictive control, and a linear system where refreshes gate actuator updates.These settings increase the causal coupling between the sampling schedule and system evolution.
- Results: −0.0005 ± 0.0066 pooled sensor-faithful effect shows no measurable time-to-refresh improvement in open-loop prediction.Under model-predictive control, schedule-information differences remain smaller than ordinary world-model error.
- Results: 80% recovery with time-to-refresh versus 33% without it occurs when refresh events control actuator updates.F recovers 7.31 of the 9.14 divergence, compared with 3.02 for F-blind; the effect grows as refresh schedules become less predictable.
2 Method
The method represents asynchronous sensing with zero-order-held observations plus backward-looking staleness and forward-looking time-to-refresh. It tests a causal prediction across schedule regularity, three system settings, capacity-matched model rungs, and multi-step rollout evaluation.
- Asynchronous observations: Each sensor channel contributes its latest zero-order-held measurement rather than an interpolated value between refreshes.The representation preserves observed values while allowing channels to refresh independently.
- Schedule features: Staleness records how long a held value has been propagated, whereas time-to-refresh records when that value will next be corrected.The two features encode past observation history and future schedule information, respectively.
- Theory and scope: Under exogenous future sensing, conditioning on time-to-refresh leaves future-state expectations unchanged for point prediction, while staleness remains part of the history.The claim is conditional on refresh processes fixed independently of the plant and excludes state-dependent or controller-coupled schedules.
- Schedules: The experiments vary refresh predictability from periodic schedules to ±50% jitter while holding mean periods fixed.Periodic timing is recoverable from phase, whereas ±50% jitter makes the next refresh unavailable from step count alone.
- Experimental design: The three acts separate refreshes that reveal state, affect actions indirectly through planning, or directly latch applied control.The predicted ordering depends on both causal strength and whether refresh timing is redundant or genuinely informative.
- Model ladder: A-ZOH, F-blind, and F receive held observations, observations plus staleness, and all three inputs, respectively.F-blind and F use the same width and parameter count because time-to-refresh is zeroed rather than removed.
- Model ladder: The primary contrast is computed with matched architecture, optimizer, data, input width, and parameter count so gains reflect information rather than capacity.The contrast F − F-blind isolates forward-looking timing.
- Evaluation: Models use a causal Transformer to predict one step ahead, then consume their own predictions during multi-step rollouts.Evaluation includes horizons H ∈ {5, 15, 50}, sensor tiers, and a primary slow-tier metric at H=50.
3 Experiments and Results
The experiments compare schedule-information variants across prediction, control, and paired-world settings. Time-to-refresh helps substantially when refresh timing changes plant dynamics, but not when refreshes merely reveal state or provide weak indirect coupling.
- Evaluation: Tables report three model runs and the paired contrast Δ(ttr) = F − F-blind, with intervals computed from paired differences.Prediction cells average three training seeds, while control cells average evaluation seeds.
- Act 1: Exogenous sensing: In open-loop prediction, pooled Δ(ttr) is +0.0310±0.0490 under cut-off rollout and −0.0005±0.0066 under sensor-faithful rollout.Only one of eighteen prediction cells excludes zero, and that effect is negative.
- Act 2: Weak coupling through control: Control provides no evidence that ttr improves planning, with no cell showing help and performance limited more by learned-dynamics quality than schedule information.On Ant, Δ(ttr) is +31 ± 102 periodically and −73 ± 44 under ±50% jitter.
- Act 3: Refresh-gated actuation: Under causal actuation with ±50% jitter, F recovers 7.31 of the target 9.14 separation versus 3.02 for F-blind, yielding Δ(ttr) = +4.29 ± 1.77.The paired-world probe sets the target by the worlds rather than by a model.
- Interpretation: The causal advantage nearly vanishes when refreshes only reveal state, reaching at most +0.14, about a thirtieth of the causal effect on the same scale.Across the three acts, schedule knowledge is useful in proportion to how strongly the schedule participates in system dynamics.
4 Related Work
World-model research largely assumes synchronous observations, while sampled-data control and related reinforcement-learning work study intermittent sensing, actuation, delays, and event-triggered sampling.
- World models: World models span recurrent state-space, discrete-token, diffusion, and latent-planning architectures, but generally assume complete observations on a common clock.This assumption persists despite deployment of some world models on physical systems.
- Adjacent literature: Sampled-data and networked-control research studies zero-order holds, intermittent observations, event-triggered sampling, and age of information.Related reinforcement-learning work likewise examines delayed or non-immediate sensing and actuation.
5 Conclusion
The study finds that schedule information has different value depending on whether refreshes merely reveal state or directly participate in system dynamics. Staleness is useful for interpreting held measurements, whereas forward-looking refresh timing becomes useful when refreshes affect state evolution.
- Conclusion: When refreshes only provide observations, time-to-refresh does not improve prediction across environments, schedule irregularities, or evaluation settings.When refreshes affect system dynamics, time-to-refresh becomes informative and improvements grow with schedule irregularity.
- Practical implication: Schedule-information choices should depend on whether a refresh only reveals the system state or directly affects how that state evolves.The conclusion distinguishes backward-looking staleness from forward-looking refresh timing.
6 Limitations
The evaluation is limited to state-based observations and a fixed family of refresh schedules, leaving broader sensing modalities, delays, and physical systems for future work.
- Scope: The evaluation covers only state-based observations and a fixed family of refresh schedules.Future work should consider multimodal world models, communication delays, and larger physical systems.
A Experimental Detail
The experiments compare schedule-information variants across prediction, control, and horizon analyses using paired contrasts and interval estimates. The evaluation varies schedules, environments, seeds, and rollout protocols to test whether time-to-refresh effects persist across conditions.
- Experimental design: The prediction grid crosses three model rungs, three environments, three schedules, and three seeds, yielding 81 training runs.The refresh-gated study separately crosses three rungs, two worlds, two schedules, and three seeds, yielding 36 training runs.
- Paired contrasts: Prediction contrasts pair F with F-blind within the same environment, schedule, and training seed, isolating the time-to-refresh feature.The prediction intervals use paired training-seed differences, while control contrasts are paired by evaluation seed.
- Horizon analysis: Table 3 measures Δ(ttr) in slow-tier ZOH skill against rollout horizon for periodic and ±50% schedules under cut-off and sensor-faithful protocols.Each cell pools nine paired differences with a 95% t-interval; the table tests whether effects grow at longer horizons.
- Horizon analysis: Under sensor-faithful evaluation, the contrast shrinks with horizon, while under cut-off it is non-monotone and returns downward by H=50.Only one intermediate cell excludes zero, and it does not persist to the reported horizon.
A.1 Aligning the training objective to the evaluation horizon
The authors test whether the open-loop null could result from training on one-step prediction while evaluating 50-step rollouts. Horizon-aligned fine-tuning changes the objective, but not the time-to-refresh result.
- Motivation: Models are trained on one-step prediction but scored on 50-step rollouts, potentially giving a horizon-scoped feature little gradient.The fine-tuning comparison matches base checkpoint, epochs, learning rate, and selection metric between rollout-aligned and teacher-forced arms.
- Gate: +0.0043 ± 0.0020 is the pooled gate between rollout-aligned and teacher-forced arms, excluding zero and showing that the intervention took effect.The gate excludes zero in five of six metric-by-environment readings; HalfCheetah’s skill column is the exception.
- Time-to-refresh effect: −0.0043 ± 0.0051 is pooled Δ(ttr) after fine-tuning, compared with +0.0007 ± 0.0064 before fine-tuning.One of twelve cells excludes zero, and it is negative; none is positive.
- Metrics: Table 4 reports the gate and Δ(ttr) on slow-tier ZOH skill and transition MSE, with means and 95% t-intervals.Positive values indicate that time-to-refresh helped; for the error metric, the difference is oriented accordingly.
B Detailed Analysis
Across the detailed analyses, the only substantial positive time-to-refresh effect appears when refresh events change the dynamics by latching an actuator. Reveal-only refreshes produce much smaller effects, while most other cells are near zero or negative.
- Causal coupling: +4.29 is the only positive and large effect among 32 cells, occurring in the irregular-schedule constructed environment where refreshes latch the actuator.The target separation is 9.14, making this the dominant contrast in the detailed analysis.
- Causal coupling: +0.14 occurs in reveal cells, roughly thirty times smaller than the causal actuator effect.Three of the other five intervals excluding zero are negative, indicating small costs rather than benefits.
- Analysis axes: The analysis varies whether refreshes change dynamics, whether timing is inferable from the schedule, and whether refreshes announce themselves in the data.These axes correspond to causal versus reveal refreshes, ±50% versus periodic timing, and cut-off versus sensor-faithful protocols.
B.1 When Refresh Timing Pays
Refresh timing helps only when refresh events causally alter the system, are not inferable from other signals, and are not already revealed by the observation stream. The effect is strong for a latched actuator, absent in open-loop prediction, and weak in model-predictive control.
- Causal coupling: In the latched-actuator setting, F separates the two worlds by 7.31 of 9.14 divergence (80%), versus 3.02 without time-to-refresh (33%).The advantage is +4.29 ± 1.77 on the raw trajectory-separation metric.
- Timing identifiability: Time-to-refresh becomes redundant when timing is inferable from a periodic schedule, reducing the contrast to +1.07 ± 3.62 despite rung spans of 4.07 to 6.05.The result indicates that the timing feature must provide information unavailable from counting or other signals.
- Observation leakage: When refreshed observations reveal that a latch occurred, the advantage disappears: the effect is −0.75 ± 0.73, with A-ZOH reaching 8.36 versus F at 7.67.Under sensor-faithful observations, the data itself announces the update event.
- Open-loop prediction: Open-loop prediction provides no measurable time-to-refresh benefit: the pooled sensor-faithful effect is −0.0005 ± 0.0066, with every per-environment interval containing zero.The MuJoCo physics never reads the refresh mask, so jitter does not make time-to-refresh causally informative.
- Closed-loop planning: Model-predictive control satisfies the causal condition only weakly, with schedule effects small and world-model fidelity remaining the binding constraint.The planner reaches 85–93% of the fresh-observation reference only on Ant, versus 8–18% on HalfCheetah and 14–28% on Walker2d.
- Design criterion: The practical criterion is an irregularly clocked held actuator observed over a separate channel, whose wiring can be checked before training.Sensor schedules report state, whereas genuinely gated actuation schedules enter the system dynamics.