In-Depth Analysis of RTT Marginal Modeling Technology for Networked Robot Sensors

Registration: European Humanoid Robot Summit 2025

Abstract:This article focuses on the marginal modeling of RTT for networked robot sensor transmission, comparing exponential, log-normal, and log-logistic distributions, improving parameter estimation and Cramér-von Mises tests, and analyzing the technical advantages and disadvantages to support the reliable real-time operation of robots.

In the context of remote control and sensor data transmission for mobile robots, “Round-Trip Time (RTT)” is a key indicator determining system stability. When a robot cannot perform complex calculations locally, sensor data must be sent to a remote site for processing, and control commands must be sent back for execution. This cycle’s RTT exhibits randomness due to network non-determinism, hardware and software delays, etc. If RTT cannot be accurately modeled, the robot may fail to complete tasks due to delays exceeding constraints (e.g., obstacle avoidance lag, control inaccuracies).

Traditional RTT modeling methods (such as time series prediction and machine learning) generally suffer from issues like “high computational cost,” “offline dependency,” and “failure to cover all sources of delay.” In contrast, RTT marginal probability modeling technology discards the time dependency of continuous RTT measurements, achieving approximate modeling at a lower computational cost, particularly suited to the core needs of networked robots for “real-time online” processing and “complete statistical information.” The following sections will deeply analyze the core logic of this technology from three aspects: technical architecture, core modules, and experimental validation, while comparing the advantages and disadvantages of the three core distribution models.

1. Technical Background: Why Choose “Marginal Modeling”? – Core Differences from Traditional Methods

The core demands for RTT modeling in networked robots are threefold: full online processing (modeling within the current RTT cycle without slowing down the control loop), covering all sources of delay (including network, sensor, and hardware/software delays), and providing probabilistic information (e.g., “the probability that the next RTT falls within [100ms, 200ms]”). Traditional methods struggle to meet these demands simultaneously, as detailed below:

Modeling Method Type	Representative Technology	Core Issues Unadaptable to Networked Robots	Advantages of Marginal Modeling
Time Series Prediction	ARIMA, LSTM	Requires a large amount of offline training data, high computational cost (e.g., LSTM training takes minutes), cannot output probability intervals in real-time	Based on the independent and identically distributed (iid) assumption, no training dependency, low computational cost, can output complete probability distributions (including interval probabilities)
Simple Statistical Filtering	EWMA (Exponentially Weighted Moving Average)	Only performs smoothing filtering, cannot detect sudden changes in RTT, cannot provide probability information (can only predict the mean of the next RTT)	Combining goodness-of-fit tests, can quickly detect state changes, outputs multidimensional statistical information (mean, variance, quantiles) based on probability distributions
Filtering and Hidden Markov Models	Kalman Filtering, HMM	Requires pre-defined network/system models (e.g., Kalman filtering requires node parameters), HMM has high computational costs (training takes 5 seconds or more, risking open-loop dangers for robots)	No prior model required, parameters estimated in real-time from samples, adaptable to diverse networked robot scenarios (e.g., intercity/local communication, different hardware)
Machine Learning (Deep Learning)	Transformer, RELM	Lacks statistical robustness (performance in sudden scenarios unverified), time-consuming offline training phase, cannot adapt to dynamic environments of robots	Based on strict hypothesis testing (goodness-of-fit), statistical significance verifiable, fully online process, dynamically adapts to environmental changes

In summary, the core advantage of marginal modeling lies in “sacrificing time dependency for low computational cost, full online capability, and complete probabilistic information.” This sacrifice is entirely acceptable in networked robot scenarios where RTT is typically stable (no long-term trends, only short-term noise and sudden changes) — as long as state changes can be accurately detected, different marginal models can be adapted to different stable segments.

2. Technical Core: Three Major Modules and Design Logic of Three Distributions

The RTT marginal modeling technology consists of three major modules: “core distribution model,” “improved parameter estimation,” and “goodness-of-fit testing,” which work together to achieve “real-time modeling + sudden change detection.” The following sections will analyze the design details of each module and compare the technical characteristics of the three distributions.

1) Core Distribution Model: Why Choose Exponential, Log-Normal, and Log-Logistic?

The foundation of marginal modeling is “fitting RTT samples with probability distributions.” The core reason for choosing these three distributions is their adaptability to the “non-negativity,” “heavy-tailed characteristics,” and “scene diversity” of networked robot RTT. The specific characteristics are compared below:

Distribution Type	Core Adaptation Scene	Key Parameters (Including Location Parameters)	Statistical Moment Characteristics (Mean / Variance)	Core Advantages	Inherent Limitations
Exponential Distribution	Fast communication scenarios (e.g., short delays between sensors), high network load states	α (location, >0), β (shape, >0)	Both exist, simple calculations (mean α+β, variance β²)	Fastest computation speed, simple parameter estimation logic	Low modeling flexibility, cannot adapt to heavy-tailed RTT (e.g., large data packet transmission from cameras)
Log-Normal Distribution	Wired networks (LAN), medium data volume transmission (e.g., small packets from sensors)	γ (location, ≥0), μ (log mean), σ (log standard deviation, >0)	Both exist, requires exponential calculations (mean γ+exp (μ+σ²/2))	Higher flexibility than exponential distribution, adaptable to most wired scenarios	Location parameter estimation may fail, poor adaptation to high-determinacy RTT (low noise)
Log-Logistic Distribution	Complex scenarios (intercity communication, large data packets from cameras), heavy-tailed RTT	a (location, ≥0), b (scale, >0), c (shape, 0.05≤c<0.5)	Expectation exists only when c<1, variance exists when c<0.5	Highest modeling accuracy, strong adaptability to heavy tails, wide scene coverage	Highest computational cost, strict constraints on shape parameters (must be controlled between 0.05~0.5)

Note: All three distributions introduce a “location parameter” (α/γ/a) to avoid RTT having “0 or negative values” — networked robot sensor transmission inevitably has a baseline delay (e.g., hardware response time), and the location parameter precisely characterizes this baseline value, making the distribution more aligned with actual scenarios.

2) Improved Parameter Estimation: Addressing the “Adaptability to Networked Robots” Issue of Traditional MLE

Traditional Maximum Likelihood Estimation (MLE) faces three major pain points in networked robot scenarios: failure of non-negative parameter constraints (e.g., unbiased estimates for exponential distribution may yield α≤0), failure of location parameter estimation (e.g., Cohen method for log-normal distribution may have no zero-crossing), and discretization errors (RTT is measured in milliseconds/nanoseconds, while models are continuous distributions). To address these issues, the technical team proposed targeted improvements, as detailed below:

• Exponential Distribution: “Unbiased Estimation + Fallback Mechanism” Ensures Parameter Validity

  The traditional unbiased estimation formula isα^=min(X)−β^/n、β^=n−1n(Xˉ−min(X))，但当样本存在极端值时，可能导致α^≤0（violating RTT non-negativity). The improvement plan: first perform unbiased estimation, if α^≤0, immediately fallback to “biased estimation” (α^=min(X),β^=Xˉ−min(X),which ensures parameter validity while minimizing bias.

• Log-Normal Distribution: “Cohen Method + Moment Method” Dual Path Safety Net

  The location parameter γ is the core difficulty in estimating the log-normal distribution — the classic Cohen method determines γ by finding the zero point of the function Θ1(γ), but in high noise/low noise scenarios for networked robots, it often results in “no zero-crossing” leading to estimation failure. Improvement plan: first use the bisection method to execute the Cohen method (constraining γ range based on the sample minimum); if it fails, switch to “moment method + trust region reflective optimization”: minimizing the 2-norm of the “first/second/third moments of the sample” and “theoretical moments,” constraining γ∈[10⁻⁹, min (X)-10⁻⁹], ensuring the estimation results align with actual RTT characteristics.

• Log-Logistic Distribution: “Two-Step Optimization” Balancing Precision and Constraints

  This distribution has three parameters: a (location), b (scale), c (shape). Traditional MLE can easily lead to numerical divergence due to parameter coupling. Improvement plan in two steps: ① Initial parameter guess: estimate c based on the moment method (minimizing the error between “sample expectation” and “theoretical expectation”), a references the exponential distribution (taking values near the sample minimum), b is based on the median (bˉ=max(median(X−aˉ),10−6)); ② Refine parameters: starting from the initial guess, minimize the 2-norm of the “log-likelihood partial derivatives,” using trust region reflective algorithms to ensure parameters meet constraints (c∈[0.05,0.5)), avoiding numerical overflow or divergence.

• Common Optimization: Addressing RTT Discretization with “Continuous Distribution Adaptation”

  RTT measurements for networked robots are often discrete values (e.g., nanosecond-level clocks). Directly modeling with discrete distributions would significantly increase computational costs. The technical team experimentally verified that at nanosecond resolution, the error in estimating continuous distribution parameters from discrete RTT is negligible (e.g., in Table 1, the error in the exponential distribution α parameter at 1e⁻⁹ second resolution is only ±0.5%), thus retaining continuous distribution modeling without pre-shifting RTT samples (to avoid cumulative floating-point representation errors, adapting to existing robot software interfaces).

3) Goodness-of-Fit Testing: “Scene-Specific Transformation” of Cramér-von Mises Statistic

The core purpose of goodness-of-fit testing is to “detect whether RTT has undergone a state change” — by testing “whether the current RTT sample comes from the current distribution model” (null hypothesis H0). If H0 is rejected, it indicates a change has occurred, necessitating re-modeling. The technical team abandoned the traditional Anderson-Darling statistic, opting for the Cramér-von Mises statistic (W²), and derived new thresholds adapted to networked robot scenarios, with specific improvements as follows:

• Reason for Statistic Selection: The Anderson-Darling statistic places too much weight on the tails of the distribution, leading to failures in testing when RTT approaches the location parameter (e.g., in low-latency scenarios). In contrast, W² treats all intervals of the distribution equally, naturally handling situations where “RTT approaches the location parameter,” adapting to the diverse delay scenarios of networked robots (from microsecond to second).

• New Threshold Derivation: Monte Carlo Simulation Adapting to “Parameter Sample Estimation”

  Traditional W² thresholds are based on “distribution parameters being known in advance,” but in marginal modeling, parameters must be estimated from current RTT samples, rendering traditional thresholds ineffective. The technical team generated new thresholds through large-scale Monte Carlo simulations:

  ① Extract parameter ranges for the three distributions from real networked robot datasets (including intercontinental communication, Arduino embedded systems, large data packets from cameras across 36 scenarios);

  ② For each sample size (20~10000), generate 10000 random distributions and corresponding RTT samples, calculating the W² statistic;

  ③ Take the 95% quantile of the statistic as the threshold for a “significance level of 0.05” (ensuring Type I error is close to 0.05);

  ④ Use piecewise polynomial interpolation to fit the threshold curve (e.g., log-normal distribution divided into 2 segments, log-logistic distribution divided into 3 segments), achieving “rapid threshold lookup for any sample size,” meeting real-time processing needs.

3. Experimental Validation: Quantitative Comparison of the Advantages and Disadvantages of Three Distributions

The technical team quantitatively compared the performance of the three distributions based on 36 “real + simulated” networked robot scenarios (covering intercontinental/local communication, embedded/high-performance hardware, data volumes from 1 byte to 786KB) across four core dimensions: “computational cost,” “online capability,” “modeling coverage,” and “sudden change detection capability,” with the following results:

1) Computational Cost: Exponential Distribution Optimal, Log-Logistic Highest

Computational cost directly determines “whether real-time online processing is possible” — networked robot RTT typically ranges from 1ms to 1s; if modeling takes longer than RTT, it will cause control loop delays. The experiment was conducted on “Intel Xeon Gold 5317 (48 threads) + NVIDIA RTX A6000” hardware, measuring single-step modeling time (parameter estimation + goodness-of-fit testing):

• Exponential Distribution: Median time ~60ms, 99% of scenarios take <100ms, fully adaptable to RTT≥100ms scenarios (e.g., most sensor transmissions);

• Log-Normal Distribution: Median time ~560ms, one order of magnitude higher than exponential distribution, requiring RTT≥600ms scenarios to be fully online;

• Log-Logistic Distribution: Median time 6~200ms (due to sample size fluctuations), slightly lower than log-normal distribution but higher than exponential distribution, adaptable to RTT≥200ms scenarios, with large sample sizes (e.g., 1000 RTT) taking up to 500ms.

2) Online Processing Capability: Exponential Distribution “Fully Online,” Log-Logistic Second

Online processing capability is measured by “batch size” (how many new RTTs must be waited for to complete current modeling), the smaller the number, the better the online capability:

• Exponential Distribution: In 34 out of 36 scenarios, achieved “batch size = 1” (fully online), only 2 scenarios (#12, #36) had “batch size = 2~4,” optimal online capability;

• Log-Normal Distribution: 22 scenarios achieved fully online, 8 scenarios had batch sizes of 5~20, and 6 scenarios (e.g., #26 high determinacy scenarios) had batch sizes of 100~110, moderate online capability;

• Log-Logistic Distribution: 25 scenarios achieved fully online, 7 scenarios had batch sizes of 5~15, and 4 scenarios had batch sizes of 50~90, online capability better than log-normal but weaker than exponential.

3) Modeling Coverage: Log-Logistic Optimal, Exponential Distribution Lowest

Modeling coverage refers to “the proportion of RTT successfully modeled”; the higher the proportion, the stronger the technology’s adaptability to the scene:

• Log-Logistic Distribution: In 36 scenarios, 35 had coverage ≥0.7, 3 scenarios (#7, #21, #32) had coverage ≥0.95, only #26 (high determinacy RTT, noise < 0.1%) had coverage = 0.68, optimal coverage;

• Log-Normal Distribution: 28 scenarios had coverage ≥0.7, 10 scenarios (e.g., #25 Arduino low data volume scenarios) had coverage = 0.4~0.6, moderate coverage;

• Exponential Distribution: 18 scenarios had coverage ≥0.7, 12 scenarios (e.g., #25, #26 high determinacy scenarios) had coverage = 0, unable to adapt to low noise, heavy-tailed RTT scenarios, lowest coverage.

4) State Change Detection: Exponential “Fast but Fragmented,” Log-Logistic “Accurate but Stable”

State change detection capability is measured by “detection delay” (the number of RTTs from when a change occurs to when it is detected) and “regime length” (the number of RTTs covered by a single modeling):

• Exponential Distribution: Shortest detection delay (most changes <5 RTTs), but very short regime length (average <50 RTTs), prone to “over-detection” (misinterpreting noise as changes), suitable for scenarios requiring extremely high response speed to changes but low continuity requirements for modeling;

• Log-Normal Distribution: Medium detection delay (changes <10 RTTs), medium regime length (average 100~200 RTTs), false detection rate <5%, suitable for balancing speed and continuity;

• Log-Logistic Distribution: Longest detection delay (some changes = 15~20 RTTs), but longest regime length (average 300~500 RTTs), false detection rate <3%, suitable for scenarios requiring high modeling accuracy and continuity, where a small detection delay is acceptable (e.g., robot navigation, precision operations).

4. Overview of Technical Advantages and Disadvantages: Scene Adaptation Determines Choice

Based on the above experiments and analyses, the core advantages and disadvantages of RTT marginal modeling technology (including the three distributions) can be summarized as “sacrificing time dependency for real-time performance, defining scene boundaries by distribution characteristics,” with specific comparisons as follows:

1) Overall Technical Advantages (vs Traditional Methods)

• Low computational cost, adaptable to real-time demands: Compared to ARIMA (offline training takes hours) and LSTM (single prediction takes milliseconds), marginal modeling takes at most 500ms per step, suitable for most networked robot RTT cycles (1ms~1s);

• Fully online processing, no offline dependency: No pre-training data required, parameters estimated in real-time from current RTT samples, goodness-of-fit testing executed simultaneously, adaptable to dynamic environments of robots (e.g., switching from WiFi to 4G);

• Complete statistical information to support decision-making: Not only can it predict the mean of the next RTT, but it can also output information such as “the probability that RTT falls within a certain range” and “variance,” aiding in optimizing robot control strategies (e.g., adjusting obstacle avoidance safety thresholds based on RTT variance);

• Strict statistical robustness: Goodness-of-fit thresholds based on Monte Carlo simulations ensure Type I error remains stable at 0.05, making it more suitable for safety-sensitive scenarios (e.g., industrial robots) than machine learning methods (which lack statistical guarantees).

2) Overall Technical Disadvantages

• Dependent on the assumption of “RTT stability”: If the robot is in extremely dynamic scenarios (e.g., frequent network switching, sudden sensor failures), the increased non-stationarity of RTT may invalidate the marginal modeling’s “independent and identically distributed” assumption, reducing modeling accuracy;

• Distribution selection must match the scene: There is no “universal distribution”; it must be selected based on scene characteristics (e.g., exponential for fast scenarios, log-logistic for complex scenarios), increasing engineering complexity;

• Poor adaptation to high determinacy scenarios: When RTT noise < 0.1% (e.g., USB-connected Arduino sensors), all three distributions struggle to fit (coverage < 0.7), requiring additional deterministic models.

3) Scene Adaptation Recommendations for the Three Distributions

Distribution Type	Optimal Adaptation Scene	Scenarios to Avoid	Engineering Implementation Suggestions
Exponential Distribution	Short delays between sensors (<100ms), high real-time demands (e.g., collision detection)	Heavy-tailed RTT (e.g., large data packets from cameras), low noise scenarios	Used for rapid change detection, can be combined with other distributions (e.g., use exponential for detecting changes, log-logistic for modeling stable segments)
Log-Normal Distribution	Wired LAN, medium data volumes (1KB~100KB), balanced demand scenarios	Intercity communication, heavy-tailed RTT, high noise scenarios	As a general solution, suitable for initial implementation in scenarios with unclear characteristics
Log-Logistic Distribution	Intercity/wireless communication, large data volumes (>100KB), high precision demands (e.g., navigation)	Ultra-short RTT (<10ms), high real-time demand scenarios	Used in core control loops, can reduce time through parallel computation

5. Technical Summary and Outlook

The RTT marginal modeling technology for networked robot sensor transmission addresses the pain points of traditional modeling methods, such as “poor real-time performance,” “incomplete information,” and “weak scene adaptability” through its architecture of “core distribution models + improved parameter estimation + scene-specific goodness-of-fit testing.” Each of the three distributions has its focus: exponential distribution pursues “speed,” log-logistic distribution pursues “accuracy,” and log-normal distribution pursues “balance.” When implementing, it is essential to comprehensively select based on RTT characteristics, real-time demands, and precision requirements.

Future optimization directions will focus on three major pain points:

① Enhancing adaptability to high determinacy scenarios (e.g., integrating deterministic models with marginal distributions);

② Reducing computational costs of log-logistic distribution (e.g., parallelizing parameter estimation);

③ Strengthening robustness in non-stationary scenarios (e.g., combining hidden Markov models to identify non-stationary segments).

We believe that with the implementation of these optimizations, RTT marginal modeling technology will become one of the core supporting technologies for remote control of networked robots and management of sensor transmission delays.

For more details, please refer to the original work: “Goodness of Fit in the Marginal Modeling of Round-Trip Times for Networked Robot Sensor Transmissions“

END

Click “Read the Original” for more