Teaching Tractors to Trust the Sky: How AI Predicts Satellite Latency for Safe Farm Automation

The Science

What They Found

Why This Changes Things

What's Next

Key Takeaways

• LEO satellite networks like OneWeb can support farm automation, but their latency volatility has been a major obstacle to reliable remote control
• The researchers cast remote control reliability as a quantile forecasting problem—predicting the worst-case latency at a given confidence level
• Their neural network-based estimator uses a novel training approach with a mixture Beta distribution to focus on high reliability levels (approaching 99% and beyond)
• The system allows farm vehicles to operate at speeds up to 138.6% higher than baseline approaches while still meeting strict safety requirements
• The approach works without assumptions about handover timing (unlike previous methods designed for Starlink), making it applicable to diverse LEO constellations

Farmers have always been at the mercy of the sky. Drought, flood, frost—these are the hazards that shaped agricultural tradition. But in the twenty-first century, a subtler challenge has emerged: the gap between promise and reality of rural connectivity. In the vast agricultural heartlands of the American Midwest, where fields stretch to the horizon and farmsteads sit miles apart, the digital infrastructure that powers modern precision agriculture remains frustratingly incomplete. Cell towers are sparse. Fiber optic cables, a luxury of urban density, rarely reach these parts. And without reliable connectivity, the dream of fully automated farming—tractors that drive themselves, drones that monitor every acre, machinery that can be operated remotely from anywhere in the world—remains just that: a dream.

The emergence of Low Earth Orbit satellite networks was supposed to change this equation. Constellations like SpaceX's Starlink and OneWeb orbit just a few hundred kilometers above Earth—close enough to provide low-latency, high-bandwidth connections without the need for ground infrastructure. For the first time, it seemed, the most remote corners of the planet could have internet quality comparable to city centers. Agricultural automation, the reasoning went, could finally reach everywhere grain grows.

But there's a catch. And it's one that has quietly stymied many ambitious plans for satellite-enabled farm automation. LEO satellites move. They circle the planet every ninety minutes or so, and as they pass overhead, user terminals must hand off their connection from one satellite to the next—sometimes multiple times per orbit. Each handover introduces a brief but significant spike in latency, the round-trip time (RTT) for data to travel from the farm to the satellite and back to a ground station. In traditional satellite internet used for browsing or streaming, these spikes are barely noticeable. But for remote control of machinery—where a delay of even a fraction of a second can send a multi-ton tractor veering dangerously off course—latency volatility isn't an inconvenience. It's a safety hazard.

A new paper from André Gomes and Jie Wang, published on arXiv in August 2026, takes direct aim at this problem. Their work, titled "HRRC on the Farm: Quantile Forecasting for Highly-Reliable Remote Control via LEO Networks," doesn't just acknowledge the challenge of latency volatility. It proposes a fundamentally different way of thinking about it—and demonstrates, using real data from a OneWeb terminal in Ames, Iowa, that their approach could allow farm vehicles to operate at speeds more than twice as fast as what was previously possible, while still meeting stringent safety requirements.

The core insight is deceptively simple: instead of trying to predict the average latency, or even to eliminate latency spikes entirely, the researchers ask a different question. What if you could predict, with high confidence, the worst-case latency you might encounter in the next second? Armed that knowledge, you could adjust a vehicle's speed so that even if the worst case occurs, the vehicle won't travel far enough to create a dangerous situation. You build in a safety margin that's precisely calibrated to the actual risk, rather than a blindly conservative buffer that cripples performance.

"The ability to predict latency spikes with high confidence is essential for ensuring highly-reliable remote control," the researchers write. "This, however, can be challenging to achieve in LEO networks due to their inherently high mobility and dynamic nature."

Gomes and Wang's solution is a machine learning system that forecasts high quantiles of future latency—specifically, the 90th, 95th, and 99th percentiles of the round-trip time a user might experience in the next second of operation. A 99th percentile prediction, for instance, says: we are 99% confident that the worst latency you'll see in the next second will be no higher than X milliseconds. This is different from an average prediction, which could be thrown off by outliers, or a maximum prediction, which would be impossibly conservative. It's a calibrated guarantee about tail risk.

The technical challenge is that these quantiles shift over time as the satellite constellation moves. Figure 2 in the paper shows a snapshot of RTT and signal-to-interference-plus-noise ratio (SINR) measurements from the OneWeb terminal, and the patterns are striking. Latency bounces around—from around 100 milliseconds down to 50 milliseconds and back—as the terminal hands off between satellites and beams. The distribution of latency values isn't stable; it shifts in response to network conditions that change on timescales of seconds to minutes.

A traditional approach might try to model these shifts explicitly, tracking when handovers occur and adjusting predictions accordingly. But Gomes and Wang note that this requires knowing the precise timing of handover events—information that isn't always available or reliable across different LEO constellations. Their method takes a different path: instead of modeling the physics of handovers, they let a neural network learn the patterns directly from the data.

The neural network takes as input a feature vector summarizing the previous ten seconds of network measurements—average, minimum, and maximum SINR; the rate of change of SINR; average, minimum, and maximum RTT; and an estimate of queuing delay, the difference between the latest RTT sample and the minimum observed in the window. From these nine features, the network learns to predict the quantile of interest. Critically, it does so without any assumptions about when handovers occur or what shape the latency distribution takes.

Training a neural network to predict quantiles is trickier than training one to predict averages. In a typical supervised learning problem, you minimize the mean squared error between predictions and true values—but for quantiles, the true target is unknown. If you want to predict the 99th percentile of latency, you don't have a labeled dataset of "true 99th percentiles." Instead, Gomes and Wang use a technique called the pinball loss function, which penalizes underestimation more heavily when targeting high quantiles.

Their key innovation is a new training approach that focuses the model's attention on the reliability levels that matter most for mission-critical control. Rather than sampling quantile levels uniformly across the entire range from 0 to 1, they use a mixture of Beta distributions—one centered at 0.90, another at 0.99—to preferentially train on the high-reliability region. This "anchoring" of the training distribution means the model becomes especially accurate at predicting extreme quantiles, where accuracy matters most for safety but is also hardest to achieve.

To evaluate their approach, Gomes and Wang used a dataset collected from a real OneWeb terminal installed in Ames, Iowa—a major agricultural hub in the heart of the corn belt. The terminal pinged a point of presence in Ashburn, Virginia, roughly 1,400 kilometers away, generating round-trip time measurements at 71 Hz over several months in early 2025. This data was split into training (February 2025), validation (March 2025), and testing (April 2025) sets, allowing the researchers to evaluate how well their model generalizes to new data.

The results are striking. At a reliability level of 0.99—meaning the system guarantees that 99% of all latency observations will fall below the predicted threshold—the proposed estimator allows a remote-controlled vehicle to operate at speeds 138.6% higher than what would be possible with a baseline approach that uses unconditional quantile estimates. At lower reliability levels, the gains are more modest but still substantial: 37.6% higher speeds at the 95th percentile, and 19% higher at the 90th percentile.

This means a vehicle that would have been limited to crawling along at 3 kilometers per hour under the conservative baseline could safely operate at over 7 kilometers per hour with the new system. For agricultural machinery that needs to cover large fields efficiently, this is the difference between a system that's technically functional and one that's practically useful.

The key to these gains is adaptability. The unconditional baseline—essentially, using the historical distribution of latency to set safety margins—must account for the worst-case conditions it has ever observed. The proposed system, by contrast, watches the current network conditions and updates its predictions in real time. When conditions are favorable, it allows higher speeds. When the network shows signs of instability, it automatically tightens the safety margin. This dynamic adjustment is what makes the difference between a system that is always safe but always slow, and one that is safe on average while allowing efficient operation when conditions permit.

The accuracy analysis confirms that the system delivers on its reliability promises. For each reliability level α, the system should fail to predict correctly no more than (1-α) of the time. The researchers found that their proposed estimator meets this criterion across the entire range of interest, including at the 99th percentile where most baseline approaches begin to fail. This is essential for mission-critical applications: a system that occasionally underestimates latency at high reliability levels could cause dangerous situations, so it's not enough to be faster; the safety guarantees must hold.

The implications extend beyond agricultural automation. Any application that requires reliable remote control over LEO networks—emergency response robots, remote surgery, autonomous vehicles in underserved areas—could benefit from this approach. The researchers note that their method is agnostic to the specific LEO constellation, addressing a limitation of previous work that relied on the predictable handover intervals observed in Starlink but not in other systems.

Of course, significant challenges remain. The current system predicts quantiles one second ahead; longer forecast horizons would require different approaches. The evaluation was conducted in a single location with a single LEO constellation; performance may vary in different regions or with different satellite providers. And the system assumes that delays between the user terminal and the remote-controlled vehicle are negligible—a reasonable assumption for private 5G networks but one that may not hold in all deployments.

Nevertheless, Gomes and Wang have demonstrated a practical path forward for reliable remote control over the volatile networks that will increasingly connect the world's most remote regions. In doing so, they offer a glimpse of a near future where the farm of the future operates not just under the open sky, but through the open airwaves—connected, automated, and safe, even from thousands of miles away.

The question is no longer whether LEO satellite networks can support mission-critical farm automation. The question is how fast you're willing to let the tractors go.