From Highways to Haul Roads: How Motion Planning Transfers Across Autonomous Domains

By Adityaveer Raswan

I've spent the last seven years building motion planning systems for machines that drive themselves. First at Waymo, working on the robotaxi fleet navigating San Francisco's streets. Then at Waabi, adapting those ideas for 80,000-pound autonomous trucks. Now at Agtonomy, I lead the motion planning team building autonomy for agricultural equipment: tractors, mowers, and sprayers operating in vineyards and orchards.

Each time I switched domains, I carried the same question: how much of what I know still applies? The answer has consistently surprised me, and it has direct implications for autonomous mining.

The Planning Pipeline Is Domain-Agnostic (Until It Isn't)

Every autonomous vehicle runs the same fundamental pipeline: perceive the environment, predict what other agents will do, plan a safe trajectory, then execute it through vehicle controls.

This architecture reflects the physics of the problem. Any machine moving through space needs to know where it is, what's around it, where things are headed, and what it should do next. That sequence holds whether you're planning at 65 mph on a highway or at 5 mph through an almond orchard. But the similarities at the pipeline level mask real differences in how each stage gets implemented.

What Changes When You Leave the Road

The most counterintuitive lesson from my career transitions is this: off-road autonomy is simultaneously harder and easier than on-road, just in completely different ways.

Localization gets harder. Urban robotaxis rely on centimeter-accurate HD maps and reliable GPS. The world is mapped in advance: lane boundaries, traffic signals, curb geometry. Move to a farm field or a mine site, and those assumptions fall apart. GPS can be degraded or denied in canyons, under tree canopy, or near large metallic structures. HD maps don't exist for most off-road environments, and the terrain itself changes. A freshly graded haul road looks different every shift.

Off-road systems work with simpler representations: aerial imagery, topographic data, sparse waypoint routes. The localization problem shifts from "match what I see to a perfect prior map" to "build understanding on the fly from limited references."

Interaction complexity drops dramatically. This is the insight that surprised me the most. On San Francisco streets, a robotaxi's motion planner has to reason about dozens of agents simultaneously: pedestrians stepping off curbs, cyclists filtering through traffic, ride-share drivers making sudden U-turns. Every agent has different goals, different dynamics, different levels of rule compliance. This is why on-road autonomy has been pushed toward increasingly ML-heavy planning architectures. You need learned models to capture the sheer unpredictability of human behavior in dense urban settings.

Off-road environments are a different world. On a mine site, the agents are known: other haul trucks, water carts, light vehicles, maybe a few pieces of ancillary equipment. They operate under dispatch systems and strict traffic management plans. In agriculture, the situation is simpler still. Your tractor might share a field with one other machine, or nothing at all.

This reduced interaction complexity has a real architectural consequence: classical motion planning works exceptionally well off-road.

Why This Matters for Mining

The on-road AV industry has spent fifteen years and tens of billions of dollars building increasingly sophisticated planning systems, with much of the recent investment going into ML-based approaches: joint prediction and planning models, end-to-end learned systems, foundation models for driving. These tools were built to handle unstructured interaction with unpredictable human agents in dense urban environments.

Mining doesn't have that problem. Mining has challenges around localization in GPS-constrained environments, planning on deformable terrain, managing vehicle dynamics for machines that weigh hundreds of tons, and operating safely around the clock. These are hard problems, but different hard problems, and many are well served by classical approaches: search-based planners, optimization-based trajectory generators, deterministic state machines for behavioral decisions.

A well-tuned classical planning stack can get you to 99% or even 99.9% operational coverage in mining. The interaction problem is structured enough that you don't need a learned model for the common case. Where ML becomes valuable in planning is pushing beyond that last fraction: the rare edge case, the unusual obstacle, the degraded conditions that deterministic rules struggle with. That is a very different role for ML than on-road, where learned models are increasingly responsible for the entire planning decision.

ML is not optional in off-road autonomy. Perception still depends heavily on learned models, regardless of domain. A haul truck needs neural networks to tell the difference between a boulder, a person, and a piece of equipment. The distinction is that once perception has done its job, the downstream planning can lean on classical methods, with ML reserved for edge cases where deterministic approaches fall short.

This split—ML‑heavy perception with predominantly classical planning—fits naturally with an open, modular architecture. With standardized interfaces (for example, ISO 23725 between fleet management and autonomous systems), mines can swap or upgrade perception, prediction, or planning modules as needs evolve—without rewriting the entire autonomy stack. It’s a practical way to bring in cross‑domain innovations while keeping deployment and certification manageable.

It also shapes strategy. Knowing where ML is essential and where classical methods carry the load informs mines how to partition its autonomy stack, what interoperability between vendors can actually look like, and how safety cases get built and certified. These are topics I plan to explore in more detail in upcoming pieces.

Building Bridges

The autonomous vehicle industry has historically operated in silos. But the planning pipeline refined for urban driving, the safety framework developed for highway trucking, and the terrain-adaptive planner built for agricultural fields are not as far apart as they seem. The core algorithmic foundations are shared, the implementation details diverge based on domain constraints, and lessons learned in one domain regularly solve problems that another hasn't encountered yet.

Mining autonomy is at an inflection point. The architectural decisions being made today will shape the industry for years. Those decisions will be better informed if the mining community engages with the broader AV ecosystem: not to copy solutions, but to adapt proven approaches with domain-specific intelligence.

The road from highways to haul roads is shorter than most people think.

---

Adityaveer Raswan is a Staff Software Engineer specializing in motion planning for autonomous systems, currently leading the motion planning team at Agtonomy. He is a contributing partner at OpenAutonomy.com.