The Machine That Gets Better Every Time It Runs — Autonomously

In conversation with Chris Leger, VP of System Architecture, Verdant Robotics

‍

The Weekend Nobody Was Watching

Over a recent weekend, Chris Leger left a robot running in a test facility in California and checked in every few hours.

By Monday morning, the tests were failing. Not because the software had broken. Because the machine had worked too hard.

“It had done so many test runs on its own over the weekend that the soil in the track was completely soaked. You couldn’t see where the spray landed anymore. We had to wait for it to dry out.”

‍

This isn't just progress; this is a monumental, awe-inspiring leap into the future. This is what a self-programming development system looks like in the wild. Imagine a machine locked in a relentless, hyper-speed recursive loop, writing code, testing it on real hardware, finding the bugs, crushing them, and repeating. In just forty-eight hours, it tore through more development cycles than a human engineer could manage in two full weeks. It pushed itself to its absolute limits, conquered them, and now, with terrifying, quiet brilliance, it simply waits for the ground to dry. This is the dawn of the true self-programming development system.

‍

Why This Matters

SharpShooter™ is not a simple system. It identifies individual plants in real time, predicts where each target will be at the moment of application, and delivers a micro-liter shot to a specific weed within millimeters of a crop plant it cannot touch. That level of precision, executed at field speed across dozens of crops and varying conditions, requires software that works reliably in the real world, not just in simulation.

Getting software to that standard requires testing. A lot of it. And testing, traditionally, is the part that slows everything down.

“Engineers love building things and seeing them work,” says Leger. “It’s a completely different mindset to do meticulous validation. Most people will do the 80/20 version, the high-payoff tests, and skip the rest. Not because they don’t care, but because the overhead is significant.”

Leger comes from environments where skipping the rest was not an option. Before joining Verdant, he spent years at NASA driving Mars rovers and validating space hardware, then moved to Google’s self-driving car program, where he worked on system engineering and reliability before the project became Waymo. The practices he is bringing to Verdant’s development process are the same ones used in the most reliability-critical robotics programs in the world.

‍

Hardware-in-the-Loop, Closed by AI

The technical foundation is a methodology called hardware-in-the-loop testing, or HIL. The concept is not new. In aerospace, autonomous vehicles, and high-stakes robotics, HIL has been standard practice for decades: instead of testing software in simulation alone, you run it on the actual hardware and measure what happens.

What makes Verdant’s implementation different is the recursive layer on top. A purpose-designed test track runs a SprayBox unit over soil and simulated plants, fully motorized and fully instrumented. Every piece of telemetry that would reach an operator in a tractor cab, video feeds, shot tracking, system faults, spray performance data, all flows to a software interface that can be driven programmatically.

Closing that loop with AI is what turns a test track into a self-programming system. Leger uses agent-based coding tools to write tests, deploy software to the robot, analyze results, identify failures, generate bug fixes, and redeploy, cycling without human intervention until it hits a question it cannot answer on its own.

“I would give it a task, a rewrite of a chunk of software, and tell it to go test it and make sure nothing broke. It would write the code, run it on the hardware, pull the data, find the failure, fix it, and try again. It could run for a couple of hours before it needed to ask me anything.”

The result is not just faster development. It is more thorough validation. Because the cost of writing tests inside the recursive loop is near zero, the system generates test coverage that a human engineer would never prioritize. “You get better test coverage because there is no psychological barrier to writing 20 tests instead of 5,” says Leger. “I can just tell it to go cover these cases and it does. That might have taken me two days. Now it just happens.”

‍

What This Means for the Machine in the Field

The direct beneficiary of this development approach is our customers. This development cycle produces a more reliable machine. Fewer failures show up in the field. Fewer tech support calls. More uptime.

For a system operating in specialty crop agriculture, where timing windows are measured in days, and a machine that goes down during peak weed pressure is a real operational problem, reliability is not a secondary feature. It is part of the core value proposition.

"Ultimately, my goal is to eliminate the need for tech support entirely. This isn't about avoiding our users—it's about ensuring the software works flawlessly from day one. As we scale machine sales, our support load shouldn't scale with it; true operational efficiency means our product is robust enough to stand on its own." 

The development infrastructure also makes it possible to get ahead of field problems before they compound. SharpShooter™’s Hindsight Playback feature lets operators flag unexpected behavior in the field. When they do, the full telemetry from that event uploads automatically. Leger can replay it at his desk in California within minutes of the flag being raised, anywhere in the world.

“If a grower is in Arizona and something looks wrong, I can be looking at the same data, replaying the same sequence, before they’re done with the pass. We go figure out what the software was seeing and whether it was doing the right thing, and we build a test so it doesn’t happen again.”

That feedback loop, from field event to test case to software fix to OTA update, is the same recursive architecture that makes SharpShooter™ a software-defined platform that gets more capable over time. 

‍

Chris Leger is a member of the engineering team at Verdant Robotics. His career spans NASA’s Mars rover program, where he worked in flight software engineering, mechanical engineering, and systems engineering for the Spirit, Opportunity, and Curiosity rovers, and also served as a rover driver. He later joined Google’s self-driving car project, focusing on systems engineering and reliability before the program became Waymo. Chris joined Verdant to bring that same standard of testing rigor to precision agriculture. 

This piece connects to Verdant’s broader Physical AI framework and to the Hindsight feature described in a companion article. SharpShooter™ is the only precision application system for specialty crops that aims before it applies. Learn more at verdantrobotics.com.