Sunny Ray interviews Guillaume Binet of Copper Robotics, who is on his fourth company, about whether an autonomous robot can be proven safe without determinism. Binet draws on his time at Argo AI, a self-driving company backed by Ford and Volkswagen, and at Skyway, where a failed drone demo in front of Japanese customers taught him to always bring a spare robot. He explains why formally proving a machine learning perception system safe is impossible across every possible image, so teams must restrict the operational domain, sample it densely, and assume continuity between samples. He argues that ROS is non-deterministic, which prevents replaying a crash for auditors, and describes how Copper, written in Rust and open source first, enables deterministic replay and a time-travel debugger. The conversation also covers why humanoids remain at the demo stage, early customers, defense as a possible driver of autonomy, and the future of ROS.
Why a robot you can't replay is a robot you can't prove safe, and how Copper Robotics wants to fix that.
What are you building at Copper Robotics, and what problem are you solving?
I'm building infrastructure for autonomous robots at a time when physical AI is everywhere on LinkedIn. From the self-driving space, I saw that safety was not solved then, and with the new wave of robots it still isn't solved today. Copper aims to close that gap.
Why can't you prove an autonomous system is safe without determinism?
Traditional safety proofs show every outcome is accounted for across every input. For a vision model, even one image has so many pixel combinations that there aren't enough atoms in the universe to cover them. So you restrict the operational domain, sample it densely, assume continuity between samples, and reproduce behavior at those points.
Early investors told you to just use ROS. How did you pitch new infrastructure?
ROS is amazing, and I have no beef with it. It was built for flexibility, not safety. But it's non-deterministic, so if a robot hits a tree, you can't replay what happened. Eventually an auditor asks you to prove safety, and you need replayable data and simulation. We trade some flexibility for that end goal.
Why Rust, and what would you say to C++ engineers?
At Argo our developers weren't C++ engineers, they were roboticists. Rust is memory safe and catches many mistakes very early, when they're cheap to fix. People coming from Python or simpler languages can move into Rust and be more confident that bugs won't reach the robot.
What does deterministic replay change for a team day to day?
Robot development starts to look like normal software development. You catalog logs of failures and turn them into red or green tests. You reproduce the problem, fix it, and see it go green, then check the others are still green. It's a ratchet effect, so you never go backward on a system with many edge cases.
Why did you go open source first?
Others tried to build better robot infrastructure and failed, partly because they lacked an ecosystem. I needed an ecosystem like ROS has, but also needed to make money. Academics and startups use the free core and contribute drivers back, and some later need a product. Our commercial side includes the red/green log testing and a time-travel debugger.
Why aren't humanoids everywhere yet?
They're at the demo stage, working well in controlled environments on one or two repetitive tasks where safety matters less. In streets or homes, with kids waving at them and no engineers nearby, nobody has a good answer for what happens if one falls or misrecognizes something. Proof of safety is needed before scaling.
If Copper wins, what happens to ROS in ten years?
I don't think ROS is going anywhere. It has enough momentum and is huge for experimentation and university settings. But I think productized autonomous robots will move to Copper or something close to it. I can't say everything will be Copper, though hopefully most of it will be.
Building something daring? Sunny talks to founders like this every day. Fifteen minutes to see if your story belongs on the stage.
Claim your pre-interviewSend a founder here and we will know it was you.