The Self-Driving Industry Finally Figured Out What We Knew in 2016: Rules Matter

Learning from bad teachers

The second paper, Plan-R1 from a team that's clearly been wrestling with the same demons, goes even further in diagnosing the core problem. Their insight is blunt: expert demonstrations "lack explicit safety awareness" and risk "inheriting undesirable behaviors such as speeding from suboptimal human driving data."

Read that again. These researchers are admitting, in an academic paper, that training self-driving cars on human driving data teaches them to drive like humans, which is to say, badly. The humans speed. The humans tailgate. The humans make choices that would get points deducted on a driving test. And the neural networks dutifully learn all of it.

Plan-R1's approach is to separate behavior learning from what they call "principle alignment." First, train the model on human data to capture how people actually drive. Then, in a second stage, fine-tune it with explicit rule-based rewards for safety, comfort, and traffic compliance. It's a two-step process that acknowledges you need both the intuition of human driving patterns and the guardrails of, well, actual rules.

They also identified a technical problem I hadn't considered: when you use group-wise normalization in their optimization approach, rare safety violations get treated with similar weight as common safe behaviors. Their fix, something called Variance-Decoupled GRPO, ensures that safety-critical objectives stay dominant during training. The details are dense, but the philosophy is simple: don't let the math accidentally deprioritize "don't crash."

So what

Here's what these papers tell me about where the autonomous vehicle industry actually is, versus where the press releases claim it is.

The good news: serious researchers are moving past the "just add more data" phase and grappling with the hard engineering problems of making these systems reliably safe. Both DriveAnchor and Plan-R1 treat safety as a first-class design constraint, not an emergent property they're hoping for. That's mature thinking.

The concerning news: we're still largely in the research phase on these approaches. DriveAnchor has been validated through "real-world vehicle testing," according to arXiv, which is encouraging but vague on the details. Plan-R1 shows state-of-the-art results on the nuPlan benchmark, but benchmarks and real roads are different beasts entirely.

And the honest truth is I don't know how quickly this research translates into the vehicles companies are actually deploying. The gap between a paper and a product is measured in years, sometimes decades. I've watched promising robotics research disappear into the void plenty of times.

What I do know is that the conversation has shifted. The kids building these systems (and yes, to me they're still kids, even if they have PhDs) are asking better questions than they were in 2016. They're not just asking "can we make it work," they're asking "can we make it safe, and how do we prove it."

That's the right question. It's just taken us a while to get here.

If you want to argue about any of this, my email's on the about page. I still check it, unlike certain messaging platforms I could name.

A wave of new papers is finally tackling the problems we've been complaining about for years, from scale drift to multi-robot coordination.

Robert "Bob" Macintosh · 2 hours ago · 4 min