TL;DR

• Same endpoint, different progress: LLMs, autonomous driving, and embodied AI all pursue “completing tasks independently, without human supervision.” But LLMs are nearly there, autonomous driving is in the thick of it, and embodied AI is still early.
• The data gap is the core bottleneck: LLMs have the internet corpus, autonomous driving has driving data, but embodied AI has no “robot internet” — the wave of startups flooding into the data space right now is essentially building the industry’s infrastructure.
• Physical AI has no “undo button”: LLMs can iterate repeatedly to improve usability; autonomous driving and embodied AI are real-time and safety-critical — they must be right the first time, with reliability requirements far above LLMs. Hence they need the next technological breakthrough even more.
• Conclusion: embodied AI is making up the coursework autonomous driving already did — not overtaking it. What both need is the next generation of learning frameworks — a method that can economically push reliability from 99% to 99.9999%.

Main text

Scrolling X recently, I saw many embodied-AI dataset startups in full swing — FPV_labs, Ropedia, and so on. Generalist also released Gen-1, demonstrating a scaling law for embodied scenarios. A wave of thoughts rushed in, and one question flashed into my mind:

Will embodied AI arrive before autonomous driving?

After some CoT, my conclusion is: No. Here is the CoT:

Three Roads, One Endpoint

LLMs, autonomous driving, and embodied AI — these three tracks ultimately pursue the same goal: completing tasks independently, without real-time human supervision.

LLMs are moving fastest. In the ChatGPT era it played an assistive role — you ask, it answers, and you bear the consequences of errors. But with the Claude Code and Cursor generation, once model accuracy is high enough and closed-loop correction exists — running tests, self-correcting — its value changed qualitatively. People began using LLMs as the engine and harnesses as the scaffolding, building projects previously unimaginable. It is no longer an assistant; it is genuinely working independently. Everyone is rushing to hand money to Claude and Zhipu.

Autonomous driving is going through a similar process. L2 assisted driving is like ChatGPT: it can help you drive, but you can’t let go — you are still the backstop. L4 is its “Claude Code moment” — the car really drives itself, and you can hand over the wheel with confidence and attend to other things. Once that state is reached, people will pay the automakers too — subscribing to a “virtual driver” is buying back your own time.

Embodied AI is the same. Today’s embodied AI is roughly at the just-works-demo stage, still far from “hand it over with confidence.”

Data: Everyone Has Their Own Hard Problem

These three tracks are in very different positions on data, and that difference explains much of today’s industry landscape.

LLMs enjoyed the dividend of internet data early on — massive text was readily available. But that dividend is long spent. The quality and density of internet data are actually mediocre; the truly useful, high-quality portion of that ocean of text is extremely low — panning for gold in sand. So LLM vendors now deliberately construct high-value data and design dedicated pipelines for capability training. Whoever constructs data better gets stronger performance.

Autonomous driving is almost the opposite. Humans naturally drive, so the base fuel for imitation learning has never been lacking. Dashcam footage, fleet-collected data, production-fleet data feedback — these are the internet-scale corpus the autonomous driving industry naturally owns. Imitation learning has advanced so fast that since everyone went end-to-end, no one can open up an absolute performance gap.

And embodied AI? There is no “robot internet” to crawl, no massive archive of robot-manipulation recordings to train on. That is why we see so many startups pouring into the embodied data space — what they are doing is, in essence, building embodied AI’s “internet.” They believe in the scaling law, and believe every data investment today is laying the foundation for the embodied industry. It is exactly this faith that fuels the sector’s heat.

This also explains an interesting phenomenon: the core narrative of embodied-AI startups is the Scaling Law and data, while the core narrative of autonomous-driving startups is algorithms. Because they are stuck on different bottlenecks.

The Ceiling of Imitation Learning

Embodied AI has indeed validated the scaling law well: more data, bigger models, better performance. Everything looks great — but it’s all far too early.

In today’s autonomous driving industry, if you can’t do 20 kilometers of zero-intervention driving in urban areas, you’d be embarrassed to call yourself first-tier. Embodied AI’s current laundry-folding and coffee-pouring abilities are, at best, “automatic parking” or point-to-point navigation inside a closed campus. If you want to benchmark against autonomous driving’s current capability, embodied AI would need to work continuously for an hour in any home environment.

The scaling law for imitation learning is expensive: linearity in log space means exponential cost, and exponential cost is something nature cannot afford.

The LLM field has already proven this in practice — pure pre-training scaling is nearly saturated, and ultimately reinforcement learning is needed to raise the performance ceiling. Moreover, startups specializing in RL environments have already emerged in the LLM space, providing training environments and tools for vendors like Claude and Gemini.

For AI in the physical world, RL is much harder: the simulation environments are harder to build, and the sim2real gap must be considered. Currently perhaps only Tesla has made large-scale RL work in autonomous driving. Embodied AI is even earlier — only task-level attempts, nowhere near the large-scale parallel RL of the LLM world.

Beyond RL, Physical AI Needs the Next Revolution

There is a fundamental distinction that splits LLMs and physical-world AI (autonomous driving, embodied AI) into two camps.

LLMs can sacrifice time for quality. They can attempt a task many times, self-correcting until it’s right. Test-time scaling is exactly this idea — spend more compute at inference so the model can think, verify, and fix repeatedly. Wrong code can be re-run; wrong reasoning can be redone; the cost is just a few extra seconds or minutes. Harness engineering pushes this to the next level, studying how to construct environments where LLMs can perform, correct, and iterate better.

But autonomous driving and embodied AI cannot. They are real-time critical and safety critical. Autonomous driving errs once and the car crashes; embodied AI errs once and something shatters.

This means physical AI cannot rely on test-time scaling as a backstop the way LLMs do. It must be right the first time, or at least recognize uncertainty and act conservatively before going wrong. This places reliability demands on the model far above LLMs.

RL will improve system reliability, but is the current technology (sample efficiency) sufficient? Judging from the progress of LLMs and Tesla, I’m not so optimistic.

Perhaps we need the next technological revolution: learning algorithms with much higher sample efficiency.

Summary

Back to the original question: will embodied AI arrive before autonomous driving?

By “arrive” I mean: entering every household, usable and loved by everyone, something you can hand tasks over to with confidence — whether Tesla’s Optimus or Physical Intelligence’s dual-arm robot.

My judgment: no.

The current boom in embodied AI — waves of startups, pouring capital, data companies everywhere — looks lively, but the essence of that liveliness is an industry actively working through the foundational problems autonomous driving began facing and solving years ago. Embodied AI is making up coursework, not overtaking.

Embodied AI carries none of the autonomous driving industry’s baggage and its route is more scaling-friendly — that is an inherent advantage.

Autonomous driving and embodied AI face the same ultimate challenge — achieving extremely high reliability in a real-time-critical, safety-critical physical world. Current imitation learning certainly cannot reach it, and current reinforcement learning barely manages. What we need is the next generation of learning frameworks: a more economical approach that lets us evolve from 99% reliability to 99.999%.

Note: this text was organized with Claude’s help.