Tesla Optimus can now walk through a warehouse

I remember when Tesla showed Optimus on stage for the first time. September 2022. A person in a robot suit danced, and then the actual robot shuffled across the stage like it was afraid of the floor. The internet was merciless. “That’s it?” “Two years of hype for this?”

Two and a half years later, Optimus is walking through a warehouse. Alone. No teleoperator. No tether. No person in a suit.

It picks up boxes. It places them on shelves. It navigates around obstacles it’s never seen before. It does this at walking speed, which is slow compared to the dedicated warehouse robots that zoom on rails at Amazon. But the warehouse robots can only do one thing. Optimus can do many things. Badly. For now.

The curve

I’ve been tracking humanoid robot progress the way some people track stock prices. And the curve looks familiar.

In 2022, the best humanoid robots could walk on flat ground. In 2023, they could handle stairs and uneven terrain. In 2024, they could manipulate objects while walking. In 2025, they’re navigating unstructured environments autonomously.

That’s the same shape as the LLM curve, shifted back about three years. In 2020, language models could write coherent paragraphs. In 2021, they could follow instructions. In 2022, they could reason about code. In 2023, they could think step by step.

If the pattern holds (and I’m not certain it will, because physical hardware has constraints that software doesn’t), then by 2027 or 2028, humanoid robots will be doing useful work in enough settings to matter economically.

What’s different about Tesla’s approach

Most robotics companies build the robot and then figure out the AI. Tesla built the AI (for self-driving) and is figuring out the robot. The neural networks that process camera feeds and make driving decisions in Tesla cars are being adapted for Optimus. Same architecture, different body.

There’s something elegant about that. A single AI approach that works in cars and robots and potentially anything with sensors and actuators. The Tesla Bot isn’t a separate project from Tesla FSD. It’s the same project wearing a different body.

Whether this approach works better than purpose-built robotics AI is still an open question. But the development speed suggests something is right about it. The gap between “can barely walk” and “navigating a warehouse” closed in 30 months.

The speed of improvement

That’s the thing I keep coming back to. Not the current capability, which is modest. The speed at which the capability is improving. 30 months from shuffling to autonomous warehouse navigation. What’s 30 months from now?

I don’t know. But I’m taking notes.

Related thinking: