The robots are walking. Like, actually walking.

I’m going to list something that would have been unthinkable three years ago:

Boston Dynamics Atlas: parkour, backflips, dancing. The gold standard.

Figure AI Figure 01: walking, coffee-making, learning from observation.

Tesla Optimus: walking, sorting objects, doing yoga (yes, yoga).

Agility Robotics Digit: walking, carrying bins, deployed in warehouses.

Unitree H1: walking, running, backflips.

1X Technologies NEO: bipedal walking, designed for home environments.

Apptronik Apollo: walking, designed for logistics.

Seven companies. Seven humanoid robots that can walk reliably on flat surfaces. Several can handle stairs. One can do parkour. One is deployed in a real warehouse doing real work.

Three years ago, none of these companies (except Boston Dynamics) had a walking humanoid prototype. Three years.

The curve

The progress curve for humanoid robots looks like the progress curve for large language models, shifted about 3-4 years back.

LLMs in 2020: GPT-3 existed. It was impressive to researchers but unusable by normal people. The interface was an API. The capabilities were uneven. Nobody’s grandmother was using it.

Humanoid robots in 2023: several exist. They’re impressive to robotics researchers but not deployed at scale. The capabilities are uneven. Nobody’s warehouse is full of them (Digit is the closest exception).

LLMs in 2023: ChatGPT exists. 100 million people use it. It writes, codes, reasons, and sees. It’s embedded in products. It’s changing industries.

Humanoid robots in 2026? If the curve shape holds?

I’m speculating. The hardware challenges for robots are different from the software challenges for LLMs. You can scale a language model by buying more GPUs. You can’t scale a robot by buying more motors. Manufacturing, materials science, mechanical reliability, these constraints don’t yield to the same scaling laws that silicon does.

But the software side of robotics is benefiting from the same AI advances that powered LLMs. Vision models for perception. Language models for instruction understanding. Imitation learning for task acquisition. The intelligence layer is riding the same exponential curve that language models ride.

If the hardware can keep up with the software (big if), the humanoid robot of 2026 might be as different from 2023 as ChatGPT is from GPT-3.

What I notice

Seven companies. That’s what I keep coming back to. Not one. Not two, with everyone else watching from the sidelines. Seven, all building, all funded, all demonstrating results.

When multiple well-funded organizations converge on the same goal at the same time, it usually means the goal is achievable. The market doesn’t coordinate this way around fantasy. It coordinates around opportunity.

The robots are walking. Actually walking. That sentence is new. It wasn’t true three years ago. In three more years, the sentence might be “the robots are working.” And a few years after that, it might be “the robots are everywhere.”

The curve is steep. The direction is clear. The timeline is the only question, and I’ve learned not to bet on timelines being slow when the curve has this shape.

Related thinking: