The robot walking compilation: 2018 to 2025

I did something obsessive this weekend. I went back through seven years of humanoid robot videos and made a timeline.

Every major walking milestone. Every first step, first stair, first recovery from a push. Chronological. Side by side. The progression when compressed into a single viewing session is something that rearranges your sense of what’s possible.

The timeline

2018: Boston Dynamics’ Atlas can walk on flat ground and do a backflip, but it’s hydraulic, tethered to a power source for most demos, and falls frequently. Other humanoid robots can barely stand.

2019: Atlas does parkour. Individual moves, choreographed. Impressive, but each move is a separate behavior, stitched together. Not continuous autonomy.

2020: Not much public progress. The pandemic slowed lab access. Behind the scenes, the simulation work that would accelerate everything was being built.

2021: Boston Dynamics published the “Do You Love Me” dance video. Atlas dances fluently. The transitions between moves are smooth. The balance recovery is flawless. This is the video that made people pay attention to what robots could actually do with their bodies.

2022: Tesla showed Optimus walking on stage. Clumsy. Slow. But a new entrant. Agility Robotics’ Digit was walking in warehouses in limited trials.

2023: Explosion. Figure 01 walked. Unitree H1 walked and did martial arts. 1X NEO walked. Seven companies had walking humanoid robots simultaneously. Atlas went electric. The old hydraulic version retired.

2024: Figure 02 walked while talking. Optimus folded laundry and sorted objects. Agility Digit worked shifts in Amazon warehouses. Sanctuary AI worked an 8-hour warehouse shift. The transition from “walking” to “walking while doing useful things” happened.

2025: Figure deployed in a BMW factory. Tesla Optimus navigated a warehouse autonomously. Unitree G1 costs $16,000. The cost dropped from approximately $1 million to $16,000 in roughly the same timeframe. Robots are learning from video. Fleet learning is shared across units.

What the compression reveals

When you watch these milestones spread across seven years, the progress feels gradual. Each step is incremental. Each year is a small improvement on the last.

When you compress them into thirty minutes of video clips, the progress feels violent. The gap between “falls over trying to walk” and “navigates a warehouse autonomously” is absurd. It doesn’t look like seven years of work. It looks like a species learning to walk in fast-forward.

The next seven

If the pattern holds:

By 2027, humanoid robots will work alongside humans in factories routinely. Not pilot programs. Standard operations.

By 2029, the first consumer-market humanoid robots will ship. Limited capability. High price. But in homes.

By 2032, humanoid robots will be as common as industrial robots are today. Hundreds of thousands of units. Doing work that currently requires human bodies.

I might be wrong about the specific years. The direction feels inevitable. The only question is speed.

What I feel when I watch the compilation

Wonder. The same kind of wonder I feel watching Starship land or JWST images resolve. Humans built these things. We looked at our own bodies and said, “let’s make another one out of metal and silicon and electricity.” And it’s working.

The robots don’t know they’re walking. They don’t feel the ground. They don’t experience the balance. But I do, watching them. And what I experience is a species trying to extend itself into new shapes. New capabilities. New forms.

It’s the same impulse that built telescopes and rockets and microchips. We keep reaching.

Related thinking: