What the Robot Games are teaching us about the state of the humanoid robots

Machines are faster than humans. That’s not the story.

So machines are faster than humans. This is not news.

At the robot games, humanoids have shown they are now faster than us at the 100m, 400m, and 1,500 m … and they can jump higher than we can.

Yeah … we’ve had cars. And planes.

Here’s the real news: robot-to-robot year over year deltas are insane: the winning 100 m went from 21.50 seconds in 2025 to 9.39 in 2026, while high jump went from roughly 0.95 m to 2.88 m.

Now project that forward in time ...

The boring events matter more

Look, the more important events are the boring ones. The Games now include EV charging, cable insertion, industrial assembly, material handling, packaging, restaurant work, hotel service, logistics and emergency response.

Yes, plugging in a cable.

You barely think about it. But for humanoids, it’s a whole chain of perception → planning → locomotion → reaching → grasping → alignment → force control → verification → error recovery.

(Which of course you do too … just subconsciously because this is a solved problem for most humans.)

So here we’re seeing real gaps.

The firefighting test

Twenty-three teams entered a firefighting challenge. Robots were given 30 minutes to identify hazardous materials, locate and close three valves, find an extinguisher and put out a fire.

Of the 12 teams that competed in the reported Sunday session, only three completed the entire mission. Rain, changing light, perception, manipulation and movement all caused problems.

What does that tell us?

Single actions are easy (ish). Robots can now perform extremely difficult individual capabilities. It is much harder to string 10 relatively ordinary capabilities together and have all 10 work correctly.

Why? 

Reliability compounds negatively. If every stage of a ten-stage operation works 90% of the time, the complete workflow works only a minority of the time. And guess what … most things we want humanoid robots to do are long-horizon tasks.

“Can do it” versus “can do it autonomously”

Another example:

A hotel competition required robots to handle luggage, push a cart, unload it, restock towels, slippers and bottled water, tidy a room and deal with waste. That all means dealing with narrow corridors, deformable objects such as towels and bedding, carts that change the robot's balance, objects in slightly different positions and long sequences of actions.

Guess what: teleoperation was allowed. It cut your scores, but it was allowed.

“A robot can do it” is very different from “a robot can do it autonomously."

Doing it once isn’t enough

Sports are cool … makes for good TV. It’s relatable for the average human. And it shows physical prowess and dexterity that are going to matter at some point. Reliability too: perhaps even more critical.

Doing a hard thing once is great. But for work, you need to do it 1000X a day. Every day for weeks and months. And … on attempt 743 when the package has moved, a human walks by, the lighting changes, and the first grasp misses?

More than a sporting event

Ultimately, this is much more than a sporting event. It’s a proving ground for commercial products and China's strategic robotics industry. It’s a pride thing … look at what we can do. And it injects even more competition into the race for great humanoid robots.

And it’s an improvement loop: 

Build → compete → fail publicly → collect data → improve → compete again.

So … we’ve seen the growth from last year to this year: 2-3X improvement.

What should we expect next year?

What should we expect for next year? Harder benchmarks, harder tasks, and longer sets of tasks. Things like:

• longer multi-stage jobs

• deformable objects

• stairs and uneven surfaces

• instruction-following

• unexpected object placement

• human-robot collaboration

• failure recovery

• endurance

• intervention rates

• task completion per hour

And maybe even robots stopping on their own without running into a big cushion or falling over in a shower of sparks. That would be nice.

:-)

Two things we know

2 things we know:

1. Humanoid robotics is advancing extremely fast
2. Humanoid robots are still nowhere close to being universally reliable workers

But at their rate of improvement … we’re not far off. Months? Probably not. Years? Definitely. 

Certainly not decades.