The New Year has just begun—earlier for us, yesterday for China—and 2025 already seems like ancient history. At least when it comes to advances in humanoid robots. For the Chinese New Year show, robots were once again integrated into the dance and kung fu choreography, and not for the first time.

On display were a handful of G1 robots from Hangzhou-based robot manufacturer Unitree, which are extremely popular with researchers and showcase their movement capabilities in many impressive videos. Here is the G1 robots’ approximately four-minute appearance at the Lunar New Year Festival:

The fluid movements of the robots and their balancing abilities are truly impressive, even when the human dancers unintentionally throw them off balance. Other manufacturers are keen to keep up, such as Agibot from Shanghai.

The progress made within a year, from 2025 to 2026, cannot be overlooked. While the Unitree G1 robots at the 2025 Chinese New Year Festival still seemed rather jerky and awkward, the 2026 presentation was definitely a big step forward in terms of movement capabilities.

We see this in walking and running as well. The comparison between the Tesla Optimus and the Figure.AI robot, as well as the participants in the 2012 DARPA Robotics Challenge, demonstrates how far the industry has come over the years.

I’ll spare us any more videos, because we’re happy to believe that great progress is being made here. But is it the right kind of progress? Are we primarily interested in robots that can run, dance, and fight kung fu, which is why investors are pouring billions of euros into the development of humanoids? Of course not.

Figure.AI CEO Brett Adcock called on robot manufacturers in a podcast—which, incidentally, is well worth watching—not to just show “loops.” A robot that does a somersault, a few kung fu jumps, and similar tricks is only performing pre-programmed routines. As impressive as this looks, and as much as it advances the stability and balance of robots, it misses the actual goal by a long shot—unless, of course, the intention is to develop robots solely for such performances.

The actual development goal should be humanoid robots that can relieve humans of dangerous, monotonous, and hazardous work. Cool dance routines and Shaolin moves on stage are probably not part of that.

But there is very little information available in the form of videos. It’s not that there aren’t any, but a quick glance raises more questions than it answers. The first question is whether this is a remote-controlled or autonomous robot. By “remote-controlled,” we mean whether the robot is operated by a human being using a remote control and carries out tasks.

Not that this is a bad thing per se. Remote control is used today to train humanoid robots and also to generate training videos for machine learning systems. NEURA Robotics in Germany, for example, has set up its own gym, or training center, for its robots.

Fourier Intelligence from Shanghai also offered a glimpse behind the scenes of how the GR-3 robot is trained. Human trainers sit in front of individual robots and perform the movements required to complete a task.

For some tasks, it will continue to make sense to call on human teleoperators in the future, such as rescue missions or research missions in inaccessible parts of the world, whether underwater, in caves, or in orbit.

But that’s not what Brett Adcock meant. He’s calling on other manufacturers to go beyond simply showing a routine (“loop”) or a remote-controlled robot, and instead to publish an unedited video showing an autonomous robot performing a complex task—and currently, every household task seems to be complex for humanoid robots—from start to finish.

This does not refer to the following video, which shows the NEURA Robotics 4NE-1. Note that the interesting parts are not visible, such as cutting the cucumbers and arranging the cucumber slices on a plate.

A video from Figure.AI demonstrating the capabilities of the VLA Model Helix 02 can serve as a benchmark for all other video demonstrations of humanoids. In the video, we see unedited footage of the Figure 03 opening a dishwasher, removing the plates and cups, stacking them, and placing them on shelves and in drawers. It then takes the plates and cups out of the sink, fills the empty dishwasher, inserts a dishwasher tablet, turns it on, and closes it.

No remote control, unedited, with the entire robot moving around the kitchen and performing complex tasks such as opening, clearing out, stacking, loading, opening and closing cabinet doors and drawers, putting things away, inserting items, switching on, and closing.

As impressive as the kung fu and dance routines are, they are nothing more than window dressing. They impress laymen and make investors loosen their purse strings, but they will disappoint entrepreneurs who expect real work performance from the humanoids.

These videos prove one thing above all else: robot hardware has been mastered, to put it simply. The VLA and operating systems that will control the robot hardware are still completely open. The true capabilities of humanoid robots will lie in how they can perform tasks autonomously. The challenges here lie in the advances made by robot manufacturers in VLA models that can recognize objects, manipulate them with robot hands, and perform tasks satisfactorily.

Anyway, Happy Lunar New Year!