This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on robotics and AI, follow Robert Hart. The Stepback arrives in our subscribers’ inboxes at 8AM ET. Opt in for The Stepback here.

We’re taking a winter break! The Stepback will be back on January 11th, 2026. Meanwhile, you catch up on past issues here.

I have a soft spot for robot fail videos. I watch them on a loop, chuckling to myself, as a kind of therapy. Maybe I’m a sadist, maybe I need to get out more — you can judge me later — but they get me every time. So naturally, I’ve been glued to a clip of Tesla’s Optimus robot falling like a felled tree at the company’s Autonomy Visualized event in Miami that’s been doing the rounds on social media this week.

According to the footage, Elon Musk’s vaunted humanoid robot was handing out water bottles from behind a table before knocking a bunch over, flailing its arms upward, and collapsing backwards like a puppet whose strings had been severed. Watch closely and you’ll see two things: a small plume of water as its arm crushes a bottle mid-fall (it made me laugh), and a motion uncannily similar to someone removing a VR headset.

It wouldn’t be the first time Tesla has faked the autonomous part of its autonomous robot, which Musk says is central to the company’s future. An early demonstration was just a dancer in a skintight bodysuit to show what the Tesla Bot, now Optimus, could be. Later demonstrations were revealed to be (rather obviously, by the sound of it) humans in disguise, operating the robots remotely with something like a VR headset, which we know Tesla uses in development.

Humans have been obsessed with robots for centuries; from ancient tales of stone golems and artificial automatons to modern science fiction, driverless cars, and Roombas, the idea of animating the inanimate with something resembling life has fascinated us. Much of the current hype for humanoids can be traced directly to Musk, so it is reasonable to be more than a little bit skeptical when he and others promise they’ll revolutionize the world. Musk, who has vowed to build a “robot army” of 1 million humanoids, has a deserved reputation as an outlandish and unreliable forecaster, and robotics has ridden more than a few hype waves in its history. In the past, technology has always lagged the enthusiasm of those eager to bring AI into the real world, but today, we’re being told that tech is finally ready to deliver.

So what does ready to deliver actually look like in 2025?

There’s definitely a gold rush of sorts for humanoid robots right now. Every major tech company has them on their roadmap in some capacity, and the likes of Nvidia, Meta, SoftBank, Google, Amazon, Microsoft, Intel, and Tesla (duh) are all throwing serious weight — and cash — behind them as the next big frontier in tech. And they’re not alone: A growing constellation of challengers want in on the action, including Apptronik, Boston Dynamics, Figure AI, and 1X.

China wants in, too. Beijing has decided embodied AI — which also includes non-humanoids like drones, quadrupeds, and other autonomous machines — is the key to future economic growth. In its trademark fashion, it has moved to position itself as a world leader in robotics through massive investment, government directives, and state subsidies. Everyone from tech giants Ant Group and Baidu to startups like Unitree and AgiBot is piling in.

If you go by the demos China and everyone else is pumping out, you’d be forgiven for thinking the humanoid robot future has already arrived. This summer, robots competed in dance, combat, track, and field events at the first World Humanoid Robot Games in China. A similar event was held in Greece — the International Humanoid Olympiad — at the birthplace of the Olympic Games. Humanoid fights are more popular than I could have imagined and are apparently taking place absolutely everywhere, from organized contests and underground fight clubs to sparring with CEOs.

Companies are also keen to get bots out of factories and into the home. These kinds of human-centric spaces are why advocates say humanoids are worth pursuing instead of other robot body types that could be easier to produce. Figure said its new Figure 03 robot can take on domestic chores, releasing a video of the bot doing dishes and folding clothes. 1X debuted Neo, which it claims is safe and “the world’s first consumer-ready humanoid robot.” There are also videos of Neo shakily completing basic household tasks. I’m not sure I can afford it, but if you’re interested, 1X is selling units for $20,000, with delivery starting in the US next year.

But actual use cases remain slim, and impressive demos are not the same as working products. For all the progress — and there have been impressive advances — demos are staged, scripted, or even teleoperated. For example, while it is technically true to say that Ant Group’s R1 was cooking for audiences at a trade show this year, it moved at a pace so glacial it would’ve thrilled Meryl Streep’s Miranda Priestly. Similarly, the idea of having 1X’s Neo sounds much less appealing once you realize you must also commit to having random humans telecommute into your home to control it remotely. And likewise, robot sport events are so fun to watch because the contestants are wobbly, uncertain, and unpredictable, not because they are proficient athletes.

All of this raises the obvious question: If the tech clearly doesn’t match the hype, why are so many people getting so excited about humanoids? Well, for the first time in a long time, it’s not that crazy to think the tech might be starting to catch up.

Historically, it has been exceptionally difficult to get robots to perform even mundane tasks that humans do without thinking, like walking or picking up a glass of water. Hardware — but mostly software — was a limiting factor, and machines were confined to highly structured environments and very narrow tasks. Advances in AI are changing that and turbocharging robot development.

The extraordinary breakthroughs in AI image and text generation from companies like OpenAI and Google are largely down to large language models, or LLMs, that power chatbots. These work by using complex networks to recognize and reproduce patterns found in the massive amounts of data they are trained on. Bigger AI models are trained on practically everything that can be scraped from the internet, and the resulting systems are capable of acting in a generalized way, rather than following rigid, hand-coded rules. To function in the real world, robots need this kind of generalization, and roboticists are trying to use the same approach to give machines the flexible understanding of the physical world they’d need to survive in it.

The challenge here is data. The kind of material used to train LLMs was abundantly available online, but the data needed to train robots is not nearly as easy to come by. Robots need detailed examples of things moving in the real world, which doesn’t really exist at scale. Companies are working hard to change that, and there is a huge effort to create the kind of data needed to train robot models at scale. On the surface, this looks ridiculous — Tesla has workers don cameras and sensors to teach Optimus to behave like a human — but it’s working. It’s also why companies like 1X are deploying less-than-autonomous bots into peoples’ homes, as doing so gives them the opportunity to gather the data needed to make autonomous bots by conducting complex tasks remotely, such as loading the dishwasher or cleaning a tricky object.

Humanoids are becoming more affordable, too — particularly in China — as hardware costs sink and economies of scale kick in. While prices vary wildly — entry-level units can be as little as $1,400 for Chinese model Bumi to around $13,500–$20,000 for models from Unitree or 1X, while industrial models can cost more than houses — they’re approaching prices affordable by some consumers, meaning more bots are being deployed in the real world. This creates a feedback loop, as companies get more data to work with, they build better models and better robots that more people would want, presuming they’re willing to welcome them into their homes.

But even with all this progress, it may still be hype. In November, China’s leading economic planning agency warned a humanoid robot bubble could be brewing, contrasting the sheer number of companies and scale of investment with the absence of viable use cases. The fact is, they’re not really autonomous yet, so, other than hobbyists and researchers, who in their right mind would buy one? If I wanted to get someone in to clean my home, I could hire a cleaner without the hefty upfront cost and save my money until someone produces a robot that can actually do the task I bought it for.

Until companies stop hiding behind glitzy promo videos and teleoperating purportedly autonomous robots, it’s going to be hard to tell where we are. Maybe the robots are finally coming, or maybe I have many more fail videos coming my way. Time will tell, I guess, but I’ll get the popcorn ready.