Robotics Is Moving Fast: IPOs, New Models, and Smarter Robots  July 9th 2026

In the Wild

What's breaking through in AI right now, from China's robot showrooms to the search charts.

• You can actually buy a home humanoid in China now. UBTech opened sales of its lifelike U1 companion robot, starting around $17,650. It's aimed at Chinese homes first, and it's one of the first humanoids you can order instead of just watch in a demo. (SCMP)
• The robot hype is getting a reality check. After a year of viral backflip clips, Chinese analysts and outlets are openly asking whether the humanoid boom turns into real sales. Handy context the next time a demo goes viral. (SCMP)
• Europe wants in on humanoids. A Paris startup called UMA, founded by a former Tesla Optimus lead, came out of stealth with a humanoid named Northstar. The robot race isn't only a US-versus-China story anymore. (Bloomberg)
• Grok 4.5 was one of the week's most-searched AI terms. Grok 4.5 shipped with Cursor as a co-developer, and interest spiked across US search. If you build with it, EU access is delayed for now. (Axios)
• OpenAI's new voice can listen while it talks. GPT-Live is rolling out to free and paid tiers with full-duplex voice, so it responds without waiting for you to finish. It's the kind of thing you notice the second you try it. (OpenAI)

Quick Hits

Humanoids Go Public

• Agility Robotics is going public via SPAC at a $2.5 billion valuation, merging with Michael Klein's Churchill Capital XI for more than $620 million in proceeds. The tell is what CEO Peggy Johnson said out loud: home robots are "10-plus years" away, so Agility is selling warehouses (GXO, Amazon, Toyota) about 1,000 Digit units on a robots-as-a-service model that has already booked over $300 million.
• Unitree cleared its Shanghai IPO, winning approval from China's securities regulator to raise about 4.2 billion yuan ($618 million) on the STAR Market at a roughly $5.9 billion valuation, with a debut possible as early as late July. It is a rare profitable name in humanoids, which is exactly why every Chinese robotics supplier is watching the pricing.
• Tesla is turning the line that built its last Model S into an Optimus factory. Musk says output will start "quite slow" because Optimus has roughly 10,000 all-new parts with no supply chain yet. The V3 reveal keeps slipping, but the factory conversion is real.

The Robot Brain Race

• Mistral shipped Robostral Navigate, an 8-billion-parameter model that lets a robot find its way with a single RGB camera, no lidar and no depth sensors. It hits 76.6% success on the R2R-CE navigation benchmark, beating the best depth-and-multi-camera systems by 4.5 points, and it runs on wheeled, legged, and flying robots. Europe just put a serious robot brain on the board.
• InternVLA-A1.5 posted the best results on all six major robot simulation benchmarks, folding language understanding and a "latent foresight" step into one policy trained on 1.2 million robot episodes. The benchmark win matters less than the cadence behind it: a new vision-language-action model clears the bar almost every week now.

Capital and Control

• Nvidia is building Halos, a safety stack meant to make a 200-pound humanoid safe enough to work next to people, and Neura and others are racing at the same problem. This is the unglamorous layer nobody demos, and it's the one standing between the videos and a robot in a real workplace. (WSJ)
• Chinese lidar maker Hesai is deepening its US business through a bigger Nvidia partnership even though it sits on the Pentagon's blacklist. Lidar is becoming the robotics version of the chip fight: one company, one component, two governments that both want control of it.
• GM- and Tencent-backed Momenta had a flat debut in Hong Kong after a $751 million IPO, landing at a $9 billion valuation to a muted reception. It's a small cold-water moment: even a well-backed autonomy name can't assume investors will pay up right now.

The IPO Is Easier Than the Home

Listen to what the founders say while the bankers cheer. Agility's CEO priced a $2.5 billion public listing this week and, in the same breath, put a robot in your house "10-plus years" out, because a warehouse has fixed aisles and a home has dogs, toddlers, and a bag left on the stairs. Tesla flipped a car line to Optimus and warned that production will crawl, with 10,000 parts that don't have a supply chain yet. Nvidia is still building the safety layer that would let any of these machines stand near a person without a cage around them.

So the money is running ahead of the machines. Public markets are pricing humanoids as if the hard problems are solved, and the people building them keep saying the opposite: the near-term business is factories and warehouses, sold by the hour, to customers who can fence the robot off. The home, the version in the keynote video, is still a decade of hands, safety cases, and world knowledge away. That gap between the valuation and the deployment is the real story of the week, and it's worth remembering the next time a demo reel goes viral.

Key Takeaways

• Three humanoid companies moved toward the public markets in one week, and the CEOs building them still put home robots 10-plus years out. The capital is ahead of the capability.
• The fastest progress in robotics right now is in the brain, not the body. Mistral's model navigates on one cheap camera, and new vision-language-action models clear the benchmarks almost weekly.
• China owns the cheap hardware and the supply chain. The fight the US and Europe can still win is over lidar, safety certification, and who deploys in factories first.
• The hard part is no longer walking. New research shows that training a model to control a robot can make it lose the everyday world knowledge it started with.

Worth Reading

• LingBot-VLA 2.0 scales robot pretraining to 60,000 hours, including 50,000 hours of real robot trajectories. (Hugging Face)
• GigaWorld-1 introduces WMBench, a benchmark for evaluating robot policies in learned world models. (arXiv)
• Can Europe compete with China and the US in humanoid robotics? A clear-eyed look at where Europe actually stands in the physical-AI race. (SCMP)
• Paradigm closed a $1.2 billion fund and is pushing from crypto into AI and robotics. A useful signal of where the next wave of money is looking. (Bloomberg)

Wait, What?

• Teaching a model to move a robot can make it dumber about the world. A new study, Does VLA Even Know the Basics?, tested 7 leading vision-language-action models and found they lose commonsense and world knowledge after robot fine-tuning, especially on the richer, more semantic questions their source models could answer. We're bolting bodies onto these models faster than we're checking what the bodies cost the mind. (arXiv)

This week's poll

We're thinking about a hands-on, gamified program that trains you to prompt, code, and leverage AI tools to optimize your workflow like the top 1% do. Would you join, and what would a full program be worth to you?

Back Friday with the regular mix.

Alexis