Rhoda AI Raises $450M to Train Robots From Web Video

The hardest problem in robotics was never the motors. It was teaching a machine what is about to happen next in a messy, unscripted world. Rhoda AI just walked out of 18 months of stealth claiming it solved the data half of that problem by watching the internet, and it raised $450 million to prove the claim at scale.

What Actually Happened

Rhoda AI exited stealth with a $450 million Series A that values the company at $1.7 billion, an unusually large first public round even by 2026 robotics standards. The investor list reads like a who is who of deep-tech capital: Capricorn Investment Group, Khosla Ventures, Leitmotif, Matter Venture Partners, Mayfield, Premji Invest, Temasek, Prelude Ventures, and Xora, with Silicon Valley figure John Doerr among the named individual backers. The company spent roughly 18 months building before saying a word publicly, then surfaced with both a funded balance sheet and a product thesis already intact.

That product is FutureVision, which Rhoda describes as a foundation model for robotic intelligence built on video-predictive control. The model studies hundreds of millions of internet videos to learn how objects move and how the physical world behaves, then uses that learned intuition to constantly anticipate what is about to happen and translate those predictions into physical movement, a perception-to-action cycle it runs dozens of times per second. Rhoda frames FutureVision as an intelligence layer that powers its own systems today and that it intends to license to partners running different robotic hardware over time. The $450 million is earmarked for continued research, expansion of industrial pilots, and growth of a team spanning generative AI, computer vision, and robotics.

Why This Matters More Than People Think

Robotics has been stuck on a data wall. Language models had the entire text internet to learn from. Image models had billions of captioned photos. Robots had almost nothing comparable, because manipulation data must usually be collected one teleoperated demonstration at a time, which is slow and brutally expensive. Rhoda’s wager is that the missing corpus was hiding in plain sight all along: video. Every cooking clip, factory walkthrough, and unboxing on the open web is a lesson in physics and cause and effect. If FutureVision can extract real manipulation skill from passive video the way a language model extracts reasoning from text, the data wall does not just shrink, it collapses.

The valuation tells you investors take that collapse seriously. A $1.7 billion price tag on a company that just revealed itself is a statement that video-predictive control is a category, not a feature. The deeper implication is about who controls robotic intelligence. If the winning approach is training on freely available internet video rather than painstakingly captured robot demonstrations, then the advantage shifts toward whoever has the best models and the most compute, not whoever owns the most robots. That reframes the entire physical-AI race around software, which is exactly the terrain venture investors understand and want to fund.

The Competitive Landscape

Rhoda lands in a fight that is getting expensive on every side. Physical Intelligence and Skild AI are building generalist robot brains. Figure AI raised another $1 billion this year and runs multi-day humanoid demos. Mind Robotics, founded by Rivian’s RJ Scaringe, raised past $1 billion to train on a captive automaker’s factory data. Nvidia underpins much of the sector with its GR00T world-action models. Tesla feeds Optimus from its own lines. Each of these players represents a different answer to the same question: where does the training data come from?

Rhoda’s answer is the most software-native of the group, and that is both its edge and its exposure. Mind Robotics owns a proprietary factory data engine that Rhoda cannot replicate. Tesla owns a fleet. Figure owns deployed humanoids generating real interaction logs. Rhoda’s counter is that internet video is effectively infinite and free, so it can scale data faster and cheaper than any competitor collecting proprietary demonstrations one robot at a time. The strategic tell is the licensing model. By positioning FutureVision as an intelligence layer sold to other hardware makers, Rhoda is trying to become the model layer of robotics, the Android of physical AI, rather than competing to sell the best individual robot.

Hidden Insight: Passive Video Is a Shortcut That May Hit a Wall

The seductive part of Rhoda’s thesis is that it turns a scarcity problem into an abundance problem overnight. But there is a deep technical gap inside the elegance, and it is the gap that will decide whether this is a $1.7 billion company or a $17 billion one. Internet video shows what happened. It does not contain the thing a robot most needs: the action labels, the forces, the motor commands, the proprioceptive feedback of doing the task. A model can watch ten thousand videos of someone pouring coffee and learn what pouring looks like, while still not knowing how hard to grip the handle or how to recover when the cup is heavier than expected. Bridging from passive observation to active control is the entire unsolved research problem, and Rhoda is betting $450 million that video-predictive control bridges it.

There is real intellectual lineage behind the bet. World models, the idea that an agent learns by predicting the next frame of reality, have been gaining ground precisely because prediction forces a system to internalize physics. If you can accurately predict what happens next, you have implicitly learned the dynamics that govern the scene, and dynamics are what control needs. This is why the approach is credible rather than hype. The hard question is the last mile: does a model trained mostly to predict pixels transfer cleanly to issuing torque commands on a real arm, or does it need so much real-world fine-tuning that the internet-video advantage mostly evaporates at deployment?

The bear case, however, is the sim-to-real gap wearing a new outfit. Critics argue that robotics has repeatedly been promised a silver-bullet data source, from simulation to teleoperation to video, and each time the messy reality of contact physics, sensor noise, and edge cases has humbled the demo. The risk the market may be underpricing is that FutureVision produces breathtaking lab videos and stalls on the unglamorous reliability required for a customer to trust it on a real line for eight hours straight. An 18-month stealth period and a marquee investor list are evidence of conviction, not of solved physics, and a $1.7 billion entry valuation leaves little room for the multi-year grind that physical reliability usually demands.

Yet the upside is why Doerr and Khosla wrote the checks. If video-predictive control even partially works, Rhoda owns a data-acquisition cost structure no demonstration-based competitor can match, and it can sell that advantage as a licensable layer across the whole industry. The company that makes robot intelligence a software API, decoupled from any single hardware maker, captures value the way Android captured mobile: not by owning the device, but by owning the layer every device depends on.

What to Watch Next

For the next 30 to 90 days, the signal that matters is specificity. Rhoda has shown a thesis and a funding round, not a benchmark. Watch for any published task-success rate on real hardware, any named industrial pilot customer, and crucially any evidence that FutureVision controls a physical robot rather than merely predicting video. The gap between a compelling world model and a reliable controller is where most robotics narratives quietly break, and the first hard number Rhoda releases will reveal which side of that gap it is on.

Over the next 180 days and into 2027, the leading indicator is the licensing strategy. If Rhoda announces a hardware partner shipping FutureVision inside someone else’s robot, the intelligence-layer thesis is real and the company is on the path its valuation implies. If instead it ends up building and selling its own robots to prove the model works, it has quietly become a vertically integrated robot company competing head-on with Figure and Mind Robotics, a harder and more capital-hungry game than the one it pitched. Track which of those two stories Rhoda is living by year end, because they value very differently.

Rhoda is betting the entire robotics data wall was an illusion, and that the training set for physical intelligence was sitting on the open web the whole time.

---