Rhoda AI leaves stealth with a $450 million Series A to run factory robots on a model that learns from internet video

Rhoda AI, a Palo Alto start-up that teaches robots by having them predict video, came out of stealth on 10 March 2026 with a $450 million Series A. The company said the money will pay for research and engineering, for more industrial deployments and customer pilots, and for a larger team.

The round drew Capricorn Investment Group, Khosla Ventures, Leitmotif, Matter Venture Partners, Mayfield, Premji Invest, Prelude Ventures, Temasek and Xora, along with individual backers including John Doerr, according to Rhoda's release. The company named no lead investor and disclosed no valuation. It said it had spent 18 months in stealth before the launch.

What Rhoda says it has built

Rhoda calls its core model Direct Video Action, or DVA, and the product layer built on it FutureVision, which the release describes as the foundation model that powers Rhoda systems today. The model is pretrained on internet-scale video; Mayfield put the corpus at hundreds of millions of videos in a post published on launch day. In operation the model predicts how the scene in front of the robot should unfold, turns that prediction into motor commands and repeats the cycle every few hundred milliseconds in a closed loop, the company said.

A technical post Rhoda published the same day, titled 'Causal Video Models Are Data-Efficient Robot Policy Learners', fills in the design. A causal video model forecasts future frames from the video history, the robot's joint readings and a language instruction. A separate inverse dynamics model then works out which actions would produce those frames; Rhoda says that component needs about 10 hours of robot-specific data and can be trained on random motions. A scheme the company calls Leapfrog Inference computes the next prediction while the robot is still executing the current one, so the model's latency does not stall the arm.

The post reports long autonomous runs rather than benchmark scores. A decanting task learned from 11 hours of robot data ran for 1.5 hours without interruption, and a container breakdown task trained on 17 hours of data ran for 160 minutes. Post-training for a new task takes 10 to 20 hours of task data, according to the post, and the release says some tasks need as little as ten hours of teleoperation. Rhoda published no success rates and no head-to-head numbers against other robot models.

As commercial evidence, the release says that in a recent high-volume manufacturing evaluation the system completed a component-processing workflow autonomously in under two minutes per cycle, beating the customer's performance targets. The customer was not named. Rhoda's website lists automotive, manufacturing, logistics and e-commerce as target sectors, shows returns processing, bearing decanting and container breakdown, and describes a wheeled robot with custom actuators rated at 25 kg of payload and 40 kg at peak, with a brake in every actuator.

The people behind it

Chief executive Jagdeep Singh co-founded QuantumScape, the solid-state battery developer, and led it until 15 February 2024, when he handed the CEO role to Siva Sivaram and stayed on as chairman, according to QuantumScape. Mayfield says Singh has now built seven companies. Chief science officer Eric Ryan Chan is a Stanford researcher in computer vision and generative modelling who previously designed generative models at World Labs, and Stanford professor Gordon Wetzstein, head of the university's Computational Imaging Lab, is part of the team, the release says. Singh said in the release that the next era of robotics requires models that understand how the world moves, which he set against models that only capture what a scene looks like or how it is described in language.

Where it sits among rivals

The round landed in the busiest quarter yet for US robot-AI financing. Skild AI raised $1.4 billion in a SoftBank-led Series C on 14 January at a valuation above $14 billion, Mind Robotics raised $500 million the day after Rhoda's launch and Sunday raised $165 million on 12 March, each according to the companies. Physical Intelligence closed a Series B led by CapitalG in November 2025. Most of these groups build vision-language-action (VLA) models, which attach an action output to a network pretrained on images and text. Rhoda's website argues that DVA generalises where conventional VLA pipelines struggle.

Why video instead of language

The bottleneck in robot learning is data. Teleoperated demonstrations, in which a person drives the robot through a task, are slow and costly to gather. Internet video is abundant and shows hands, tools and objects behaving under gravity and contact. A model trained to predict what happens next in such footage acquires a rough sense of physics, and Rhoda's bet is that this prior cuts the robot data each new task needs to a few working days of teleoperation. The cost is compute, because generating video frames is heavier than emitting a short action sequence, which is why the company engineered its inference around latency.

Risks and open questions

Rhoda has shown long runs on a handful of tasks, not reliability statistics across a fleet. Without success rates, interventions per hour or comparisons with VLA baselines, outsiders cannot measure how much the video prior adds. The manufacturing evaluation has no named customer, and the company has not said who builds its hardware. Mayfield describes the model as labour as a service and writes that 1,000 deployed units would generate about $100 million in recurring annual revenue, which implies roughly $100,000 per robot per year, a price that has to beat the fully loaded cost of the shifts it replaces.

What to watch

Rhoda followed up on 10 September 2026 with a research post asking whether scaling web-video pretraining helps real robots do real work, so its scaling evidence is now public to test. A first named production customer, published intervention rates and the price of a deployed unit are the facts that would turn the round into a business.