Rhoda AI, the Palo Alto robotics startup led by co-founder and CEO Jagdeep Singh, is betting that the fastest way to teach robots about the physical world is to have them watch the internet. In a November 13 interview with The Robot Report, Singh explained how the company's direct video action model, or DVA, predicts what a scene will look like a fraction of a second ahead and then turns that prediction into robot movements. It is one of the clearer explanations yet of a training approach that a lot of physical AI teams are watching.
The company has the funding to test the idea at scale. SiliconANGLE reported in March that Rhoda raised a $450 million Series A led by Premji Invest at a $1.7 billion valuation when it came out of stealth. Here is how the model works, what Rhoda shipped this month, and what it means for robotics founders and engineers around the South Bay and Peninsula.
How Rhoda AI's direct video action model works
Many physical AI developers build vision-language-action models, which map camera images and text instructions to robot actions. Rhoda takes a different route. Singh told The Robot Report that the DVA model looks at a history of video frames, action tokens and language tokens, then predicts in video what should happen over the next half second or so. A second part of the model, an action converter, takes that predicted scene and outputs robot actions.
The robot executes those actions for about half a second, takes a new observation and makes a new prediction. That loop of observe, predict, convert and act repeats continuously, so the robot keeps correcting as the world changes around it. For engineers used to separate perception, planning and control stacks, it is a notable design choice because one learned model carries most of the work.
Why internet video is the training bet
Language models improved quickly because they could train on text from across the internet. Robotics has lacked an equivalent dataset, and Singh argued that teams collecting their own data through teleoperation produced lab demos that rarely held up in real deployments. Rhoda's answer is to pretrain on internet-scale video, which Singh said covered hundreds of years of footage, from baseball games to waves on a beach.
Singh also said the team tested more curated datasets that showed only human hands manipulating objects, and found that the best models were the ones trained on the broadest data. After pretraining, Rhoda applies a small amount of task-specific training, which he put at 10 to 20 hours, comparing it to teaching an experienced worker a new assembly line workflow. Rhoda's website describes the split as more than a year of web-scale video in pretraining and 1 to 10 hours of trajectory data in post-training. Robot embodiment only matters in that post-training phase, which is why Rhoda says one model can run on different hardware.
Mobile DVA adds movement to manipulation
On November 3, Rhoda published research on Mobile DVA, which extends the model so a single policy controls both the arms and the mobile base. The company says the model can learn new mobile tasks from as little as 15 minutes of additional robot data, building on its video pretraining and existing manipulation experience. It needs no maps, SLAM or separate navigation planner, and Rhoda says the robot adapts to new obstacles and changed destinations because it keeps updating its video predictions from fresh camera images.
That matters for anyone deploying robots in warehouses or factories, where a manipulation arm that can also reposition itself covers far more work than one bolted to a station. Rhoda told The Robot Report it focuses internally on manufacturing and logistics, and lists automotive and e-commerce customers as well.
What it means for Palo Alto and South Bay robotics builders
Rhoda describes itself as full stack. Singh said the company always owns the AI layer, but customers can run its models on Rhoda's hardware or a third-party robot, and skills can be trained by Rhoda or by partners. For hardware founders around Palo Alto, Mountain View and Sunnyvale, that opens a path where a startup builds the body or the workflow and licenses the intelligence, similar to how software startups build on cloud models.
It also raises the bar on data strategy. If a model pretrained on video needs only hours of task data, the advantage shifts toward teams with access to real customer sites and well-defined jobs. That fits a wider wave of capital into the category, from Agile Robots' $500M Series F to the demand we tracked in Q3 2026 robot orders. If you are prototyping your own hardware, our Circuit Launch Mountain View guide covers one of the local spaces where robotics teams build, and our post on Serve Robotics' delivery API shows another way software teams plug into robots. For a parallel story in AI security, see our breakdown of Corma's $125M Series A.
The South Bay and Peninsula are where a lot of this work happens, and Tribe organizes South Bay Renaissance to rebuild the region's startup and social scene in person, one gathering at a time, with founders' dinners, AI deep dives and demo nights from San Mateo to San Jose. Back in October, Tribe hosted a Physical AI and Robotics Mixer at Circuit Launch in Mountain View for people building in this space.
Find your people in robotics
Singh told The Robot Report he expects robotics to have its own "ChatGPT moment" sooner than many people think, and the best way to judge that claim is to compare notes with people who build and study robots. If you are building robots or love following the field, apply to join Futurists in Robotics, a group for founders, engineers, operators and fans. If you do robotics research as a professor, PhD student or lab researcher, Embodied Frontier is the group built for researchers, and you can apply there. For in-person meetups, follow the Tribe Events calendar on Luma, which lists upcoming South Bay Renaissance gatherings like Designing AI for Humans in Redwood City on November 19.