Rhoda AI Exits Stealth with $450 Million Series A to Build Robot Foundation Models That Learn from Internet Video
Palo Alto startup Rhoda AI emerged from 18 months of stealth with a $450 million Series A at a $1.7 billion valuation, unveiling a robotics intelligence platform trained on hundreds of millions of internet videos.
Overview
Palo Alto-based Rhoda AI emerged from eighteen months of stealth on March 10, unveiling a $450 million Series A round that values the company at $1.7 billion. The startup is building what it calls a robot foundation model, trained not on conventional teleoperation data but on hundreds of millions of publicly available internet videos, in an effort to close the gap between laboratory demonstrations and real-world industrial deployment.
The round was led by Premji Invest, with participation from Khosla Ventures, Temasek Holdings, Capricorn Investment Group, Mayfield, Matter Venture Partners, Leitmotif, Prelude Ventures, Xora, and venture capitalist John Doerr.
What We Know
Rhoda is led by CEO and co-founder Jagdeep Singh, a serial deep-tech entrepreneur best known for founding QuantumScape, the solid-state battery company. The scientific leadership includes Chief Science Officer Eric Ryan Chan, a Stanford researcher and former generative model architect at WorldLabs, and Gordon Wetzstein, a Stanford professor who heads the university’s Computational Imaging Lab.
The company’s core product is FutureVision, a robotics intelligence platform built around what Rhoda calls its Direct Video Action architecture. Unlike conventional approaches that rely primarily on teleoperation—where a human operator guides a robot through tasks to generate training data—Rhoda’s system first pretrains models on internet-scale video libraries to learn how objects move and how the physical world behaves. The model then fine-tunes with smaller amounts of actual robot data and integrates perception and control through continuous feedback, running predictive cycles dozens of times per second.
Singh has argued that internet video offers a decisive advantage over teleoperation alone. In conventional setups, if an object’s orientation changes from what the robot saw during training, the model can fail. By learning from millions of video examples showing objects at different orientations, in different lighting, and under varied conditions, Rhoda’s system is designed to generalize more robustly.
Rhoda has already demonstrated its technology in industrial settings. The company has completed component-processing workflows in under two minutes per cycle and reports reliability exceeding internal performance targets in high-variability environments, including a successful pilot with an automotive manufacturer using off-the-shelf robotic hardware.
What We Don’t Know
Rhoda has disclosed little about the specific industrial partners involved in its pilot deployments beyond a reference to an automotive firm. The company has not published benchmarks comparing FutureVision’s performance against competing approaches from firms such as Google DeepMind, Physical Intelligence, or Covariant. Details on the scale of the team, the size of its video training dataset, and the compute infrastructure underpinning the platform remain undisclosed.
The company has signaled plans to build its own humanoid-style robots and develop proprietary hardware, but no timeline or specifications have been shared. How Rhoda intends to balance in-house hardware development with its stated plan to license AI models across third-party platforms is also unclear.
Analysis
The $450 million Series A is among the largest early-stage raises in the robotics sector and reflects growing investor conviction that the “physical AI” thesis—applying foundation model techniques to real-world robotic manipulation—is ready to move beyond research. The round places Rhoda alongside a small cohort of heavily capitalized robotics AI startups including Figure AI, which is reportedly in talks for funding at a $39.5 billion valuation, and Physical Intelligence, which raised $400 million in late 2024.
Singh’s track record at QuantumScape—where he raised billions for a deep-tech bet on solid-state batteries—likely contributed to the scale of the round. However, QuantumScape also illustrates the risks of deep-tech timelines: that company took years longer than projected to reach commercial production. Whether Rhoda can translate laboratory demonstrations into reliable, scalable industrial deployments will be the central test of the investment.
The broader context is favorable. Humanoid robots from Boston Dynamics, Figure AI, and Agility Robotics are entering factory floors in 2026, and major manufacturers including BMW, Toyota, and Hyundai have committed to expanding robotic deployments. An intelligence layer that can make these machines more adaptable to unpredictable real-world conditions represents a potentially large addressable market—if Rhoda’s video-trained models can deliver on the promise of generalization that has eluded the field for decades.