Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchRhoda AI emerges from 18 months of stealth with a $450M Series A at $1.7B valuation — training robots on hundreds of millions of internet videos instead of teleoperation, backed by Khosla Ventures and Premji Invest.
AI, startups & growth insights. No spam.
TL;DR: Rhoda AI emerged from 18 months of stealth on March 18, 2026, announcing a $450 million Series A at a $1.7 billion valuation. The round was led by Premji Invest, with participation from Khosla Ventures, Temasek, Mayfield, and Capricorn. The company's core thesis is that robots should learn from hundreds of millions of internet videos — the same data that taught humans how the physical world behaves — rather than from expensive human teleoperation. Rhoda's exit lands during a single week in which the robotics sector collectively raised over $1.2 billion, signaling a market-wide conviction shift toward embodied AI at scale.
On March 18, 2026, Rhoda AI stepped out of an 18-month stealth period with a fully formed fundraising story: a $450 million Series A at a $1.7 billion post-money valuation — making it a unicorn on its first institutional round.
The round was led by Premji Invest, the family office of Wipro founder Azim Premji, which has steadily expanded its AI infrastructure portfolio over the last 24 months. Co-investors include:
| Investor | Type |
|---|---|
| Premji Invest (lead) | Family office |
| Khosla Ventures | Venture capital |
| Temasek | Sovereign wealth |
| Mayfield | Venture capital |
| Capricorn Investment Group | Impact / deep tech |
The investor mix is notable for its diversity of mandate. Khosla Ventures has a long history of early-stage deep tech bets and has backed robotics companies including Agility Robotics. Temasek brings a long-duration sovereign capital lens particularly suited to hardware and infrastructure cycles. Capricorn's presence suggests Rhoda's work intersects with industrial sustainability themes. And Premji Invest's decision to lead — rather than merely participate — is an implicit statement about conviction in the team.
A $450 million Series A is not typical. The median Series A in AI in 2025 was approximately $35 million. A round nearly 13 times the median, taken at unicorn valuation, as the first institutional raise, means investors are pricing a category-defining outcome — not a product experiment.
Sources: TechFundingNews, Bloomberg
Rhoda AI was founded approximately 18 months before its public announcement, placing its founding in late 2024. The company operated in complete stealth — no press releases, no conference appearances, no product leaks — for the entirety of that period.
In robotics, stealth is often forced by hardware development timelines. Physical systems require iteration cycles that are slower than software. A company that surfaces too early will be evaluated on pre-production prototypes, creating reputational and fundraising risk if the hardware does not perform. Rhoda's 18-month silence is most plausibly explained by the time required to:
The company's choice to emerge only when it had a complete funding story — not a bridge round or seed extension — suggests the stealth period was used to reach a point where the scientific thesis was substantiated enough to support a unicorn valuation argument.
The central claim Rhoda AI makes is this: robots should learn about the physical world from video, not from human operators moving robot bodies around.
To understand why this is significant, you need to understand the dominant alternative.
Most leading robotics AI companies today — including Physical Intelligence (PI), Figure, and Apptronik — rely heavily on teleoperation to generate training data. A human operator wears a haptic glove or VR interface and moves a robot arm or full-body robot through a task in real time. The robot's sensor readings and motor commands are recorded. Thousands or tens of thousands of these demonstrations are then used to train imitation learning or reinforcement learning models.
Teleoperation produces high-quality, task-specific data. The robot learns from demonstrations that actually happened in the physical world with real physics and real objects. But it has critical scaling limitations:
The internet contains hundreds of millions of hours of video showing humans and other agents interacting with the physical world. Cooking videos show hands manipulating objects, responding to weight, slippage, and deformation. Sports footage encodes complex multi-limb coordination, balance recovery, and reaction to unexpected forces. Industrial and craft videos document fine motor skills, tool use, and workspace organization.
Rhoda AI's core claim is that this data, if the right learning signal can be extracted from it, contains the physics and motion priors that robots need — at a scale that teleoperation could never match and at a cost approaching zero per demonstration.
This is not a new idea theoretically. Researchers have explored video-based robot learning for years. What has changed is:
Rhoda AI's $450 million is a bet that the time when this approach becomes the dominant training paradigm is now.
Rhoda's video pretraining approach targets three interrelated categories of knowledge:
Watching humans and animals move across millions of hours of video encodes statistical regularities about how bodies navigate space — weight shifting before a step, arm extension before a grasp, head orientation before a directional change. A model that learns these priors can generate physically plausible motion plans without needing to exhaustively sample from a physics simulator or collect robot-specific teleoperation data for every variation.
Video contains implicit physical information that supervised physics simulators often miss: the way a full glass moves differently from an empty one, the deformation of soft objects, the recovery dynamics when an agent loses balance. This implicit physics is encoded in pixel sequences in ways that large vision models have proven capable of extracting. Rhoda is training on this data at a scale sufficient to build a generalizable internal physics model — not a lookup table of specific scenarios, but a learned intuition about how the world responds to force.
The highest-value learning target in Rhoda's framework is physical interaction: contact-rich manipulation, tool use, bimanual coordination, and reactive grasping. These are the behaviors that require robots to understand not just their own motion but the causal chain between their actions and the state of objects in their environment. Video of human hands assembling furniture, preparing food, or operating machinery is rich with this information at a density that teleoperation datasets, even large ones, cannot match simply due to collection cost.
The embodied AI field in 2026 is crowded with well-capitalized competitors. Rhoda AI is entering a landscape that includes:
| Company | Training Approach | Total Capital | Key Differentiation |
|---|---|---|---|
| Physical Intelligence (PI) | Teleoperation + generalist policy | $510M | π₀ foundation policy, multi-robot |
| Figure AI | Teleoperation + OpenAI partnership | $754M | BMW manufacturing deployment |
| Agility Robotics (Amazon) | Teleoperation + simulation | Amazon-backed | Warehouse logistics at scale |
| Boston Dynamics / Google DeepMind | Simulation + real-world fine-tuning | Hyundai-backed | Locomotion leadership |
| Rhoda AI | Video pretraining (internet scale) | $450M | Data efficiency, coverage breadth |
The key differentiator Rhoda is betting on is data efficiency at the frontier. If video pretraining works as a foundation, Rhoda can potentially train a general robot policy on a broader coverage of physical scenarios than any teleoperation-based competitor — at a lower marginal cost per new scenario. The competitive moat, if it materializes, is not hardware or any specific application — it is the ability to rapidly generalize to new tasks without expensive data collection campaigns.
Physical Intelligence, the most direct comparison as a generalist robot policy company, has demonstrated strong results with teleoperation-based approaches but has publicly acknowledged the scaling cost challenge. PI's π₀ model, released in late 2024, was trained on a large multi-robot teleoperation dataset that required significant investment to assemble. Rhoda is proposing a different scaling curve.
Rhoda AI's announcement lands in a week that saw the robotics sector collectively raise more than $1.2 billion across multiple major rounds. This concentration of capital is not coincidental — it reflects a shift in investor conviction that has been building since late 2024.
Several factors are driving the surge:
Foundation model transfer is proving out. The research community has accumulated evidence that large-scale pretraining — whether on text, images, or video — transfers useful priors to downstream physical tasks. Physical Intelligence's π₀, Google DeepMind's work on generalist manipulation, and Embodied Intelligence's demonstration results have collectively moved robotics from "science project" to "infrastructure investment" in the minds of institutional allocators.
Labor economics are making the case. With persistent skilled labor shortages in manufacturing, logistics, elder care, and construction, the ROI calculus for deployable robots has improved materially. Companies deploying robots in warehouses and factories are now reporting positive unit economics at scale — a threshold the industry had been approaching for years but not consistently clearing.
The hardware cost curve is finally moving. Actuators, depth cameras, and on-device inference chips have reached price points where humanoid or semi-humanoid platforms can be manufactured at costs that support enterprise pricing. This hardware maturation is a prerequisite for software-first robotics companies — you cannot build a data moat on platforms that cannot be profitably deployed.
The $1.2 billion week is a sentiment marker. When sovereign wealth funds, major growth equity firms, and strategic corporates all move in the same week, it signals that the asset class has crossed a credibility threshold, not just that individual companies are promising.
That Premji Invest led a $450 million Series A in a stealth robotics company deserves analysis. Family offices typically participate in rounds — they rarely lead them at this scale.
Azim Premji's investment philosophy, as executed through Premji Invest, has consistently favored companies with structural technological differentiation over near-term commercial traction. The portfolio includes Anthropic, Recursion Pharmaceuticals, and a range of AI infrastructure bets where the scientific thesis precedes the revenue curve by years.
Leading Rhoda's round fits this pattern exactly. Video pretraining for robots is a scientific thesis that has not yet produced a commercially deployed product at scale. Premji Invest's willingness to lead — and to price the company at $1.7 billion without shipping revenue — indicates a conviction about the correctness of the underlying approach, not just the market size.
The co-investors reinforce this read. Khosla Ventures has historically backed science-first deep tech companies — Willow Garage in robotics, OpenAI in AI, QuantumScape in batteries. Temasek brings patient capital with a 10-20 year investment horizon typical of sovereign wealth. Capricorn's deep tech focus and Mayfield's enterprise software heritage complete an investor group that collectively has managed transformative technology transitions before.
The case for video pretraining is compelling on paper. The risks are real.
The embodiment gap is unsolved. Human video shows human bodies performing tasks with human morphology. Transferring those motion and physics priors to a robot with different joint configurations, actuator torque limits, and sensor modalities is a hard generalization problem. Early research results show partial transfer but not seamless transfer. Rhoda's success depends on closing this gap at scale — and it is not clear the gap has a clean solution.
Action labels are absent from most video. Video shows what happens, not what commands produced it. A video of a hand opening a jar encodes the outcome but not the precise force trajectory, contact schedule, or controller state that achieved it. Learning an actionable policy from this data requires inference about the action space — a significantly harder problem than supervised imitation learning from teleoperation, where the commands are recorded directly.
Distribution mismatch between video and robot deployment contexts. The internet is heavily biased toward human-scale tasks, indoor environments, and consumer product interactions. Industrial automation, clean-room manufacturing, outdoor construction, and other high-value robot deployment scenarios are underrepresented in the video corpus. Rhoda will need to supplement internet video with targeted collection or simulation to cover its commercial use cases.
Competition is not standing still. Physical Intelligence, Figure, and Agility Robotics are not ignoring video as a data source. PI's research team has published work on video-based policy pretraining. If the video pretraining thesis validates, the competitive advantage shifts from "we figured it out first" to "we built more compute and collected more data" — an arms race where Rhoda's $450 million may not be sufficient insulation.
Rhoda AI now has the capital and the public profile. The next 12 months will reveal whether the video pretraining thesis holds up under production conditions.
Key indicators:
What is Rhoda AI? Rhoda AI is a robotics AI company founded in late 2024 that trains robots using hundreds of millions of internet videos rather than relying primarily on human teleoperation. It exited stealth on March 18, 2026, announcing a $450 million Series A at a $1.7 billion valuation.
How much did Rhoda AI raise? Rhoda AI raised $450 million in a Series A round, its first institutional fundraise, announced upon exiting stealth in March 2026.
What is Rhoda AI's valuation? The Series A values Rhoda AI at $1.7 billion post-money, making it a unicorn on its first institutional round.
Who led Rhoda AI's Series A? Premji Invest led the round. Co-investors include Khosla Ventures, Temasek, Mayfield, and Capricorn Investment Group.
What does it mean to train robots on internet video? Instead of recording humans manually operating robots (teleoperation) to generate training data, Rhoda AI trains its models on video from the internet — cooking tutorials, sports footage, industrial and craft videos, and other content showing physical interaction. The models learn motion priors, physics intuition, and interaction patterns from this data before being adapted for robot deployment.
How long was Rhoda AI in stealth? Rhoda AI operated in stealth for approximately 18 months before its public announcement in March 2026.
What is teleoperation and why does Rhoda AI avoid it? Teleoperation involves a human operator controlling a robot remotely to demonstrate tasks while the robot's sensor and motor data is recorded for training. It produces high-quality task-specific data but is expensive, slow, and hard to scale. Rhoda AI argues that internet video provides equivalent or superior coverage of physical interaction at far greater scale and lower cost.
How does Rhoda AI compare to Physical Intelligence? Physical Intelligence (PI) raised $510 million and trains its generalist robot policy (π₀) primarily on large-scale teleoperation data. Rhoda AI proposes video pretraining as an alternative or complementary data source, arguing it enables broader coverage of physical scenarios without proportional teleoperation cost. Both companies target general-purpose robot manipulation but differ on training data philosophy.
What is the $1.2 billion robotics mega-round week? The week of March 18, 2026, saw multiple major robotics AI companies announce or close large funding rounds totaling over $1.2 billion in aggregate — including Rhoda AI's $450 million — reflecting a sustained surge in investor conviction about embodied AI as the next major infrastructure category.
What are the risks of video-based robot training? Key risks include the embodiment gap (human videos show human bodies, not robot bodies), the absence of action labels in video (policies must infer commands from outcomes), distribution mismatch between available video and target deployment environments, and competition from well-capitalized incumbents who may adopt similar approaches.
What is Premji Invest's typical investment style? Premji Invest, the family office of Wipro founder Azim Premji, is known for long-duration, science-first bets in deep tech and AI. Portfolio companies include Anthropic and Recursion Pharmaceuticals. Leading Rhoda AI's round follows a consistent pattern of early, high-conviction bets where commercial traction lags the technical thesis.
Which industries could Rhoda AI's robots serve? While Rhoda AI has not disclosed specific vertical targets, video-pretrained robots with broad physical understanding are applicable to manufacturing, logistics, elder care, food service, construction, and any domain involving unstructured object manipulation and locomotion.
Does Rhoda AI manufacture its own robots? No public information indicates Rhoda AI manufactures hardware. The company appears to focus on the training and policy layer — the AI models that govern robot behavior — and may target deployment on third-party robot platforms or through partnerships with robot OEMs.
How does video pretraining compare to simulation-based robot training? Physics simulators generate training data at low cost but suffer from sim-to-real gaps — behaviors that work in simulation often fail on real hardware due to modeling inaccuracies in contact physics, friction, and material deformation. Internet video, while noisier and lacking action labels, encodes real-world physics as it actually occurs. Rhoda AI's approach positions video as a complement or alternative to simulation for building real-world physical priors.
What is the next milestone to watch for Rhoda AI? The most consequential near-term milestone is a public product demonstration or research publication showing that video-pretrained robot behavior generalizes to tasks the model was not explicitly trained for — proving the core scientific thesis under real conditions rather than curated demo scenarios.
Sunday Robotics raises $165M Series B at $1.15B valuation, planning to ship consumer humanoid robots for under $25,000 by Thanksgiving 2026 — the first unicorn focused exclusively on household robotics.
Mind Robotics — spun out of electric vehicle maker Rivian — raises $500M in a Series A co-led by Accel and a16z at roughly $2B valuation, applying EV manufacturing expertise to build industrial AI robots.
Four robotics startups — Mind Robotics, Rhoda AI, Sunday, and Oxa — collectively raised over $1.2 billion in a single week, signaling that robotics has entered the same mega-round investment cycle that defined LLMs in 2024-2025.