I'm thrilled to announce that DYNA has raised $120 million in new funding just as DYNA turned one year old. We are honored to be backed by a group of world-class partners, including Robostrategy, CRV, and First Round Capital, with participation from NVIDIA Ventures, Amazon Innovation Fund, Salesforce Ventures, Samsung Next, and LG Technology Ventures, among others.
[ Updates ]
DYNA Robotics Closes $120M Series A: How We Think About Scaling Robotic Foundation Models
Category:
Updates
Author:
Lindon Gao
Date:
Sep 2025
[ Overview ]
This is my second time building an AI-powered hardware company, and I'll admit it sounds a little crazy to do it again (and I did say that Caper would be my last hardware company). After all, what could be tougher than mixing the headaches of hardware scalability with the research intensity of frontier foundation models?
Yet I've never been more energized, because this moment in history feels truly singular. The convergence of AI breakthroughs, accelerating hardware progress, and the urgency of real-world labor challenges has created a once-in-a-generation opportunity.
Unlocking physical agents is the last frontier to creating unlimited abundance for humanity.
The challenge of this last frontier has drawn world-class talent, each with their own philosophy — from humanoids to world models and everything in between. With so much entropy and no consensus on the right path, it is easy to get lost in the noise.
This is why first principles matter: they anchor decisions in what does not change and serve as the compass through complexity. Robot development cycles are unforgiving — one wrong call can ripple into months of wasted data, while a misguided research direction can set back years of progress. First principles are not just philosophy; they are survival.
While embodied AI companies are grounded in research, the model itself isn't the ultimate product — the robots are. So for once, we're not talking about research principles, but about the first principles guiding us in building a world-class robotics business, something strangely absent from the conversation in our industry today.
[ First Principles ]
DYNA's First Principles
Principle 1 — Physical AGI requires Generalization AND Performance
Generalist models can't commercialize and specialist models can't scale. LLMs benefit from the flexibility of language, where words don't need to be exact, but VLAs demand physical precision, where every action must be precise.
Our approach goes broad through research and data to drive generalization, and deep through commercialization to push for performance. Our commercial deployments don't rely on singular specialist policies; they are generalist models. Both breadth and depth are required to reach Physical AGI.
Principle 2 — Distribution is King
Most people assume the winner in AI is whoever has the best model. But if you've watched SOTA leaderboards, you know they reshuffle weekly. OpenAI may only be on top part of the time, yet they're still considered the king because distribution beats scoreboards.
In robotics, many think solving embodied GPT-3 is the finish line. In reality, it's just the starting gun. Robots can't be shipped overnight on a webpage, and even a perfect model tomorrow would still leave the hard part: productionizing. The company that masters distribution will win. That's why at DYNA we focus on deployment — it jump-starts distribution, forces us to harden research where theory meets reality, shows us what customers actually care about, and spins up the best flywheel of all: real production data.
Principle 3 — ROI is Product-Market Fit
Solving the model does not equate to product-market fit for embodied AI. Our customers never ask us about success rates. Their sole concern is ROI. For a robotics foundation model company to thrive, every research endeavor must answer a fundamental question: how does our research translate to business outcomes?
ROI is value over cost. Value comes from throughput and quality, which in the physical world depend on the entire robot stack, not just the model. That's why “model-only” companies struggle: meeting real-world requirements demands end-to-end control across data, inference, control, and hardware. DYNA-1 has already reached 60% of human throughput at a stringent quality bar, a milestone unmatched today. On the cost side, total ownership — hardware, durability, maintenance, and downtime — must stay below thresholds like $50k, or humans remain cheaper.
Scalable ROI is the ultimate goal, as one-off ROI is meaningless. This demands generalization, which we categorize into four layers: environment generalization, SKU generalization, embodied reasoning, and task generalization.
Principle 4 — Effortless Data, Enduring Quality
The key obstacle to achieving the performance and generalization needed for sustained ROI is scaling the right kind of data.
Embodied AI lags far behind language and multimodal models for a simple reason: there isn't enough data, and unlike text on the internet, we have to generate it ourselves. This makes data the highest-entropy battleground in robotics, and unlike model architectures — which can be tweaked or scrapped — the wrong data strategy can sink the whole ship. Two things we care about the most:
Quality. Success rate alone is not enough. In our deployments, success is table stakes; true quality means data that improves generalization, consistency, throughput, and robustness. That's why DYNA-1 was built with a reward model that optimizes beyond “did it succeed” to include robustness and speed — the only scalable way to measure and improve across millions of actions.
Ease of curation. A million hours of brute-force data may look good in a pitch deck, but the real breakthrough is the ability to effortlessly generate 10,000 hours of high-quality data every day — understanding what quality truly means, then scaling it aggressively and efficiently.
Principle 5 — Iteration Speed Wins the Race
In embodied AI, there's often debate about which research path or architecture will ultimately succeed. But in a high-entropy environment, the only way to discover the truth is to move swiftly. Iteration reveals what works.
Our focus has been on developing infrastructure that enables rapid training, evaluation, and deployment, at both the architectural and data levels. At DYNA, we run hundreds of model variants weekly across offline and online evals. Over time, this compounding cycle of rapid iteration, evaluation, and deployment becomes our ultimate competitive advantage.
[ What's Next ]
What's next
This $120 million will accelerate four critical areas of our mission:
01
Building a world-class team. Attracting top researchers and engineers who want to build the foundational intelligence for the physical world and see their work make a real impact in production environments.
02
Infrastructure expansion. Doubling down on our data pipeline and model iteration capabilities to handle the exponential growth of data and deployment.
03
Use-case diversification. Expanding to more diverse environments and tasks to accelerate generalization across our foundation model.
04
Continued commercialization. Building the GTM flywheel to continue scaling deployment.
[ Join Us ]
Join us
Our mission is to build high-performance, general-purpose robots.
If you're an enterprise looking to solve real-world labor challenges, let's set up a pilot. If you're a supplier who believes in building for the long term, let's talk.
And if you're a builder who wants to ship — we're hiring. At DYNA, we're offering you a chance to be part of the core engine building physical-world AGI: to ship your research and see it live on robots, solving real problems, within days.
Join us. Visit us at dyna.co/careers.
[ Stay Updated ]
Our research straight to your inbox.
