The startup building foundation models that teach robots to fold, assemble and plate - and to get better with every shift they work.
In a demo that has become the company's calling card, a Dyna Robotics machine ran for more than 24 hours without a human touching it, folding upward of 900 napkins at a 99% success rate. It is a deliberately unglamorous feat, and that is the point. Dyna Robotics is betting that the road to intelligent machines runs not through flashy humanoid stunts but through the repetitive, low-margin labor that keeps hotels, laundries, restaurants and factories running - and that a business will pay for.
Founded in 2024 and headquartered in Redwood City, California, Dyna Robotics builds embodied-AI foundation models: the software brain that lets a robot perceive a messy real-world scene and act with dexterity. Its flagship, DYNA-1, is what the company calls a single-weight, general-purpose model - one neural network that can be pointed at different jobs without being reprogrammed for each. The hardware is comparatively ordinary: two industrial arms mounted on a wheeled base. The intelligence is the product.
The company was assembled by people who have done this before. Co-founders Lindon Gao and York Yang previously built Caper AI, whose AI-powered smart shopping carts sold for a reported $350 million. They paired with Jason Ma, a former Google DeepMind researcher who worked on foundation models for robotics. That mix - repeat operators plus frontier AI research - helps explain how a company barely a year old attracted checks from NVIDIA, Amazon, Salesforce, Samsung and LG.
Most industrial robots are precise but brittle. They repeat a programmed motion flawlessly and fail the moment the world shifts an inch. Dyna's approach inverts that. Instead of scripting a task, its robots run DYNA-1, a model trained on millions of real manipulations that aims for what the company describes as zero-shot performance - the ability to walk into a new environment and work without task-specific programming. Field teams map a workspace, often with an iPad, and calibrate the robot in about ten minutes. Then it starts.
The tasks Dyna showcases are pointedly mundane: folding 40-plus shirts an hour at human-level quality, folding napkins, plating food, and handling assembly, quality inspection and packaging on factory lines at what the company reports as 99%-plus reliability. The unifying thread is dexterity under variation - the soft, fiddly, semi-predictable work that has resisted automation even as arms and grippers got cheaper.
That last claim points to Dyna's real ambition. Each deployment - every shirt folded in a real laundromat, every part placed on a real line - generates data that feeds back into the foundation model. Improvements made from one customer's floor propagate to the whole fleet. Dyna frames the endpoint as physical AGI: a single robot intelligence general enough to handle the diversity of daily physical work.
Dyna aims squarely at labor-intensive service and industrial businesses - hotels, restaurants, laundromats, gyms and factories - where wages have climbed and workers are hard to keep. Analysts tracking the company note service and manufacturing wage inflation of roughly 15-30% since 2022, a pressure that makes reliable automation newly attractive. In early commercial deployments, Dyna has reported robots running about 16 hours a day.
The problem Dyna solves is not "can a robot do this once" but "can a robot do this all day, every day, without an engineer babysitting it." Endurance and self-correction are the features. DYNA-1 is designed to catch its own mistakes mid-task and recover in real time across multi-step workflows - the difference between a demo and a deployment.
For a business owner, the pitch is concrete: a machine that shows up, learns the room, and handles a task that is chronically understaffed. For Dyna, each of those installations is also a sensor collecting the training data that widens its lead. The customer buys reliable hours of work; the company buys a better model.
That framing also shapes which markets Dyna enters first. Rather than chase the hardest, most prestigious robotics problems, it targets high-repetition tasks with clear economics - places where "good enough, all the time" beats "brilliant, sometimes."
Dyna sells its robots as a service. Rather than a large upfront hardware purchase, customers pay a recurring monthly fee per robot that bundles the machine, the DYNA foundation model, maintenance and software updates. That Robots-as-a-Service (RaaS) structure turns a capital expense into an operating one and lowers the barrier to trying automation - a familiar SaaS playbook applied to the physical world.
The model has a second edge. Because every deployment enriches the shared model, Dyna's fleet gets collectively smarter as it grows - a data flywheel that is hard for a pure hardware seller to replicate. Observers have floated a longer-term path in which Dyna licenses its foundation model to third-party robot makers, a role analogous to NVIDIA's CUDA layer for AI. For now, the recurring per-robot subscription is the engine.
A production-ready, single-weight robot foundation model for sustained autonomous manipulation. Demonstrated 99%+ success over 24-hour continuous runs, folds 40+ shirts an hour at human-level quality, and self-corrects in real time on multi-step tasks.
Dual industrial arms on a wheeled base, deployed as Robots-as-a-Service. Field teams map the workspace and calibrate in minutes; robots then run across factories, laundries, restaurants and hospitality with no task-specific programming.
Dyna sits in the emerging category of robot foundation models, alongside rivals such as Covariant, whose RFM-1 targets warehouses; Alphabet's Intrinsic; and well-funded newcomers like Physical Intelligence and Skild AI. It also brushes against established collaborative-robot makers like Universal Robots and vertical specialists like the kitchen-focused Miso Robotics.
Dyna's differentiation is its insistence on learning in production. Where some competitors lean on simulation or warehouse-scale datasets, Dyna positions real commercial deployments - the actual laundry, the actual line - as both its go-to-market and its data source. The risk is real: rivals like Covariant already hold large warehouse datasets, hardware can commoditize, and a loosening labor market could soften the ROI case. Dyna's answer is to compress the loop between deployment and improvement, and to sell reliability that a business can measure in uptime.
| Round | Amount | Date | Lead & Notable Investors |
|---|---|---|---|
| Seed | $23.5M | Mar 2025 | CRV, First Round Capital |
| Series A | $120M | Sep 2025 | RoboStrategy, CRV, First Round Capital, Salesforce Ventures, NVentures (NVIDIA), Amazon Industrial Innovation Fund, Samsung Next, LG Technology Ventures |
Lindon Gao and York Yang sell Caper AI, their AI-powered smart-cart company, for a reported $350 million.
Gao, Yang and ex-DeepMind researcher Jason Ma launch Dyna in Redwood City to build robot foundation models.
CRV and First Round Capital co-lead a seed round to build DYNA-1.
The company reveals DYNA-1, billed as the first commercial-ready robot foundation model with round-the-clock autonomous dexterity.
A round led by RoboStrategy, CRV and First Round pushes the valuation past $600M, with NVIDIA, Amazon, Salesforce, Samsung and LG joining.
It builds embodied-AI foundation models - led by DYNA-1 - that power general-purpose robots to perform dexterous, real-world manipulation tasks continuously in commercial settings.
Repeat founders Lindon Gao (CEO) and York Yang, who previously sold Caper AI for a reported $350M, together with former DeepMind researcher Jason Ma. The company was founded in 2024.
About $143.5M total - a $23.5M seed in March 2025 and a $120M Series A in September 2025 - at a valuation above $600M.
RoboStrategy, CRV and First Round Capital led its rounds, with NVIDIA's NVentures, Amazon Industrial Innovation Fund, Salesforce Ventures, Samsung Next and LG Technology Ventures participating.
Through a Robots-as-a-Service model - customers pay a recurring monthly per-robot fee bundling hardware, the DYNA model, maintenance and updates.