Dyna Robotics Founder Lindon Gao on Moving From Robot Tasks to Workflows

Zhang Xiaojun speaks with Dyna Robotics founder and CEO Lindon Gao about entrepreneurship, robot product strategy and the mindset needed for a long journey. Gao compares robotics to an unfinished map: teams must explore technical frontiers while finding products that work sustainably in the real world.

Gao recalls telling his team, after repeated fundraising rejections, that they were like cockroaches: hard to kill. He sees no contradiction between an ambitious ideal and commercial survival; a distant dream still needs a practical path.

Visiting Shanghai, he sees rapid releases and fundraising in both China and Silicon Valley, alongside different emphases on long-term strategy.

He describes some teams as focusing on the robot’s “brain” and others on hardware and supply chains; Dyna calls its own approach full-stack.

Born in China, Gao moved to the United States at 11.

Small businesses were common around him. At 14 he sold game accounts and items and built a trading site; he says what attracted him was the feedback from doing something well, not money alone.

In college he started a jewelry business, first selling inventory online and later connecting Chinese suppliers with American designers.

A family member helped manage it for years, but he eventually closed it to direct his time elsewhere.

Gao initially studied medicine, switched to finance and spent more than two years in investment banking.

He says joining banking partly reflected the competitive expectations around him, but he eventually left to start a company.

His earlier company Caper explored cashierless retail.

While many startups wanted to retrofit entire stores, Gao interviewed supermarket operators and found little appetite for renovation, maintenance and uncertain returns.

The team put computer vision and sensors in the shopping cart instead. The cart travels with the person paying and identifies items placed inside, a potentially easier deployment path for determining who pays and how much.

Caper entered Y Combinator but struggled to raise money afterward.

Gao recalls founders and early employees living together and keeping expenses low for a long period before securing a seed round.

He says he contacted more than a thousand angel investors and found only a few supporters, while many institutional investors also declined.

He remembers it as a test of endurance, not a guarantee that persistence alone makes a product succeed.

Hardware development and supermarket sales cycles were long.

Gao says the company eventually found backers and made progress; later, while raising another round, it received an acquisition offer and chose to sell.

On product building, he emphasizes understanding what customers actually need to accomplish rather than merely showing a technology.

A feature’s lasting value becomes visible in the customer’s real workflow.

Gao calls founders cognitive athletes: physical condition affects judgment.

He describes developing exercise and dietary routines to sustain his energy over a long period of work.

After Caper, he worked inside a larger organization alongside leaders from major technology companies. The experience let him observe different ways teams coordinate, communicate and make decisions.

When he returned to founding, Gao saw robotics as a multidimensional problem involving models, hardware, motion control and actual use.

Choosing a direction and assembling a complementary team were early priorities.

Gao considers choosing a direction one of his strengths, but acknowledges blind spots.

He explains his reasoning to others and invites disagreement; in technical partners, he values depth and structured thinking.

Zhang asks whether a team can split between near-term deployment and foundational research.

Gao says Dyna faced that tension when its direction was unclear, and members struggled to define what kind of company they were building.

He argues that deployment alone does not prove scalable product-market fit, while research pursued without application may not reveal the capabilities a product needs.

The team sought a longer-term goal that could advance intelligence and lead to deployment.

Gao calls the approach product research. Research remains central, but the team works backward from a future product to set boundaries and priorities. With limited time, it tries to avoid spreading its effort across too many directions at once.

Around the end of 2024, the company assembled its team and began speaking with customers. Gao recalls that early foundation models could perform individual actions but were slow and struggled with sequences. The team set reliability and speed targets for its research.

He describes the Dyna One folding demonstration and says the team achieved high success rates under particular conditions. The speed and reliability figures in the interview are company-reported and do not establish performance in every setting or long-term deployment.

Why folding?

Gao says the task allows some recovery from errors: a misfold can be retried, unlike a spilled drink. He also saw repetitive work in laundries and hotels and treated folding as a way to explore a first commercial loop.

Cloth differs from a rigid object: each step changes the state faced by the next, and sizes, fabrics and settings vary.

He says the team used T-shirt folding to investigate generalization, quality and speed, including what data helped in new environments.

Deployment in a real restaurant exposed the gap between a demonstration and work. The desired napkin required five folds, but the early robot completed only four, leaving staff to finish. Although the arrangement had been discussed, customers were unhappy in practice.

The customer needed many napkins a day, and initial output was insufficient.

After improving speed, the team found edge cases: picking up two napkins could take much longer to resolve than a normal fold, reducing the day’s production.

Even placing folded napkins neatly in a container was a problem.

Gao says the team had to teach the model about placement order and quality, and reconsider what pretraining and post-training should each address.

Individual rare failures can be patched with special engineering rules, but the team worried that this would not scale. Gao prefers the model to handle more variation, while acknowledging that real deployment exposes small issues that matter to delivery.

He concluded that performing the fold was not yet a scalable product.

After finishing a stack, the robot needed an employee to replenish it; if the person did not return in time, the machine could remain idle.

Gao calls this the robot “babysitter” problem. Customers buy output, not just the success rate of a single movement. Scaling an incomplete product too early can also scale the burden of supporting it.

He distinguishes a task from a workflow.

A robot may generalize an action to a new setting without completing a commercial loop. The full workflow includes finding supplies, carrying them, folding, putting finished items away and starting again.

Gao says chaining steps compounds the effect of failures and requires models to understand language and multistep instructions.

Improving pretraining and post-training methods therefore became important to the next stage.

He rejects the idea that Dyna is simply a deployment company.

Product needs direct the research, while research determines whether a product can complete the job; in his view, the two must constrain and inform each other.

Discussing generalization, Gao looks at performance across objects and environments as well as the ability to follow language instructions. Speed, reliability, safety and cost also determine whether a robot becomes a viable product.

Commercial value is not just the number of actions a model can perform. Customers also ask how much work it completes in a day, how much human intervention it needs and whether it is worth buying. Those expectations vary.

Gao sees hotels as environments with many related workflows and compares them to a more standardized household.

Garments, rooms and materials vary, but the boundaries can be easier to define than in an entirely open home; tasks still differ in difficulty.

He prefers to classify problems by capability and workflow complexity rather than simply by commercial versus household labels.

Edge cases encountered in a setting can feed back into model and data research.

The team works on both pretraining and post-training, organizing the latter around specific workflows.

Gao says real tasks help set research priorities and test whether general capabilities are actually useful.

On data, he emphasizes the relationship between quality and the training objective, including first-person and third-person perspectives.

The right data format depends on what the model must learn, rather than treating any method as an end in itself.

In his account, language following is a basic link for connecting a workflow.

He acknowledges that the timing of a major leap in embodied intelligence is uncertain: it could be near or far away.

Gao argues that expanding from one workflow to many requires more data, stronger models and better deployment.

He is optimistic about the path, though his scaling claims remain expectations to be tested in practice.

Asked where robotics stands, Gao says an intelligence demonstration is not enough.

Making a system useful to a broad audience, as ChatGPT did in its domain, also requires the right product form and way to use it.

He treats whether a robot should be humanoid as a product question that includes return on investment.

Form should serve the task and use, rather than becoming the goal simply because a shape attracts attention.

Looking back at Dyna One, he says the company initially emphasized deployment publicly, then realized that a single task did not complete the commercial loop.

Its strategy shifted toward full workflows alongside stronger models and deployment infrastructure.

Gao links product-market fit to actual output and return. If a robot cannot create enough value within its assigned work, a technical demonstration cannot support wide deployment. Building a deployment feedback loop also takes time.

On China’s robotics industry, he says teams are still experimenting and have much to learn from one another.

He expects some convergence in models, data and product practices over the next one or two years, while recognizing that this is a prediction.

He sees a close relationship between robot intelligence and embodiment: hardware, motion control and data collection all affect what can be learned.

Researchers therefore help define camera frame rates, shutter types and fields of view.

For Gao, hardware and models must be developed together.

Waiting until the model is completely solved before defining hardware would not necessarily produce a deployable, manufacturable machine; each side needs feedback from the other.

Describing the company’s changes, he says its models, data, reasoning and hardware direction all evolved.

After a period of uncertainty, the team brought research and deployment under one strategic goal rather than treating them as mutually exclusive.

Organizationally, Gao favors a relatively flat team where leaders can speak directly with engineers and researchers about execution.

He wants a focused group working on the most important problems with a shared understanding of the goal.

He thinks better engineering productivity can let a team explore more directions in the same time, though the organization’s future shape is uncertain.

Faster tools do not replace the responsibility to choose a direction and coordinate people.

The conversation turns to inner discipline. Gao says he once relied more on external measures such as fundraising and valuation; now he pays more attention to thinking clearly and understanding underlying problems. He describes this as a long practice, not a single revelation.

He recounts a personal health difficulty and changes to his routine, focusing on how he changed his attitude toward the problem.

His experience is personal and should not be taken as evidence for someone else’s treatment or health decisions.

Gao says he used to worry intensely about others’ judgments, replaying pauses in a presentation or an imperfect answer. He later tried to rely less on external approval and return his attention to the work itself.

He likens robotics entrepreneurship to a World Cup full of strong competitors and to a marathon with no visible finish line.

Finding a direction through uncertainty has value to him even before the industry settles on a technical or market path.

Gao admits he burdened himself with the idea that a CEO must not make mistakes.

Decisions made primarily to avoid error can themselves be unwise, he says; he thanks his team and family for supporting him through that period.

Looking back, he says the team lacked an independent judgment about its next strategic step and rushed to follow new directions appearing outside, causing drift and wasted time.

It became steadier after clarifying its own goal.

Some limits in one’s understanding require help from others to overcome, he says.

For him, inner discipline means examining his own attachments and learning to see a problem from another person’s perspective.

Zhang asks how he would feel if this venture achieved less than his previous one.

Gao says he still wants to give it everything, but life cannot be measured only by the finish line; without enjoyment of the journey, every stage can feel like an ordeal.

In a quick closing round, he recommends a book that influenced him and mentions the robotics research project Eureka and the importance of infrastructure.

Language, vision and other sensory inputs, he says, help a robot understand the world and turn instructions into action.