Robot Demos Are Easy. Reliability Is Hard — Jason Ma, Dyna Robotics

A robot that can perform a useful physical task once has cleared an important hurdle. A robot that can keep doing it in a customer's workplace faces a different test: changing conditions, imperfect hardware, and mistakes that alter what happens next. Jason Ma, co-founder and CTO of Dyna Robotics, introduces the company's goal as one platform that can perform many economically useful tasks with enough reliability for commercial use.

The research and deployment cycle starts with what the models and hardware can already do. Those capabilities determine which customer workflows Dyna can attempt. Deployment then supplies both training data and a sharper diagnosis: which parts of the model or hardware still fail? Ma reports more than five deployment sites at the time of the talk. Their value goes beyond demonstrating progress; they help choose which of robotics' many research problems deserve attention next.

The basic manipulation pipeline begins with demonstrations. A human controls the robot to perform a task, such as folding a T-shirt, and the resulting data trains a neural network. At execution time, the network receives a representation of the scene—in Ma's introductory example, a camera view of the table—and outputs joint positions or torques that move the robot. Learning the visual task and learning the physical action are connected through those demonstrations.

Scaling that pipeline runs into a data problem: robot demonstrations are scarce. Dyna's pre-training data pyramid combines three sources with different jobs:

• Off-robot data: Human-worn camera recordings and public datasets provide diverse experience without requiring every example to be collected on a robot. Simulation data is described as something the team is considering.
• Robot-task data: Demonstrations across industrial, household, laundromat, and hotel tasks teach the precise actions that off-robot observations alone cannot supply.
• Deployment data: Experience from customer environments helps close the gap between training in a laboratory and operating somewhere else.

Ma reports more than 200,000 hours of data in the training pipeline, across these sources.

The architecture pairs a high-level reasoning model with a low-level world action model. The reasoning component supplies semantic understanding of the situation; the action component produces fine-grained, high-frequency movements. Physical work requires both: understanding what should happen does not by itself supply the precise interaction needed to complete a fold or recover from a bad grasp. The talk gives this division of responsibility without specifying the interface or control frequency.

A useful pre-trained model makes new tasks cheaper to learn. Ma presents two dexterous tool-use examples trained with less than one hour of task-specific data, including tasks absent from pre-training. The office gallery also includes cleaning trash, opening a box, and folding towels and T-shirts. These examples establish the appeal of a general foundation: each new task can build on capabilities already learned. But repeated commercial work raises the next question—how often does the robot fail, and what happens when it does?

Ma places deployed generalist manipulation models at roughly 80–90% success. That leaves frequent interruptions in a repetitive job. To see the compounding effect, under an illustrative assumption of independent attempts with constant success probability (p), ten consecutive successes have probability (p^{10}): about 10.7% at 80% success and 34.9% at 90%. The recording's captioned figure of less than 0.1% for ten repetitions is inconsistent with that calculation, so it cannot support the numerical comparison. The underlying concern remains: reliable individual attempts and long uninterrupted runs are different requirements.

Specialized systems offer another path: build a robot and learning pipeline for one operation, such as pick and place. The tradeoff is reuse. A pipeline built around one task cannot necessarily adapt quickly to an arbitrary new one. Dyna's goal is to retain the generalist model's range while reaching the reliability a customer needs for a specific workflow.

Restaurant napkin folding gives that goal a concrete test. Employees fold napkins individually in the back of the restaurant, and customers Dyna spoke with wanted to automate the work. The task is repetitive, commercially useful, and difficult enough to expose weaknesses in manipulation. Narrowing the research to this workflow makes the desired result tangible: keep producing napkins the restaurant will accept.

Dyna-1, a generalist robot foundation model fine-tuned for napkin folding, achieved a reported 99.4% success rate over a 24-hour run. Ma also presents four distinct 24-hour office trials. In the time-lapse, daylight arrives around 7 a.m. and changes the lighting while the robot keeps folding. These are speaker-reported results; the supplied material gives neither evaluation sample sizes nor independent verification, so the percentage should remain attached to this napkin-folding evaluation rather than become a reliability claim for every task or site.

The operation has three nominal steps: pick exactly one napkin from the stack, fold it, and place it in a bin. The first step already creates trouble. Parallel-jaw grippers may pull out extra napkins, leaving the robot with a scene that differs from a clean demonstration. The fold also has to meet a customer's quality standard. Ma contrasts an acceptable grade-five fold with an unacceptable grade-three fold, separated by roughly one inch in the position of the second fold seam. Moving cloth into a folded shape is only part of success; the final geometry matters.

Dyna's initial pre-training and post-training recipe reached roughly 80% success on this task. The recurring problem appeared after a mistake: the robot entered an unfamiliar state, got stuck, and could not recover. More successful demonstrations would teach the normal path, but the missing behavior was what to do after leaving that path. Reliability therefore required training on the situations the robot created through its own errors.

Dyna adds a reward model that watches robot video and estimates progress toward completing a task. In the examples, progress moves from zero toward one as the operation finishes. A successful napkin fold produces a roughly rising estimate; mistakes produce dips. During the ninth napkin in one autonomous run, the estimate stops rising monotonically as the robot makes an error. The useful signal is the change in progress, which helps identify where the operation went wrong.

Once the policy succeeds around 90% of the time, having a person watch every attempt is an inefficient way to find the remaining failures. The reward model runs in the background and directs researchers or operators to the video cases that need attention. Humans then collect targeted demonstrations of recovery from those states, and Dyna fine-tunes the policy again. This is a human-in-the-loop active-learning cycle: autonomous operation finds the gaps, progress scoring helps locate them, and human demonstrations supply the missing actions.

Where does the progress signal change the training process? The diagram follows the loop from a napkin-folding attempt back to an improved policy. Its key relationship is the handoff from detecting trouble to collecting recovery data: the reward model helps choose what to teach, while the policy learns how to act. Ma reports that repeated cycles produced a model able to recover from a wider range of errors.

Return to the single-napkin pickup. In a recovery highlight, the robot accidentally takes more than one napkin. Instead of remaining stuck with the extra cloth, it separates the napkins and continues making progress. The observable change is from a failed selection to a state in which folding can proceed. Targeted recovery training addresses precisely this gap: the policy needs useful actions for the aftermath of a bad grasp, as well as actions for the clean pickup.

Cloth makes exhaustive recovery coverage impractical. A napkin can deform into an enormous range of configurations, so the team cannot demonstrate every possible error. Ma reports that repeated active learning produced recovery behavior that generalized to new mistakes. The strongest example is a robot pulling over the entire napkin stack, then using pulling and stretching movements to recover and keep working. Continuous operation depends on handling the mess the robot creates, rather than assuming the table always remains in its expected state.

The commercial examples move the same recipe into customer workflows. Dyna deploys napkin-folding robots in restaurant back offices and fine-tunes a model for towel folding at a Sacramento laundromat. In the laundromat example, a worker repeatedly replenishes the bin while the robot folds stacks of towels. The automation handles the folding operation within a workflow that still includes human material handling.

The restaurant and laundromat examples come with an important qualification: Dyna collected data at the customer sites used for those deployments. Mastering a task in one location does not automatically establish that the model will perform it somewhere unfamiliar. Collecting and adapting at every new site adds work to expansion, so the next objective is to preserve task performance without additional site data.

Ma describes a diverse-task data recipe intended to enable that transfer, then presents T-shirt folding at CoRL 2025 in Korea as the example. The robot arrived at an exhibition booth and began folding different T-shirts without additional site data. It faced attendees, including people who deliberately covered its camera with a shirt, and reportedly continued folding over three days at the conference. This tests environmental transfer and resilience to disruption; it does not establish the same 99.4% figure reported for napkin folding.

A Red Bull partnership extends the examples to opening cans at live events. Ma describes repeated operation at a music festival despite changing background lighting, but the demonstration video did not play properly during the presentation. This example rests on his account. Together, the deployments point toward the desired product experience: bring a robot to a site and have it start useful work, while retaining the ability to perform a chosen commercial task well.

The audience questions clarify what Dyna is building now and what remains an ambition:

• Developer tools: Dyna is building its own hardware stack and is not currently working on a developer kit. A kit may become part of the roadmap later.
• Education: Teaching children to train robots is an interesting future direction for Ma, but Dyna has not explored that use case.
• Consumer robots: The long-term goal includes a model and hardware platform usable by anyone, including households. The current go-to-market focus is enterprise customers.

Enterprise customers let Dyna deploy and iterate at the site. A consumer product would need more polish, tolerate fewer mistakes, and address privacy and safety concerns that the company is not taking on at this stage. Ma treats commercial environments as a useful testing ground for the immediate bottleneck: getting the AI and hardware to work well together. The sequencing follows the research cycle introduced at the start—choose a workable deployment, learn from it, and improve the system.

Voice commands could fit above the manipulation stack. Ma sketches an approach in which onboard speech-to-text converts a person's speech into text, then the reasoning and world models interpret the instruction and produce actions. He offers this as a possible integration rather than a demonstrated voice system. Recognizing the command would still leave the harder capability problem: Dyna's models and other robot foundation models are not ready to execute arbitrary instructions. A robot that anyone can teach or steer through language remains the goal; low-level manipulation is still a limiting step.

The final question asks how intelligence should be divided between general pre-training, task-specific training, and adaptation at a deployment site. Ma adds a consequential explanation for the recovery demonstrations: some of the behavior came from interpolating recovery experience gathered on other tasks. A model exposed to thousands of tasks can draw on physical knowledge that a model specialized from the outset would miss.

That gives the three training stages distinct purposes. General pre-training builds semantic and physical understanding, including experience with recovery. Post-training adapts that foundation to the chosen job. Active learning then concentrates demonstrations on the mistakes that remain during operation. For napkin folding, the resulting policy needs both a precise routine and enough transferable physical knowledge to find a way forward when the routine goes wrong. Dyna's intended reusable product is this recipe across tasks, rather than a separate learning pipeline for each one.