The Bitter Lesson of Robotics: Why Scaling Wins — Jason Ma, Dyna Robotics

Summary

• Dyna Robotics claims a proof point many robotics startups lack: validated multi-site commercial scale, not just pilots. Its robots fold thousands of napkins per day across many Din Tai Fung locations, and Dyna also folds laundry for Best Western. Jason says Din Tai Fung signed a much bigger contract and is considering food cutting as another use case; the transcript does not establish that cutting is already part of the contract. Jason’s formulation is that “what’s really hard is having customers validate that your robot creates enough value that they want to deploy it across many of their sites”—something that, “to the best of my knowledge,” has not happened for similar general-purpose robot companies.
• The technical bet is Dyna-2, which Jason calls the first “world action model” trained on more than 1 million hours of data. It predicts future video or a representation of the future world and uses that to guide actions: “analogous to chain of thought—the models are reasoning in visual and physics space, then outputting actions.” Under the hood, it’s a diffusion transformer using flow matching. Vibhu describes an action space on the order of 100 dimensions; the model predicts a vector of at least 50 dimensions at 10–20 times a second.
• Task adaptation takes tens of minutes to a couple of hours of robot data collection. The chopping task took 45 minutes; opening the fridge is an engineer-onboarding interview task. Vibhu says the new robot learned tasks quickly despite having little of its own data, crediting the decision to pretrain on human data before the hardware was ready. The robot’s proportions were also chosen with human proportions in mind for better transfer.
• Dyna 2.1 is presented as hardware with a “physical agentic harness” for longer workflows. Vibhu says a complete laundry loop can take about an hour, and argues that robotics’ value lies in workflows that run unsupervised for hours. The demo footage shows parts of the laundry workflow; the narration says the entire workflow would be released later. The company’s demo narration describes the ambition as a robot becoming “a whole employee at a workplace.”
• Dyna’s approach is full-stack: it builds its own inference stack and is working toward owning its actuators, alongside its semi-humanoid hardware with a wheeled base and 4-DOF waist. Vibhu says legs bring trade-offs in energy use, stability, and safety, and argues that iterating a semi-humanoid with real customers is preferable to spending five or ten years trying to build a final-form warehouse humanoid. On supply chains, Jason raises magnets as a bottleneck, saying only two countries make them; Vibhu refers to rare-earth materials, while Jason says he doesn’t know the specific chemistry or material. Vibhu adds that no U.S. company is truly building everything from scratch in the U.S.
• Data is a “non-stationary problem,” and synthetic data may be another scaling axis. Jason says one company’s collection is nowhere near enough—“it’s kind of like trying to build an internet by just having Reddit or just having Facebook.” Dyna is soliciting egocentric, robot, video, and other useful data at [email protected] , while also exploring custom task data. Jason sees frontier models’ real-to-sim abilities at a tipping point, and says Dyna’s world action model could generate video for its own training-data flywheel.
• Vibhu identifies a possible weakness in vision-only world models: missing modalities. He says a vision-plus-tactile model that is 100× more data-efficient could be an alternative, since tactile information may be denser than highly redundant pixels. He argues that robotics companies may be better positioned than video-generation companies to add sensing modalities. Hardware is another bottleneck: dexterous hands can overheat quickly.
• Deployment targets are tens of robots already and hundreds next year, with millions of robots as the longer-term vision. Jason warns that “there’s a whole graveyard of robotics companies” with shiny demos, and says teams focused on only part of the problem risk “fog of war”—building something impressive that isn’t what customers need.

Deep dive

1. Dyna-2 is a world action model: predict the future, then act

• Jason defines it as a model that predicts “the future state of the world” as video, or a representation of it, and uses that to produce executable robot actions. He compares this to how humans “implicitly think about how the world will evolve based on our actions.” He says the result is a model that is more dexterous, robust, and able to generalize to unfamiliar environments.
• Inputs include camera video from the past several seconds, the robot’s joint-encoder readings, and language instructions. Outputs include actions and future video or a representation of it—“the models are reasoning in visual and physics space, then outputting actions,” in Jason’s chain-of-thought analogy.
• Dyna-2 is described as the first world action model trained on more than 1 million hours of data. Jason says the million hours are “not enough” and “just the beginning.”

2. Why a diffusion transformer, and why flow matching fits robotics

• Jason describes Dyna-2’s architecture as a diffusion transformer. Transformers can ingest a lot of data; diffusion iteratively refines noise into a precise output. For robotics, the generated actions need to be smooth and precise, much as generated images need fine-grained features.
• Flow matching is a simpler-to-tune variant of diffusion. Jason says it works well for action generation because the robot can reevaluate in real time; the host contrasts that with image generation, which is more of a one-shot process.
• Vibhu puts the robot’s action space at around 100 dimensions. He also says the model predicts a vector of at least 50 dimensions at 10–20 times a second. In the laundry demo, a speaker says one model controls every joint, without a separation between classical planning and acting that would make the motion less smooth.

3. One general base model, then minutes of fine-tuning per task

• The recipe is to train a general model on as much data as possible, then adapt it. Vibhu says a new task can take a couple of hours of robot data collection, or tens of minutes for some tasks. Jason says chopping took 45 minutes; the fridge-opening task is an engineer-onboarding interview task. Jason’s human analogy: if you or he went to work at the Rosewood Hotel, they would need training to make the bed the way the hotel wants.
• The evaluation loop is to train, deploy in the real world, and observe. Vibhu says some models work out of the box; when one is “70% or 80% there,” the team identifies shortcomings, collects more data, and iterates. He compares this to robotics’ version of reinforcement learning from human feedback.
• Vibhu says the new semi-humanoid had comparatively little of its own data but learned tasks quickly. He credits the decision to pretrain on human data before the hardware was ready. Separately, he says the robot’s proportions were measured against real people to improve transfer.

4. Data: no internet for robotics, and quality is a moving target

• Jason says robotics lacks the internet-scale data available to language models. An individual company’s collection is “nowhere near sufficient”—it is like trying to build an internet from only Reddit or Facebook.
• His proposed recipe combines egocentric human data, which covers diverse physical activities and gives the model a prior on how the world evolves around an embodied actor, with robot data to bridge the gap between human and robot actions. Robot data is more expensive and harder to collect. Jason calls data quality “a non-stationary problem”: what counts as high-quality data changes as the model develops.
• Jason invites vendors with egocentric, robot, video, or other potentially useful data to contact [email protected] . He also says task-specific data, like the work Mercor is doing for language models, could be valuable for robotics. The transcript describes data acquisition and an open call; it does not establish that Dyna is already buying these data sources.

5. Synthetic data may be a new scaling axis—and the model could feed itself

• Jason says his PhD-era Eureka project used GPT-4 to generate reinforcement-learning reward functions for simulation, with a robot-spinning-pencils demo. At the time, models had limited vision and 3D understanding. He now sees frontier models’ vision capabilities—including real-to-sim CAD modeling and 3D reconstruction—as a potential tipping point for robotics.
• Dyna is exploring using frontier models to generate tailored synthetic environments and data. Jason also says Dyna’s world action model can generate video that could serve as its own training data, creating a potential flywheel.
• Jason describes Dyna as AI-native. When GPT-6 came out, the company held a hackathon that week, and some resulting features were already in production. He says the ability to construct virtual environments and turn them into simulations was a particularly exciting capability.

6. Full-stack hardware: own what carries intelligence, buy what’s mature

• Vibhu’s own-versus-buy rule: buy components with clear inputs and outputs and mature solutions, such as LiDAR and some cameras; build components crucial to overall system performance. Dyna is working toward owning its actuators because their quality and design affect how well the robot can be modeled in simulation. Vibhu says the robot’s proportions were measured against real people because Dyna is pretraining on human data, which helps transfer.
• The robot is a semi-humanoid, with a humanoid-like upper body, 4 degrees of freedom in the waist, and wheels below. Vibhu says legs trade off against wheels on energy consumption, stability, and safety. He argues that torso movement matters for everyday tasks and for transferring human data: a robot that cannot move its torso would be less capable and more awkward.
• Jason raises magnets as a supply-chain bottleneck, saying only two countries make them; Vibhu refers to rare-earth materials, and Jason says he does not know the specific chemistry or material. Vibhu says there is no U.S. company truly building everything from scratch in the U.S., and that Dyna is exploring options that combine more U.S. manufacturing with global sourcing despite geopolitical and shipping risks.

7. Din Tai Fung napkins: the difference between a pilot and a business

• Jason says Dyna’s robots fold thousands of napkins per day across many Din Tai Fung locations in the U.S., its biggest scaled customer. He distinguishes a pilot from multi-site adoption: “what’s really hard is having customers validate that your robot creates enough value that they want to deploy it across many of their sites.” He says that, to the best of his knowledge, this has not really happened for similar general-purpose robot companies.
• The napkin standard is precise: the second fold seam should be a couple of centimeters below the pocket so chopsticks can fit. Asked why a general-purpose robot rather than a specialized machine, Jason says deformable objects have “infinitely many possible configurations,” and a general-purpose model adapted to the task can reach production quality faster.
• Jason says Din Tai Fung signed a much bigger contract and is considering food cutting as another use case. The transcript presents this as a use case they are thinking about, not as a confirmed contracted deployment. Jason also says locations in LA, New York, and Seattle fold napkins differently, while the model’s shared weights can be updated over the air across robots. The host calls the result human parity, and Jason agrees that the robot does not get tired.

8. Longer workflows need an agentic harness, and inference must be fast

• Vibhu says real physical work needs high-level reasoning as well as low-level actions. For the laundry workflow, the robot must decide to do another task while the washer runs, then move the towels to the dryer when it finishes. Learning such long-horizon reasoning solely with a low-level action model would be inefficient, so Dyna uses an agentic system around its world action model.
• His comparison is to language models progressing from autocomplete to agents that perform end-to-end tasks. Vibhu argues that robotics’ value lies in workflows that run unsupervised for hours, and that robotics is a systems problem involving the agent, harness, model, and hardware.
• Latency constrains deployment: the robot cannot hold a knife motionless for several seconds, and dynamic physical tasks cannot wait for network calls. Vibhu says deployed models need to be small enough to run locally; one path is to train a large model and distill it. He says Dyna builds its own inference stack.
• The Dyna 2.1 narration presents the hardware and harness as supporting end-to-end workflows. The footage described in the transcript includes transferring towels from a washer to a dryer, handling a dropped towel, operating the dryer, and retrieving a drink using a reasoning layer and an action layer. The narration says the laundry footage is only one part of the whole workflow and that the complete workflow would be released later. Vibhu says a complete laundry loop takes about an hour.

9. The landscape—and how the world-action-model approach could be wrong

• Jason argues that predicting the raw camera inputs avoids some of the bias involved in predicting an abstract representation of the world. He says robotics needs to predict as much of the world as possible, rather than encode assumptions about which parts matter.
• When swyx asks what could be a better alternative, Vibhu points out a caveat: video may miss important aspects of the world, including sound and touch. A vision-plus-tactile world model that is 100× more data-efficient is one possibility; tactile information may be denser than highly redundant visual pixels.
• Vibhu says that if multimodal sensing is what ultimately solves robotics, robotics companies may be better positioned than video-generation companies to develop it, because understanding and combining sensors is complicated. In a separate exchange, swyx suggests that one might “vibe-code” around hardware knowledge; Vibhu responds that it is “just not in the distribution.”
• Jason says there are many approaches, including scaling human data, scaling robot data, and using intermediate data-collection devices, and that it is early to tell which will win. He warns that teams working on only one part of the problem may face “fog of war,” building something impressive that customers do not need; he points to a “graveyard of robotics companies” with shiny robots.

10. Bitter lessons for robotics, and why iteration matters

• Jason’s “bitter lesson” is that robotics will scale like other machine-learning domains, but finding the right data is harder because sensor choices determine what data is available. He argues for simpler algorithms and scaling the right data rather than relying on specific priors, such as object boxes.
• Jason says a human form factor may ultimately be best for a general-purpose robot, but recommends an iterative approach because hardware decisions are expensive and slow to change. He cites Optimus as having gone through multiple iterations and scrapped designs, and argues that deploying semi-humanoid robots with real customers can provide a faster feedback loop than spending five or ten years on a warehouse humanoid.
• Vibhu identifies hardware stability and data as bottlenecks to broader generality. Dexterous hands can overheat; hardware limits also slow evaluation and research. He says missing sensors such as sound and touch could leave critical task information uncaptured, even with millions of hours of data. For workflows where high-fidelity data can be collected today, he says there is not necessarily a fundamental bottleneck.

11. From AlphaGo to Din Tai Fung: the founder arc and roadmap

• Vibhu says AlphaGo sparked his interest in AI while he was a high-school senior and lifelong Go player. At Penn, he worked on reinforcement learning for robotics at the GRASP Lab, including connecting reinforcement learning with human data. At NVIDIA, he worked on pen-spinning and the Eureka project with Jim Fan and Yuke Zhu; later, at Google DeepMind, he explored using Gemini for robotics. He says he wanted to move from lab evaluations to work that could be used in the real world.
• Vibhu says he named Dyna partly after Richard Sutton’s paper “Dyna: An Integrated Architecture for Learning, Planning, and Reacting,” which concerns an agent that improves in a world model. The other inspiration was an Ultraman character called Dyna in the year he was born.
• Jason says Dyna has deployed tens of robots and aims to deploy hundreds next year. The company’s longer-term mission, stated in the closing narration, is to reach thousands of customer sites and eventually millions of robots. A company speaker says roughly 30%–40% of the company is on the hardware side and around 10 people work on models.
• Vibhu describes a product-driven culture: identify problems such as generalization, whole-body control, fast adaptation, language following, and dexterity, then let engineers create tasks that stress-test them. The drink-grabbing task tests whether the robot follows instructions.
• The closing call to action is for people motivated to get robots deployed to join the company. The speaker says Dyna is hiring in the Bay Area for now.