A robot holding a knife cannot stop and think for three seconds. Jason Ma, co-founder and chief scientist of Dyna Robotics, used that image to explain why a company known for its AI models also designs and builds the machine those models run on. People will put up with a chatbot that takes a couple of seconds to answer, he said. A robot that holds a knife and does nothing for three seconds is "going to freak people out."
Ma made the case in an on-site interview at Dyna's office for the Latent Space podcast, published October 13, 2026. The episode followed Dyna-2, the company's model trained on more than a million hours of video, and Dyna-2.1, a physical agent built around the company's own robot. His answer to why Dyna doesn't simply supply software for other companies' robots covered speed, which parts to make, magnets, and why the robot rolls on wheels.
Physics doesn't wait
Most of the argument comes down to timing. Ma said dynamic tasks leave no room for delay. His example was a robot that holds a tennis racket and juggles a ball: "physics gravity doesn't wait for you." The fastest way to get a model's answer, he said, is to serve it on the robot itself or on the customer's premises. Calling a model over a network adds latency, and there is "no guarantee" of how long it will take.
That sets a limit on size. The model a robot runs has to be small enough to serve locally. Ma said this doesn't mean the model a company trains has to be small. He pointed to a method from language models called distillation, in which a large "teacher" model is used to make a smaller one smart. "You train the biggest model then you distill," he said. Robotics is "getting there," he added, and he expects the field to converge on that approach.
Dyna also builds its own inference stack, the software that runs a trained model and turns inputs into actions. Because latency matters so much, Ma said, the company optimizes that software for the kind of model it deploys, and "there's a lot of work in there."
What to make and what to buy
The conversation then moved down the hardware stack, from industrial design to the motors. Ma described a simple rule. If a component's inputs and outputs are clearly defined and a very good, mature solution already exists, there is no point building it in-house. The black circles on the robot are lidars, laser sensors that help it detect obstacles and navigate. The technology has been used for decades, including on self-driving cars, so Dyna doesn't build lidar from scratch. The same goes for some cameras.
Dyna wants to own anything it considers crucial to the whole system's performance. Actuators, the motors that move the joints, are on that list, and Ma said the company is working on its own. Actuator design affects how well the robot can be modeled, for example in simulation, so Dyna wants actuators that are good and also easily modelable.
The robot's proportions follow the same logic. Ma said Dyna measures them against real people because it pretrains on human data and wants the best possible transfer. Dyna's Dyna-2.1 page describes its robot, Taku, the same way: human-like proportions support motion transfer, and a shared description of wrist, elbow, chest and footprint poses lets human recordings supply training targets. "It's hardware that carries intelligence," Ma said, and the hardware has to serve that intelligence.
Magnets and a long supply chain
The discussion turned to a less visible bottleneck: magnets for actuators, which the United States doesn't make. "No, we don't," Ma agreed, adding that they depend on rare-earth materials. He said bringing manufacturing back to the US is much harder for robots than it seems. In his words, no company in the US is "truly building everything from scratch" there. Even companies that say they assemble robots in the US source raw materials from elsewhere.
Building the full robot showed Dyna "just how long the supply chain is," Ma said. He thinks the field's conversation centers on foundation models and data while the difficulty of building good hardware is "very under emphasized." Asked whether Dyna will build in the US or rely on a global supply chain, he said the company is exploring several options. Global sourcing brings geopolitical issues and shipping delays, and getting US production going from zero will take time. Dyna wants an approach that lets it build more in the US while still using global suppliers.
The discussion noted that some companies choose to build only the model and deploy it on any platform, such as Tesla's Optimus. Ma said his earlier work at Google DeepMind and NVIDIA, mainly building models, showed him "the upper bound of the system when you don't control the hardware." Robots exist to perform physical tasks for customers, he argued, so a company that wants to do those tasks well needs to be able to optimize everything. A May 2026 essay by Dyna's York Yang makes a related argument: models, hardware and deployment infrastructure should be developed together while the interfaces between them are still unsettled.
A semi-humanoid on wheels
Ma calls the result a "semi-humanoid." From the waist up it is humanlike: shoulders, arms and four degrees of freedom (independent directions of movement) at the waist, so it can make the upper-body motions a person makes. Below the waist it has wheels. According to the Dyna-2.1 page, Taku has four steerable wheels and two arms with seven degrees of freedom each.
Ma described the choice as a trade-off between practicality and performance. With legs, he said, energy consumption is much higher, the robot is less stable, and safety is extremely hard to guarantee. Today a legged robot has "no guarantee" of staying upright the whole time, he said, pointing to robot dancing videos. The conversation also turned to demos in which a robot is boxed in and pushed over and stays up, which show good balance but not necessarily readiness for production.
What a pulled muscle taught him
One distinctive feature of the robot is a torso that can lean, which matters in the laundry demo, where it has to reach into machines. An earlier prototype only moved up and down. Ma explained the change with a pulled muscle he got while exercising a couple of years ago. For days afterward, he said, he realized how many muscles and degrees of freedom people use daily and take for granted, and how awkward natural tasks become when some are disabled.
A torso that only goes up and down, he said, misses how much people use their waist and back. Even reaching for a bottle, he does not move only his arms, and holding his upper body still on purpose would feel strange. Since humans move their torsos all the time, human training data is full of that movement. A robot body that cannot reproduce it gets less from the transfer and becomes "less dexterous and it's just more awkward."
Iterating toward the human form
Ma does not reject full humanoids. The conversation took up the argument that the world is made for humans, so the robot should be humanoid. Ma agreed that for a general-purpose robot, the human form "will be the best one." His disagreement is about how to get there. Hardware decisions are costly and slow to revise, and it is hard to know the right choices in advance, so he favors an incremental approach over going straight to the final form. He said it is "public news" that Optimus has gone through many iterations and scrapped designs.
Putting semi-humanoids in front of real customers and deploying new models onto them creates a feedback loop, he said, that a team spending five or ten years on a warehouse humanoid would not get. The iterative approach, Ma said, will get Dyna "closer to the final robot faster." According to the Dyna-2.1 page, the company's next milestone is deploying the new system at customer sites.