This is why we started Robai and what we set out to build. I am writing it down early, before later descriptions smooth away the details. A company's true origin is usually more specific than its eventual pitch. That specificity keeps it honest.
The Problem We Kept Seeing
Seokjae and I spent years building control systems for industrial robots: motion controllers, embedded autonomy, and perception pipelines for factory arms. We understood these systems from the inside. Yet manufacturers kept encountering the same pattern. A robot would be installed, configured, and running a production task. Then that task would change, and all the value already created would sit idle while the plant waited for an integrator to return and reprogram the robot.
The delay was not caused by integrators being scarce or incompetent. It was structural. Giving a robot a new task is specialist work. It calls for someone familiar with the robot's programming language, controller interface, safety configuration, and the details of the new task. Small or mid-size manufacturers almost never have that person on staff. The integrator must be scheduled, travel to the plant, and be paid, after which the process must be validated before the line can run again.
Over multiple years, we saw this cycle at multiple plants: different robot arms, products, and countries, but the same result. Each task change cost weeks. The hardware was fine, and the robot arm could perform the new task. Configuration, not capability, was the bottleneck.
Why Learning from Demonstration Made Sense
The research direction we kept returning to was imitation learning: training a robot to perform a task by watching a demonstration instead of programming each step. Academic work had been building for several years. Research on behavior cloning, GAIL, and later diffusion-based policy learning showed that robots could acquire surprisingly capable motor behaviors from relatively small numbers of demonstrations in controlled settings.
The difference between those academic results and production manufacturing was real. Academic demonstrations took place in clean labs with consistent lighting, controlled part placement, and tasks designed to be learnable. Factory floors bring variable lighting, inconsistently oriented parts, vibration noise in sensor data, and robots expected to run eight to twelve hours a day without falling over statistically. Moving from lab to factory did not happen automatically.
Still, the central idea made sense. Reprogramming takes time because programming translates task knowledge into code, and an expert is needed to perform that translation. If the person who already performs the task can demonstrate it, that bottleneck goes away. The person who knows the assembly or pick-and-place task shows the robot how it is done. The robot learns the policy, then runs the task.
We believed the gap between academic results and production deployment could be closed. We also believed it called for people who understood both robot control systems and manufacturing production constraints. That combination was less common in robotics research, which generally focused on academic benchmarks, and manufacturing automation, which generally focused on existing programmed solutions.
Starting in Seoul, 2023
Robai began in Seoul in 2023. The location was a deliberate choice. Korea has a dense concentration of electronics contract manufacturers and automotive parts suppliers. They operate the variable, high-mix production environments most affected by reprogramming. They were the natural first customers for what we were building.
Gangnam-gu, where we established the initial office and lab, is in the middle of Korea's technology and venture capital corridor. Being able to reach early manufacturing partners for pilot testing mattered practically. We needed to stay close enough to plant floors to test under real production conditions rather than only in simulation.
The initial team had three people: the two co-founders and Larisa, who joined early to build the control layer between AI policy outputs and physical robot controllers. Making a learned policy reliably drive a FANUC or UR arm in production proved harder than training the policy itself. Industrial robot controllers have quirks. Communication latency, joint speed limits, and fault recovery behaviors all need to shape how policy outputs become motor commands.
The First Pilots
Our first pilot engagements involved small manufacturers with cobots installed but frustrated by changeover costs. We were clear about the system's strengths and limits. Pick-and-place tasks with moderate part variation worked. Tasks needing positional tolerances under 1mm were more difficult. Long assembly sequences with many steps needed more demonstrations than shorter tasks before reaching production-stable behavior.
Early pilot feedback substantially shaped the product. The request we heard most consistently concerned visibility into policy confidence: not only whether the robot was executing the task, but whether it remained within its reliable range or was extrapolating in a way likely to cause failures. That input went directly into the telemetry layer created for steady-state production monitoring.
What We Are Not Trying to Be
We are not trying to replace all robot programming. In high-volume production with stable tasks, scripted control has a decades-long reliability record, and that record was earned. We are not trying to persuade an automotive body shop that has run the same weld program for three years to adopt demonstration learning. Scripted control has the right performance envelope for that use case.
We are building for the other side of manufacturing: variable production lines, contract manufacturers, and shops running 15 different configurations per quarter on the same robot cell. For these operations, the programming model is already broken. The issue is not whether to fix it, but how. We believe demonstration learning is the right answer, and we are building the production system to prove it.
That is the reason we started Robai. Anyone spending time with robots on plant floors could see the specific problem. We had the technical background needed to address it. The manufacturers who needed a solution were close enough to work with directly. Together, those three conditions were enough to begin.