Hardware Projects

What it is

B1 is a self-balancing robot built as a real-world testbed for the Dojo platform. It was designed to answer a simple question: can neural network policies trained in simulation survive contact with reality without extensive hand-tuning?

The chassis is an old wine box my wife happened to have lying around. Using a repurposed 25-year-old wooden case as a robot platform introduced its own set of challenges, but also made it possible to iterate quickly without waiting on custom hardware.

The B1 can be driven through teleoperation and maintains dynamic balance using its learned policy on physical hardware.

Why it matters

Sim-to-real transfer is the central unsolved problem in embodied AI. Most controllers that perform well in simulation fail when exposed to sensor noise, latency, friction differences, and imperfect modeling.

B1 was built outside of a controlled lab environment using commercially available components and consumer-grade compute. Its purpose is not to be a product, but to serve as physical evidence that the Dojo training approach can produce behaviors that remain stable under real-world conditions.

What we learned

Developing B1 required treating sensors, timing, and deployment constraints as first-class parts of the learning problem rather than post-processing details. Policies must operate only on information available to the physical system, tolerate communication delays, and remain stable despite modeling error.

The project confirmed that robust performance depends less on perfect simulation and more on alignment between training assumptions and deployment reality.