— as seen on
— intro
Successor to the 3D printed zeroth bot I initially created. After seeing massive demand for a low-cost humanoid platform for RL development, we turned it into a consumer developer product: more robust, cost-effective, and scalable.
I led the hardware across two distinct iterations, focused on making this the best platform for learning and deploying RL control. Studied limb morphologies and kinematics in sim, stress-tested through hackathons and stanford's CS234, and refined the sim2real pipeline for accessibility.
— outcome
A compact 20-DOF humanoid platform with a sub-$2k BOM, open-source SDK, and sim2real pipeline. 20k+ waitlist signups, 50+ units deployed. Hosted hackathons with 150–200 attendees, where teams built features that expanded the SDK. Used in stanford's CS234 as a teaching platform for RL.
Partnered with joyln tech to manufacture and distribute. Launched at CES, covered by CNET and PCMAG.
— devlog
After the original 3D printed zeroth bot gained traction online, we partnered with joyln tech and designed the consumer version in a two-week sprint. We received the first batch of around 15 prototypes and immediately put them to work: ran a hackathon and deployed units to stanford's CS234 course. 70% of the hackathon robots failed. Between both, we mapped every failure mode and gap in the platform.
We spent weeks improving the v2. Structural fixes, actuator characterization, SDK improvements. It got better but we kept hitting fundamental limitations in the original architecture.
Everything we learned from those deployments gave us a clear understanding of what this platform actually needed to be. We set a new design philosophy and rebuilt from scratch. The v3 came back with 6-DOF legs for 20 total DOF, raspberry pi compute, shared subassemblies, removable electronics for debugging, and silicone covers on contact points.
The result was a huge step up. Far more reliable policy deployment and a real dent in closing the sim2real gap for these robots.
— reflection
I wish I got to see this one through. The company refocused on the k-bot and sometimes you have to align with the team and move together, even when it means shelving something you care about.
One thing I'd change: open-sourcing earlier and more broadly. We got so much value from hackathons, but bringing the community into the development process itself, not just events, would have made this platform so much stronger.
This project taught me a lot. Hopefully one day I get to build something like it again and take it all the way.