The Challenge
Bambu Lab reached out over the summer and challenged me to build something cool with their brand-new P2S 3D printer. The catch: I was leaving for a robotics internship in exactly seven days, and if I couldn't get the robot walking before I left, the printer had to go back.
The safe choice would have been something I knew I could finish in time, like a wheeled robot or a small quadruped. Instead I went with a bipedal robot. Eventually I want to build a full humanoid, and starting with just two legs felt like the right first step toward that.
Design and Hardware
I found inspiration in Fauna Robotics' Walker, a six-degree-of-freedom bipedal robot: three degrees of freedom per leg, two hip motors and one knee motor, six motors total. Simpler and cheaper than a full humanoid, while still giving me two legs that could actually walk.
Stompy is built around six Robstride O2 motors (17 Nm of torque, about 400 grams each, dual encoders so I only have to zero them once), an NVIDIA Jetson Orin Nano as the brain, an IMU for sensing, and a Robstride CAN hub so all six motors share a single communication line instead of six separate connections. The entire structure is 3D printed, including a handle on top so I can actually carry it around.
Teaching It to Walk
I could have hand-coded walking trajectories, but that takes time I didn't have. Instead, I used reinforcement learning: thousands of simulated versions of the robot try to move forward while staying upright, get rewarded or penalized based on how well they do, and gradually a policy emerges that decides what the motors should do next based on the robot's current state.
After fixing the foot geometry and retraining, the policy transferred from simulation straight to the real hardware on the Jetson. By day three, Stompy was walking.
Taking It Outside
Once it walked indoors, the real test was taking it outside onto grass, concrete, and uneven ground. It handled all three. I also pushed it off balance mid-stride to see whether the learned policy could recover, and it did, repeatedly, which says more about what reinforcement learning generalizes to than anything I could have hand-programmed in a week.
Tech Stack Used
Key Features
- Built and Walking in Under a Week: From an empty CAD file to outdoor walking and push recovery.
- Fully 3D Printed: Structure printed on Bambu Lab's P2S, roughly a thousand dollars in parts.
- Sim-to-Real RL: Trained entirely in simulation, deployed straight to the Jetson.