System design · v1 target
Onboard architecture
The airframe keeps a proven quadrotor control stack: an off-the-shelf flight controller (PX4) runs the inner attitude loops, and a companion computer runs the neural-net policy that sends setpoints over uXRCE-DDS at 100–200 Hz. The leg is passive — spring and damper, no actuators; the props do the active work.
Reading the stack
- Inner loop stays classical. PX4 on an H7 board (Kakute H7 mini class) keeps attitude and rate control;
COM_DISARM_LANDand ELRS kill-switch give two independent safety paths. - The policy is the new part. A PPO policy trained in Isaac Lab runs on the Orin and emits position/velocity setpoints at 100–200 Hz through uXRCE-DDS — never raw motor commands.
- The leg is dumb on purpose. Spring + damper with a tip mass, like the Science Robotics paper. Actively-actuated legs are a later iteration.
- Voice only lands when quiet. The XVF3800 array localizes speech, but at ~20 W hover props are too loud: the pet listens and speaks during rest phases between hops.
- Training stays off-board. All learning happens on the 3090 workstation in simulation; the airframe only receives exported policies (ONNX/TensorRT).