Pet robot · wayfinder effort
Hopcopter.pet
A 3D-printed, monopedal hopping quadcopter that lives at home: it flies where flying is fast, hops where hopping is cheap, survives big drops, and answers to its name. Everything — design files, firmware, simulation, training, journal — lives in pet_project, steered by one wayfinder map.
Decisions locked
- Clone the Science Robotics monopedal hopping quad, then scale to 0.8–1.0 kg / 5″ class — paper's 35 g Crazyflie is not flyable with a Jetson aboard.
- Classical inner loop, learned setpoints: PX4 on H7, PPO policy on Jetson Orin Nano Super over uXRCE-DDS at 100–200 Hz.
- Sim: Isaac Lab (Newton) primary, MuJoCo/MJWarp fallback on one shared MJCF model.
- Leg is passive (spring + damper + tip mass); props do the active work.
- Vision: one forward-down RGB-D (D435i). Voice: XVF3800, used at rest — too loud next to props.
- Hop vs fly is an energy decision — the paper's headline: hopping stretched endurance 50 min vs 6 min flying.
In flight now
- #21 — docs site shipped: Astro pages + first diagrams (Cloudflare connect pending).
- #8 — bill of materials & sourcing; gate: all-up weight ≤ 1.0 kg.
- #11 — simulator: V1 quadcopter model + hover policy spike on both 3090s.
Hardware on hand
- Bambu Lab P1S (PLA · PETG · TPU)
- 2× RTX 3090 24 GB training box
- Indoor test space, 18 ft ceiling
The ladder
- Hover stable
- Hop in place (passive leg)
- Survive big drops, self-recover
- Traverse terrain
- See & hear onboard
- Follow · explore · come back — as a pet
Why this is tractable
The control problem was proven on real hardware by Bai et al., Science Robotics 9, eadi8912 (2024) — hopping quadruped-on-a-stick aerobatics with only classical control. What this project adds is the unproven half: perception, learned setpoint policies, and autonomous behaviour — which is exactly why the plan is a ladder of gated tickets rather than one heroic build.
Start reading at architecture, or skim the five research digests that closed on day one.