Closed on day one · issues #2–#6

Research digests

Five investigations that turned the fog into decisions. Full notes live in docs/research/; the tickets hold the resolution comments.

#2 · Paper teardown → notes

The reference machine is Bai et al., Science Robotics 9, eadi8912: a 35 g Crazyflie carrying a passive spring leg — hopping endurance 50 min vs 6 min flying, and no neural nets in the original control stack. Consequence: we clone the mechanism but must scale the airframe to ~1 kg for a useful payload, and the learned-perception layer is genuinely new work.

#3 · Onboard compute, camera & mic → notes

Jetson Orin Nano Super 8 GB (~30 g, 7–15 W) + RealSense D435i + XMOS XVF3800 mic array ⇒ 200–230 g payload at ~20 W, forcing a 5″-class, 0.8–1.0 kg all-up build. RAM (8 GB) is the binding constraint for vision + voice co-hosting. Raspberry Pi 5 + Hailo remains a challenger only. Speech is only reliable with props stopped — voice is a rest-phase sense.

#4 · Simulation platform → notes

Primary: Isaac Lab 3.0 (Newton / newton_mjwarp) on the 3090 pair; fallback MuJoCo + MJWarp — both consume one shared MJCF model as source of truth. Commitment gate: a quadcopter-hover spike (Isaac-Quadcopter-Direct style) must pass on our machine before further sim investment.

#5 · Flight controller & companion interface → notes

PX4 v1.16+ on H7 boards (Kakute H7 mini / ARK FPV class) with uXRCE-DDS carrying offboard setpoints at 100–200 Hz; COM_DISARM_LAND=-1, ELRS kill switch as the hardware safety path. Betaflight rejected — no offboard setpoint API worth building on.

#6 · NN ramp curriculum & sim-to-real → notes

Eight-stage curriculum S0–S7 (state hover → velocity control → hops → drops → terrain), trained with rsl_rl PPO on the workstation. Domain randomization kept narrow around measured system-ID values. Transfer ladder: hardware-in-loop → tethered → free flight → drops → hops.