May 2026 was a setup-and-first-results month. The three weeks covered here represent the project moving from “we have a directory and a plan” to “we have a working simulation pipeline producing PPO checkpoints.”
Week 1 — ArduPilot SITL + Gazebo + ROS 2 Jazzy got installed and talking to each other for the first time. The first MAVLink heartbeat from the simulated drone was the headline moment. Everything before that point was infrastructure; everything after it was actual engineering.
Week 2 — the first PPO training run finished, then the first frontier-reward variant, then the first attempt at RND-driven exploration. This is the week that proved the RL track is real, not aspirational.
Week 3 — physics-tuning week. Motor dynamics in Gazebo are not trivial when you want them realistic enough to transfer to a real drone, and a non-trivial amount of effort went into matching the simulated airframe to what the physical drone will (eventually) behave like.
A note on cadence: weekly digests are written after the work is done, not as a live feed. The raw events live in the agent reports (append-only); these digests are the curated narrative version.
Contents
Auto-generated from child entries during build (update-indexes.mjs).