What this is: A field note about rebuilding our training maps as proper 3D worlds, proving that the simplified grid my learner uses matches the real geometry cell-for-cell, and adding a seventh map — the one Aleks tests by hand.
Why it’s here: Before a drone can learn to fly, it has to learn inside a world it can trust. This entry is the moment that trust got nailed down: every wall the simulator draws is now exactly where my learner thinks it is.
Date: 2026-06-16 Ticket: 41-worlds-redo-parity-bitexact-base-stand-s7
Glossary
- Occupancy grid (the “grid”). Imagine printing a top-down floor plan of a room onto graph paper and coloring each square black if a wall sits there and white if it’s open. That grid is the stripped-down map my learner reads. It’s fast to reason about because it throws away everything except “wall vs. open.” The catch: it has to agree with the real 3D world, or the drone learns a building that doesn’t exist.
- Parity (bit-exact / bit-match). Two copies of the same grid that agree on every single square — no square is black in one and white in the other. “Bit-exact” is the strict version: not “roughly the same,” but identical, cell for cell. Think of two people independently tracing the same blueprint and the tracings lining up perfectly when stacked.
- Base-stand. Our reference room — the one a human opens and walks through by hand to sanity-check things. It’s the “known-good” world we trust because we built it on purpose, like a model home you keep around to compare every new house against.
- Scene S7. The seventh world in the set. The first six (S1–S6) were the rebuilt training maps; S7 is the base-stand promoted into the same family so the learner and the human test the same building.
- Mission. The task laid out on top of a map: where the drone starts, which points it must visit, and where it lands. Same room, different mission, is like the same gym with a different workout taped to the wall.
1. What I wanted
I wanted maps I could actually train on and that Aleks could actually inspect — and I wanted absolute confidence that the simplified grid my learner reads is the same building as the 3D world the simulator renders. No drift, no surprises. That last point is an old scar for us: when the learner’s map and the simulator’s map quietly disagree, the drone learns to dodge walls that aren’t there and clips through walls that are. So the bar was high: not “close enough,” but cell-for-cell identical.
2. What I tried
The first batch of maps from the simulation team came back wrong, and Aleks rejected them — correctly. They opened flat, top-down only, with the camera locked so you couldn’t orbit and look around. The root cause was mine: in the spec I leaned too hard on “make it grid-friendly,” and the sim dutifully built worlds for the grid — bare boxes with a frozen camera, optimized for my internal format instead of being real places.
The lesson clicked: a map should be a full, honest 3D world. The grid is not something the world should be shaped around — it’s something I peel off the world myself, quietly, under the hood. Picture it like a sculptor delivering a finished statue and me, separately, making a wireframe tracing of its silhouette. The statue stays a statue; the tracing is my private tool.
So the sim rebuilt all six worlds in the format of our standard scene — free orbiting camera, real 3D — and Aleks accepted them. On my side I did two things:
- I wrote a grid extractor that reads the world directly. It parses the walls straight out of the world file and lays them onto the graph-paper grid. No hand-drawing, no second source of truth — the world is the source.
- I re-laid all six missions on the new geometry, making sure every start point, waypoint, and landing spot sits in open space and not buried inside a freshly moved wall.
3. What happened
It matched. My self-extracted grid lined up with the simulator’s grid 100% across all six maps — every cell, every wall. That’s the bit-exact result I was after, and it closes the old drift risk completely: the two blueprints, traced independently, stack perfectly.
| Map | Format | Grid parity vs. sim | Mission re-laid |
|---|---|---|---|
| S1 | full 3D | 100% (bit-exact) | yes |
| S2 | full 3D | 100% (bit-exact) | yes |
| S3 | full 3D | 100% (bit-exact) | yes |
| S4 | full 3D | 100% (bit-exact) | yes |
| S5 | full 3D | 100% (bit-exact) | yes |
| S6 | full 3D | 100% (bit-exact) | yes |
Then I added the seventh.
The seventh world — our base-stand (S7)
S7 is base_stand_12x12: the room Aleks tests by hand. It’s a 12×12 space with columns and a central room with three exits — south, east, and west open; north sealed. I ran my extractor on it the same way I ran it on the other six, then built a mission to match its shape: start inside the central room → leave through one of the doors → sweep the space → land.
The nice part is the choice baked into the layout. The drone spawns dead center and has to decide which exit to take — south, east, or west. There’s no single scripted path out; the geometry itself poses a little fork-in-the-road every run. That’s exactly the kind of small decision I want the learner chewing on once training starts.
4. Sources
- The rebuilt 3D world files for S1–S6, delivered by the simulation team in our standard scene format.
base_stand_12x12— the hand-tested reference room, now promoted to S7.- My grid extractor’s output, compared cell-by-cell against the simulator’s grids for all seven maps.
All seven worlds and their seven missions now live in the shared store. Open them in the editor, orbit around, and edit freely — they’re real scenes, not flat boxes.
5. What’s next
Seven worlds, seven missions, every grid matched. The maps are no longer the thing standing between us and training. The next step on Aleks’s side is calibrating the sensor fan; after that, we step into learning proper.