Roadmap¶
Each milestone has a deliverable and a pass criterion. Results, good or bad, get published on this site and in the repository, flight logs included.
timeline
title Ariadne
Done : v1 flood fill, pivot planner, 10 kHz physics, ESP32 firmware
: v2 clothoid turns, diagonals, path following, wall observer
: v3 design studies, website, Mk II proposal
M1 : Mk I built and brought up
: model fitted from flight logs
M2 : smooth fast runs on a real maze
M3 : smooth search turns
M4 : Mk II custom PCB
M5 : 1 kHz IR estimator, diagonals on hardware
M6 : write-up and open datasetDone¶
| Version | What |
|---|---|
| v1 | Flood-fill search that stops when the path is provably optimal; time-optimal pivot planner; 10 kHz physics model running the firmware control law; Monte Carlo over tolerances; ESP32 + FreeRTOS firmware; schematic checked against the pin map |
| v2 | Clothoid turn designer with worst-case clearance; Dijkstra on edges × 8 headings and a two-pass speed plan, bit-identical in C++; path following and the distance-domain wall observer; gyro scale calibration |
| v3 | Design studies (control margins, reliability with confidence intervals, sensors, gyro, speed/reliability trade-off, chassis width, angular acceleration); this website; the Mk II proposal |
Next¶
M1. Mk I on real wheels¶
Deliverable: the module-based robot assembled and through the bring-up plan steps 1 to 6.
Pass: ten clean pivot fast runs in a row on a practice maze; flight logs published.
Research part: fit the motor and traction model (\(k\), \(R\), \(\eta\), \(\mu\)) to the logs by least squares and feed it back into config.h and the simulator. The simulator is then judged by one number: cross-track error, simulated against measured.
M2. Smooth fast runs on a real maze¶
Deliverable: the "smooth turns" mode tuned at the safe, then normal preset. Pass: ten clean runs out of ten; measured run time within 5 % of the planned time; real cross-track RMS within 2 mm of the simulated value.
M3. Smooth search turns¶
Deliverable: the search turns on the move instead of stopping and pivoting. Simulated first in a closed-loop search model (ToF-based wall detection, decision point before the turn). Pass: search time on the practice maze reduced by a third with no wall touch in 20 searches.
M4. Mk II custom PCB¶
Deliverable: the Mk II board: ESP32-S3, ICM-42688-P, pulsed IR at 1 kHz, current sense, PCB chassis 64 to 72 mm wide with the battery over the axle (lower yaw inertia). Pass: board bring-up (rails, IMU noise and drift, IR settling and calibration residuals, motor model) documented; Mk I software running unchanged above the new HAL.
M5. IR estimator and diagonals on hardware¶
Deliverable: observer updates at 1 kHz; wall-edge (post) detection for along-track correction; post-based lateral correction on diagonals. Pass: diagonal mode at least as reliable on the real maze as smooth mode, ten clean runs out of ten.
M6. Write-up and open dataset¶
Deliverable: a technical report on simulation-first design of a micromouse (turn synthesis, planning, distance-domain estimation, sim-to-real gap), and the dataset of flight logs with the matching simulated runs.
Research questions this project can answer¶
- How small can the sim-to-real gap be made for a small differential-drive robot with consumer sensors, and which model terms matter most (motor, tyre, sensor timing)?
- A search cost that matches the speed plan. The Dijkstra cost underestimates the exact time by 15 % on average and ranks close routes imperfectly (width study). Candidates: speed-aware states, or re-ranking the k best paths by their exact speed plan, within the firmware's 70 kB and planning-time budget.
- Chance-constrained planning: choosing turn speeds so that the Monte Carlo probability of finishing stays above a target, instead of a fixed lateral-acceleration limit.
- Variable-speed turns: accelerating out of a turn inside the friction circle rather than holding constant speed through it.
- Jerk-limited profiles and their effect on wheel slip measured by motor current.
- Search that targets the unknown cells on the optimistic shortest path instead of returning to the start, measured in search cells over thousands of mazes.