Project Details

What it is

NeuroMouse V2.0 is a micromouse: a small autonomous robot that has to find its own way from a corner of a 16×16 maze to the four centre cells, with no map given in advance. It runs on an ESP32-S3, sees walls with four infrared sensors, holds heading with an MPU-6500 gyro, and measures distance with magnetic encoders on two N20 gear motors. I built it with Yu Hong (Elijah) Chen. Our SJSU team, registered as "Neuromouse", placed 3rd overall at the 2026 All American Micromouse Competition (AAMC), held by IEEE at UCLA on May 24, 2026.

NeuroMouse V2.0 from above in a maze cell.

Why

Most open-source micromice run on STM32; the ESP32 is still fairly new in this contest. NeuroMouse is an ESP32-S3 stack that placed on a full-size 16×16 maze, and two months after AAMC the firmware went open source.

We built two mice. V1.0 was a custom PCB, and every mistake (a wrong resistor, a noisy power rail, a sensor too close to a wheel) meant another board spin of about a week. V2.0, the one that placed, is point-to-point wiring on protoboard. Moving a sensor and re-testing took about fifteen minutes.

V2.0 (hand-wired protoboard) next to V1.0 (custom PCB).

How it works

Hardware

• ESP32-S3-WROOM dev board, PlatformIO + Arduino 2.x
• 2× GA-N20 1:30 gear motors (500 RPM at 6 V) with 7 CPR magnetic encoders, on a DRV8833
• 4 IR channels (SFH4545 emitters + TEFT4300 phototransistors), all facing their wall at 90°: two forward, two sideways
• MPU-6500 gyro and 0.96" SSD1306 OLED on one I2C bus
• 300 mAh 2S LiPo; 95 × 85 mm; about 161 g

Top of the V2.0 protoboard stack: ESP32-S3 dev board, OLED, keyswitch.

The motors run 10-bit PWM at 200 Hz. It whines, but on this DRV8833 + N20 combination 200 Hz gave the most breakaway torque; 10–20 kHz gave weaker low-PWM response. The repo has no V2.0 schematic, so include/PinConfig.h is the wiring reference.

Close-up of the hand-wired protoboard and OLED.

Firmware

• One translation unit: src/main.cpp includes every header. Pins live in PinConfig.h, tuning knobs in Tuning.h.
• The ESP32-S3 PCNT peripheral decodes the encoders 4×, at no CPU cost.
• A 16×16 flood-fill BFS picks moves, preferring straight, then left, right and U-turn on ties, and penalising visited cells.
• Each move is a short phase script. PH_FORWARD runs a trapezoidal profile under position-PID, steered by IR centering, encoder balance and gyro yaw-hold. PH_SPOT turns in place on a gyro yaw-PID. PH_ALIGN_FRONT creeps to a calibrated front-wall gap in dead-ends and, during Explore, before a 90° turn with a wall ahead.
• The loop runs at about 200 Hz, paced with micros(). The repo ships with TELEMETRY = false, because Serial prints slow that loop.
• Explore drives to the centre and back, saving walls to NVS at each end. Outbound it brakes at every cell; on the way home it chains straights only through cells it already sensed. Fast Run reloads the map, walls off cells Explore never visited, and fuses straight runs into one move.
• The OLED shows the menu, IR counts, encoder ticks, run state and battery. Dump Walls prints the map to serial. Spinning the right wheel scrolls the menu.

The OLED IR Test screen with live LF/L/R/RF counts and front distance.

Bench debugging with the laptop next to the mouse.

The web-debugger screenshots below come from an earlier firmware sketch that has since been removed from the repo. The repo marks the HTTP debug server in include/WifiDebug.h dormant. It isn't part of the main build, and its page looks different.

Web debugger (earlier sketch, since removed): run state.

Web debugger (earlier sketch, since removed): turn PID.

Web debugger (earlier sketch, since removed): full telemetry and live PID.

Results

On the organizers' results sheet, Neuromouse is ranked 3rd Overall: maze complete, 54 cells visited, 59.715 s. Three teams completed the maze; ours was the third. The UCLA IEEE livestream's results panel shows the same numbers, plus our other two runs: 86 cells in 180.856 s (complete) and 7 cells in 8.166 s (incomplete). My own video labels the 59.7 s run as our third attempt.

My competition run video.

Explore run on the AAMC 2026 maze, from my run video.

Fast run with straight-chain acceleration, from my run video.

Reaching the goal, then the celebration spin.

UCLA IEEE AAMC 2026 livestream, starting just before our run ends.

What I learned

• Iteration beats polish. The hand-wired protoboard mouse is the one that placed.
• A practice maze is not the contest maze. Ours was a 6×3 grid of dollar-store foam board. On 16×16, millimetres add up. At the venue we re-measured a 166 mm wall-to-wall opening (178 mm pitch) and pushed the mouse 15 cells (19,856 ticks over 2,575 mm) to set CELL_TICKS = 1373.
• Some ideas failed on site. Fusing straights during Explore, tried at UCLA the day before, let lateral drift build up, so the outbound leg brakes every cell. Right-hand single-wheel pivots bumped side walls, so the planner dropped pivot turns.
• Put a screen on the robot. You get about five minutes between runs. Plug in, read the OLED, go again.

Wiring taped down during the overnight fixes before the contest.

Links

• Code, docs, photos (MIT): https://github.com/enkhbold470/neuromouse26
• My run video: https://www.youtube.com/watch?v=2M4ZANPrZ4s
• UCLA IEEE livestream (starts a minute before our run ends at 38:37): https://www.youtube.com/watch?v=dCLFL7v4gUA&t=2257s
• AAMC: https://projects.ieeebruins.com/micromouse/aamc