ESP32-S3 Open-Source Micromouse Firmware Deep Dive: From Perfboard to AAMC Third Place
AAMC 2026 third-place Micromouse team open-sources their ESP32-S3 perfboard build and full firmware.
A team competing in UCLA's All American Micromouse Competition (AAMC 2026) took third place and open-sourced their complete technical solution. They abandoned custom PCBs and the standard STM32 in favor of an ESP32-S3 on hand-soldered perfboard, enabling 15-minute hardware iterations. Their stack includes flood-fill BFS path planning, a 200Hz bare-metal PID loop, trapezoidal velocity profiling for cell-merging on straights, and NVS Flash maze caching for a fast-run sprint strategy. The ESP32-S3's hardware PCNT peripheral handles quadrature encoder decoding without CPU overhead. A 0.96-inch OLED proved critical during the 5-minute competition prep window.
Micromouse — the classic and hardcore sport of building a tiny autonomous robot to explore and sprint through a 16×16 maze as fast as possible — sits at the intersection of embedded systems, robotics, and competitive engineering. Recently, a team that competed in the All American Micromouse Competition (AAMC 2026), hosted by UCLA IEEE, shared their complete technical solution on Reddit after taking third place. They open-sourced everything: firmware, schematics, and build logs.
What makes their story particularly compelling is that they broke convention — ditching the industry-standard STM32 in favor of an ESP32-S3, and ultimately competing with a hand-soldered perfboard build.
From Clean PCB to Perfboard: A Pragmatic Choice for Rapid Iteration
The team started with a clean, custom PCB as their V1. But in practice, they quickly ran into a painful engineering reality: every time they encountered power-rail noise or needed to nudge a sensor's position, they faced a full week-long PCB re-spin cycle.
For a time-pressured competition project, that iteration speed is fatal. So the team made a bold call — scrap the polished PCB entirely and switch to raw perfboard with point-to-point soldering, even using a mechanical keyboard blue-switch clicker for mode selection.
They openly admitted that V2 looked like a "rat's nest." But it was precisely that flexibility — being able to desolder and reposition an IR emitter in 15 minutes — that got them to the competition floor. This decision is a vivid illustration of a core engineering principle: during prototyping, iterability beats aesthetics. The ability to fail fast and correct course quickly often matters more than theoretical perfection.
Why ESP32-S3 Instead of STM32?
In competitive Micromouse, nearly every serious team runs STM32. This team went the other direction, choosing ESP32-S3 with PlatformIO. The reasoning was clear-eyed.
Hardware Pulse Counter for Encoder Decoding
The team made full use of the ESP32-S3's hardware capabilities. They used the ESP32's hardware Pulse Counter (PCNT) peripheral to implement quadrature encoder decoding, offloading the work entirely from the CPU and preventing interrupt storms. This is a textbook example of "use hardware to unburden software."
NVS Flash Caching for Sprint Runs
Even smarter: they cached the maze wall data discovered during the exploration phase into ESP32's NVS (Non-Volatile Storage) Flash. This means that during the final fast run, the mouse can skip sensor-based wall detection entirely and drive full-speed based on the stored map. This "explore → memorize → sprint" strategy is a key differentiator in high-scoring Micromouse runs.
Motion Control and Core Algorithms
A mouse's performance ultimately comes down to the precision of its motion control and the intelligence of its path planning. The team put serious work into both.
Flood-Fill Algorithm and 200Hz PID Control Loop
Path planning uses a classic 16×16 flood-fill BFS solver — the standard Micromouse approach — to compute the shortest path from start to finish.
For the control loop, the team implemented a 200Hz PID cycle. Notably, the loop timing is driven entirely by micros() — no RTOS tasks, no delay() calls anywhere in the control path. This "bare metal" approach to timing, while primitive, guarantees determinism and real-time responsiveness by eliminating OS scheduling jitter.
Trapezoidal Velocity Profiling: Blasting Through Straights
The most impressive algorithmic work is their trapezoidal velocity profiling. A naive implementation would decelerate the mouse at every 180mm cell boundary, dragging down overall speed. Their approach merges consecutive straight-line cells into a unified acceleration corridor, letting the mouse accelerate continuously through long straights rather than braking at every cell boundary. This optimization is critical for fast-run performance.
IR Sensor Fusion and Gyroscope Yaw Hold
The sensing system uses four pairs of IR emitters/receivers (SFH4545 + TEFT4300) with lookup-table-based distance calibration, combined with an MPU-6500 gyroscope for yaw-hold. The quality of sensor fusion directly determines whether the mouse can stay stable at high speed.
The Scaling Challenge: From a 6×3 Home Grid to the 16×16 Competition Maze
The team shared a very honest lesson learned: they tested at home using a roughly $10 Home Depot whiteboard and 3D-printed walls, assembled into a compact 6×3 test grid. When it came time to scale up to the official 16×16 UCLA maze (256 cells), the challenges multiplied fast.
The core issue is error accumulation — millimeter-scale inaccuracies that are invisible on a small grid compound rapidly over long straights. A solution that performs flawlessly on a small test grid can cause wall collisions or disorientation in a full-size maze due to accumulated drift. This is a reminder for all hardware developers: there's often an unpredictable gap between small-scale validation and real-world deployment. Scaling itself is an engineering challenge.
The MVP of Competition Day: A 0.96-Inch OLED Display
The team specifically highlighted that the real "most valuable player" on competition day was the 0.96-inch OLED display. With only a 5-minute prep window at the event, being able to read battery voltage, IR sensor values, and the flood-fill map directly on the robot — without opening a laptop — made all the difference.
This detail seems minor but represents genuinely valuable competition hardware wisdom: building intuitive on-device diagnostics into your system can be decisive in high-pressure moments.
Open-Source Repository and Engineering Takeaways
The team has open-sourced everything — firmware, schematics, and build logs — in their GitHub repository (neuromouse26), and posted full competition run videos on YouTube. For anyone looking to get into Micromouse or embedded real-time control, this is an exceptionally practical reference.
From this project, a few broadly applicable engineering lessons emerge:
- Iterability sometimes beats polish — perfboard is ugly, but it's flexible
- Dig deep into your chip's hardware peripherals — ESP32-S3's PCNT and NVS delivered unexpected performance gains
- On-device observability is invaluable — a tiny OLED screen is worth far more than expected under competition pressure
These lessons extend well beyond Micromouse and apply to embedded systems and robotics development broadly.
Related articles
Hacktron Automations: A Deep Dive into AI-Powered Closed-Loop Security with Automatic Vulnerability Remediation
A deep dive into how Hacktron Automations uses AI for closed-loop security — covering automatic vulnerability detection, dynamic validation, intelligent patch generation, and comparisons with traditional SAST tools.
Desert Ant Labs: On-Device AI Model Local Inference Solutions
Desert Ant Labs builds AI models that run fast on local devices, offering data privacy, zero latency, and offline availability through advanced model optimization techniques.
Claude Credits Gone in 10 Minutes? A Guide to Token Consumption Analysis and Optimization
Why does Claude drain your quota so fast? We break down context accumulation, coding tool costs, and share token tracking tools and optimization tips for developers.