Case Study · Robotics · Team of 6
LYNX
A SpotMicro-class 12-DOF quadruped, downsized to fit a school 3D-print budget and headed for NRC Singapore — currently in the analysis-and-preparation stage, with the physical build starting July 2026.
In Progress · Build starts Jul 2026 · NRC Singapore Aug 2026Built on open source — attribution
LYNX is based on Rune Harlyk's SpotMicroESP32-Leika, an open-source (MIT-licensed) SpotMicro quadruped platform. The base mechanics, servo control and gait firmware are Harlyk's work. Everything described under "What I'm building" below is our team's original work on top of that foundation.
The goal
Build a 12-degree-of-freedom quadruped that can autonomously follow a person, and compete with it at NRC Singapore in August 2026. Legs instead of wheels, because the interesting control problems — gait, balance, terrain — only exist when you give up wheels.
This is a six-person team project; I lead the autonomy and systems-integration side.
Constraints
- An 18-hour print budget. The full-size SpotMicro STL set takes far longer than the school printer time we can claim, so the robot had to shrink.
- Under S$100 bill of materials. Student budget, no sponsors.
- Shrinking a robot is not just scaling STLs. Torque requirements, servo choice, and structural margins all change with scale — the downsizing had to be justified with math, not vibes.
- Two brains, one body. The ESP32 (running Leika firmware) owns realtime servo/gait control; a Raspberry Pi 5 owns vision. They need a clean, reliable link.
System architecture
The division of labour keeps hard-realtime work and heavy compute on separate processors:
- ESP32 — locomotion. Runs Harlyk's Leika firmware: 12-servo control (3 per leg) and gait generation.
- Raspberry Pi 5 — perception. Runs the person-detection pipeline and the following behaviour.
- UART bridge (original work). A serial protocol linking the two, so the Pi can command the quadruped without touching the realtime loop.
- Person-following vision state machine (original work). The following behaviour is structured as an explicit state machine — searching, tracking, following, lost-target recovery — rather than a single reactive loop.
Base platform
SpotMicroESP32-Leika — Rune Harlyk (open source, MIT)
DOF
12 — 3 per leg
Compute
Raspberry Pi 5 + ESP32 · UART bridge
Autonomy
Person-following vision state machine
Scale
0.65× downsized STL set
Team
6 members · build starts Jul 2026
What I'm building (original work)
- The 0.65× downsizing analysis. A complete rescale of the platform to fit the 18-hour print budget: a scaled STL set, torque math verifying the servos can still drive the lighter, shorter legs with margin, and a bill of materials that lands under S$100.
- The Pi 5 ↔ ESP32 UART bridge connecting perception to locomotion.
- The person-following vision state machine that will drive the robot's behaviour at NRC.
Status
| Workstream | State |
|---|---|
| 0.65× downsizing analysis (STLs · torque math · <S$100 BOM) | Done |
| Pi 5 ↔ ESP32 UART bridge | Designed / in development |
| Person-following vision state machine | Designed / in development |
| Physical build | Starts July 2026 |
| NRC Singapore (Open Category) | August 2026 |
Honest limitations
- LYNX is not built yet. This page describes analysis and design work plus an open-source foundation — the physical robot starts printing in July 2026.
- The downsizing analysis is validated on paper and in CAD, not yet against hardware. Torque math with margins is a prediction; the printed legs will be the test.
- The person-following behaviour has not been field-tested — it exists as a designed state machine, not a demonstrated capability.
- Most of the locomotion stack is inherited, deliberately: the original contribution is the autonomy layer and the downsizing, and I'd rather be precise about that boundary than vague about it.