PAAMB — Bioinspired Robotic Snake

Pipelines get inspected from the inside, where people and wheeled rigs don't fit well. A snake robot's cross-section is its entire footprint — if seven joints can be made to move like one body, hold a camera steady, and carry its own sensing, then a pipe no person can safely enter becomes inspectable from within.

The question this project had to answer: could a three-person team build the whole loop — mechanism, firmware, gaits, telemetry, video, and on-device defect detection — and have it work outside a simulator?

Seven 3D-printed joint modules, modelled in Fusion 360 and validated with static-stress FEA before printing. An ESP32 drives seven MG996R servos through a PCA9685; the gait engine plays serpenoid curves — lateral undulation, sidewinding, and concertina — as phase-shifted joint angles.

A second ESP32 carries the sensing: MQ135 gas readings and BMP280 pressure/temperature streamed as telemetry. An ESP32-CAM at the head streams MJPEG to a Flask dashboard, which overlays live defect inference and exposes gait and e-stop controls.

Early in CoppeliaSim the robot did exactly one thing: it rolled onto its side and stopped. Seven iterations chased it. The fix was less control, not more — instead of forcing every joint to stay upright, we corrected orientation only at the root link and let the rest of the body follow the serpenoid curve. Once the head stayed the right way up, the tail sorted itself out.

The model ran. It detected nothing. The dashboard's TFLite inference was silently wrong until the model's real input contract was matched — grayscale, [1,96,96,1], INT8, softmax head. No error, no crash; just confident silence until the tensor shape and quantization matched what the model was actually trained on.

The first FEA was wrong and we knew it: wrong material assigned, force pointing the wrong way. Rather than trust a number that looked fine, the whole study was thrown out and re-run with honest boundary conditions — component by component — until every structural part cleared a factor of safety above 2.6 in ABS.

A repo audit found WiFi credentials hardcoded in the firmware source. They moved to a gitignored secrets.h, and the lesson stuck: hardware projects leak secrets through convenience.