Autonomous Rock Detection — Mars Rover

Led computer vision for autonomous life detection — 2nd place ARC Turkey, Excellence Award IRC India

Context

International rover challenges (ARC, IRC, ERC) are robotics competitions where university teams build Mars-analog rovers and compete on autonomous traversal, sample collection, equipment servicing, and scientific exploration. The science task is the most open-ended: given a Mars-like terrain, find evidence of past or present life.

I led the autonomous vision and life-detection subsystem for the BITS Pilani rover team across two competitions in 2022-23.

The hard part

You’re not looking for life directly — you’re looking for biosignature-candidate rocks: surface textures, mineral coatings, fossiliferous patterns that astrobiologists flag as worth deeper investigation. The judges score you on whether you autonomously route the rover to scientifically interesting rocks within a limited time window, not random rocks.

Three subtasks:

  1. Detect rocks against highly-textured terrain (sand, gravel, basalt)
  2. Classify rock types by biological significance (sedimentary > igneous for biosignatures; certain mineralogies > others)
  3. Plan rover path to the highest-scoring detections

Pipeline

  • Detection — YOLOv5 fine-tuned on a custom labeled dataset of rover-perspective rock images
  • Classification — ResNet-50 transfer-learned on multi-class rock dataset (sedimentary, igneous, biosignature-candidate, etc.) sourced from geology references and augmented for lighting/dust variation common in field deployments
  • Scoring + planning — each detected rock gets a biosignature score; rover priority queue routes to the highest scorers via local path planning

Critical engineering: deployment under field conditions (uneven lighting, dust, vibration, intermittent network). The model had to run on the rover’s onboard compute, not stream frames back to a base station.

Results

  • 2nd Place — Anatolian Rover Challenge (ARC), Turkey — 25+ international teams
  • Excellence Award — International Rover Challenge (IRC), India

What came of it

This was my entry point into computer vision and autonomous systems. The lessons were less about model architecture (YOLO + ResNet is standard) and more about:

  • Ground truth is the bottleneck. Months of labeling rover-perspective rock images by geology students. Models are easy; datasets are not.
  • Deployment teaches you things training doesn’t. Field testing exposed silent failure modes (camera glare, dust occlusion, rover vibration blur) that didn’t show up in clean test sets.
  • Multidisciplinary teams are hard. Coordinating 7 people across mechanical, electrical, embedded, and software subsystems for 6 months taught me as much as the technical work did.

Stack: Python · TensorFlow · OpenCV · YOLOv5 · ROS · Linux

Period: Oct 2022 – Jan 2023