← all challenges

// challenge brief · hack a ton 2026

Cat Tracker

From a single robot camera, find the cat, keep its identity, and steer the robot to follow it at a safe distance, all in real time.

// Proposed by
Monsson
// Industry
Computer vision / robotics (Unitree Go2 Edu)
// Difficulty
🟡 Intermediate

★ Own jury, own €1,000 prize. Pick any one of the three to compete.

// the problem

A quadruped robot that follows a moving target needs to do three things at once: see the target, keep track of which target it is across frames, and turn raw perception into smooth motion. Do any one badly and the robot loses the cat or jerks around chasing shadows.

// your mission

Build a module that, from a monocular video stream recorded by the robot, detects and tracks cats in real time and produces control commands so the robot follows them while keeping a safe distance.

// how it works

  1. 1Perceive: detect cats in the frame (single and multiple).
  2. 2Track: maintain each cat's identity between frames through motion and partial occlusion.
  3. 3Estimate: infer the target's direction and relative speed.
  4. 4Act: output a control vector (dx, dy, rotation) at ≥15 FPS on a consumer GPU.

// what we provide

  • ~15 clips (30–60s) filmed with the Go2 camera, with varied lighting and angles plus partial occlusions.
  • A hidden test set for scoring.

// suggested approach

  • Pretrained detector (e.g. YOLO). "cat" is a standard class, so no training required.
  • An off-the-shelf tracker (ByteTrack / DeepSORT) for identity across frames.
  • Simple control logic mapping target position/size to (dx, dy, rotation), with smoothing to kill oscillations.
  • Floor version: single cat, detect + track + a basic proportional follow command. Add multi-object and speed estimation from there.

// how you'll be judged

  • Accuracy on the hidden test set.
  • Robustness to occlusions and sudden movements.
  • End-to-end latency.
  • Command quality: smooth, no oscillations.

// stretch goals

  • Reliable multi-cat tracking with re-identification after full occlusion.
  • Predictive following (anticipate the path rather than chase the last position).

// deliverables

  • Git repo: code + README + run instructions.
  • Demo: notebook or script on the provided test set.
  • 5-minute pitch: architecture, decisions, trade-offs.