// PROJECTS · 29 WEEKS

Ten projects from a home lab.

Start on the laptop, then move to the arm on the desk, then to a rover in the garage. Projects 1–7 are the spine. Projects 8–10 are optional depth if the core is solid. Each one ends with a folder you can open, a short clip, and a row in the shared eval log.

PLANNEDP07 · Intermediate → Advanced · core

Tennis-ball collector

The P6 rover plus a detector plus an intake. A working Tennibot, built so you know which parts were actually hard.

Weeks 16–19 · P6 M3 Pro + intake · ~$90 · 07_ball_collector/

Gazebo Harmonic (extends the P6 world)Train · RTX 5090 (detector)Run · Yahboom M3 Pro Orin NX Super 16GBBuy · Intake motor ~$25 · printed hopper/roller ~$20 · beam-break ~$15 · tennis balls ~$30 → ~$90ROS · YesGazebo · Yes
Tennis-ball collector
PLANNED
DEMO 07 · PLANNED
DEMO 07 · PLANNEDVideo lands here when this milestone ships.
Physical rover clearing 20 tennis balls, unattended, plus GR04.
Rover on an indoor course
Rover on an indoor course
Waypoint course
Waypoint course
YOLOHSVNav2IntakeTennibot
01

Goal

Tennibot’s Rover is $2,275. Most of that robot is P6. Start perception cheap: HSV → blob → ray to ground plane → map-frame pose with covariance. Task loop: nearest ball → Nav2 → intake → beam-break confirm → repeat. Then train a small YOLO on your own court footage and compare against HSV on the same tape. Iterate the printed intake and log jam rate per 50 balls. Stretch: frontier exploration and a ball-density heatmap. Honest surprise: vision is the easy part. The hard parts are the intake, the hopper on turns, and a battery that lasts.

02

Why it matters

You took a commercial capability list and built a working version from scratch, and you know which parts were hard. That is a different kind of knowledge from following a tutorial. Tennibot’s 4–5 hour battery figure is a hardware achievement, not a software one.

03

The work

  1. 01

    HSV segmentation → centroid → ground-plane ray → map-frame detection with covariance.

  2. 02

    Task loop: nearest ball → Nav2 goal → intake → beam-break → mark done → repeat.

  3. 03

    Train YOLO on the 5090 from your own court images. Compare vs HSV: precision, recall, FP/min, Orin NX latency.

  4. 04

    Printed roller + hopper. Log jam rate per 50 balls.

  5. 05

    Stretch: frontier exploration of an unmapped court; density heatmap.

04

Done looks like

  • Detector table: HSV vs YOLO (precision, recall, false positives/min, Orin NX latency).
  • Sim: 20 balls in the Gazebo court, cleared autonomously, metrics logged.
  • Real demo: physical rover clearing 20 balls from a court or driveway, unattended, on video.
  • Metrics: collected/total, time-to-clear, path efficiency per ball, jam rate, collisions.
  • Battery runtime under load, power budget by subsystem, honest comparison to Tennibot’s 4–5 hours.
  • docs/gap_reports/GR04-ball-collector.md
05

Stack card

SimulatorGazebo Harmonic (extends the P6 world)
Trains onRTX 5090 (detector)
Runs onYahboom M3 Pro Orin NX Super 16GB
BuyIntake motor ~$25 · printed hopper/roller ~$20 · beam-break ~$15 · tennis balls ~$30 → ~$90
ROS 2Yes
GazeboYes
Compute / targetYahboom M3 Pro Orin NX Super 16GB
InstallEverything from P6, plus ultralytics (YOLO) or OpenCV HSV, image_transport, vision_msgs
06

Repos to study

07

Watch out

  • Do not treat HSV working as the finish line. The intake jam rate is the number.
  • Keep P6 succeeding on its own terms. This project stands on that rover.
08

Workstreams

BUILD STREAM

Perception node, task loop, intake mechanism, hardware demo.

EVAL STREAM

HSV-vs-YOLO table, clear metrics, battery budget, GR04.

Cross-check. Same 20-ball layout in Gazebo and on the driveway.

09

What this proves

You built a working version of a $2,275 product and can say which parts were hard, and why.

The vision is the gift. The intake and the battery are the project.

10

Study videos

11

Gear / BOM

Collector intake kit
Collector intake kitRoller, hopper, beam-break, tennis balls on the P6 M3 Pro
~$90PLANNED
Gears: hardware + stack →