Skip to content
← The Workbook

No. 48 · 2026Concept

Fetch

A computer-vision dog trainer that runs while you're at work. Camera watches the dog. Recognizes good behavior (sitting calmly, not chewing the couch, going to the spot) and rewards it with a treat dispenser or a tossed ball. Recognizes the bad stuff and either redirects (toy launches in another room) or just logs it for the human to see at end of day. Replaces the eight hours of dog-being-alone with eight hours of dog-being-coached.

Status
Concept
Year
2026
Stack
Computer Vision · Hardware · Pets · Local Inference · AI · Pi · Jetson

The Idea

A dog spends most of the workday alone. Eight hours of waiting for someone to come home and notice them again. The expensive dog training you paid for at week one fades because nothing reinforces it for the rest of the week.

Fetch is the trainer that stays home with the dog.

A small camera watches the room. A computer-vision model recognizes a vocabulary of dog behaviors. When the dog does something good, a hopper drops a treat. When the dog finds the spot, a ball launches across the floor for a chase. When the dog goes after the couch, an attention-redirector noise plays from the speaker on the other side of the apartment.

You come home and the dog is calmer than you left it. The training is reinforced every day instead of weekly.

What The Camera Watches For

The model is trained on a behavior taxonomy, not a generic object detector. The categories that matter:

Reward triggers:

  • Sitting calmly in the designated spot (mat, bed, crate)
  • Lying down without anxiety
  • Coming when called by a recorded voice cue
  • Going to a target (paw on the touch pad, nose at the bell)
  • NOT doing the things they normally do when the human leaves

Redirect triggers:

  • Approaching forbidden zones (couch, trash, counters, off-limits doorways)
  • Chewing on furniture or shoes
  • Excessive barking at outside stimuli (delivery, mailman, other dogs)
  • Pacing or destructive anxiety patterns

Just-log triggers:

  • Sleeping (so the human sees how much rest the dog actually got)
  • Drinking water (intake monitoring matters for older dogs)
  • Bathroom accidents (location + timing for crate-training adjustment)

The vocabulary is per-dog. You start with a base set and tune it during a setup week, the system asks you to label things it sees, you confirm or correct, the model adapts.

The Hardware

  • Camera node. A small Pi-class device (Raspberry Pi 5, Jetson Orin Nano, or off-the-shelf ESP32-CAM) running a quantized vision model. Local inference, not cloud, privacy plus latency, both matter here.
  • Treat dispenser. Servo-driven hopper, holds a day's worth of small training treats. Triggers on reward events.
  • Ball launcher. Off-the-shelf automatic ball thrower (PetSafe makes a few good ones) wired to a smart relay. Triggers on chase-cue events.
  • Attention-redirect speaker. Small bluetooth speaker positioned in another room. Plays a soft, non-scary "look over here" cue. Pulls the dog away from forbidden zones without scaring them.
  • Phone bridge. App on the human's phone shows a live feed, lets the human override (treat now, ball now, voice cue), and sends end-of-day summaries.

What The Human Sees

End-of-day summary email or push notification:

  • "7h 14m alone. 5h sleeping, 1h roaming, 2h at the spot. 3 reward events. 1 redirect event (counter at 11:42). No accidents. Sleep quality looked good. Three short barks at the mailman. Highlight reel attached."

A weekly report shows trends: time-to-settle is dropping (the dog is calming faster). Couch-redirects are down by half (training is sticking). Water intake is consistent.

The system isn't replacing the human. It's filling the eight hours the human is away with consistent reinforcement so the relationship continues progressing instead of regressing.

Privacy + Honesty

  • Local inference only. Camera frames don't leave the home. The vision model runs on the Pi or Jetson; only structured event logs (timestamps, behavior labels, occasional summary clips the human approves) ever sync to the cloud.
  • No social-media-shaped engagement tricks. Notifications are end-of-day summary plus opt-in highlights. No streaks, no badges, no "your dog is lonely" guilt prompts.
  • The dog's relationship is with the human, not the box. The reward dispenser is a coach during the workday. Time with the human is still the high-value reinforcer. The system is designed not to replace that bond, just to keep training consistent during the gap.

What Could Go Wrong

The honest list:

  • Treat-machine dependence. A dog can fixate on the dispenser instead of relating to the room. Mitigation: variable-ratio rewards, periodic dispenser-off days, end-of-day in-person reinforcement.
  • Misclassified behavior. A scratch confused for an attack. A sit confused for a sit-stay. Wrong reward at the wrong moment teaches the wrong thing. Mitigation: high-confidence threshold, human-approval queue for ambiguous events during setup.
  • Wrong hardware for the wrong dog. A 90-pound lab gets nothing from a launcher that throws a ping-pong-sized ball. A 6-pound Yorkie can't lift the treats meant for a Shepherd. The system needs to size to the dog, not the average dog.
  • Surveillance creep. A camera watching the dog is also a camera watching the room. Local-only inference helps. Periodic full-system audits matter.
  • Reinforcement of the wrong behavior. A bored dog could discover that pacing in front of the camera looks like "active engagement" and get rewarded for it. Mitigation: behavior taxonomy reviewed weekly, model retrained on actual outcomes.

Pairs With

Barkangel (the GPS dog collar that the dog charges by walking) is the away-from-home companion. Fetch is the at-home companion. Together: the dog is trained at home during the day, traceable when they're outside.

Status

Concept. The hardware is straightforward off-the-shelf or already-prototyped (Pi or Jetson camera node, servo dispenser, smart-plug for the ball launcher). The interesting work is the behavior taxonomy and the per-dog tuning loop. A v1 with five reward triggers, two redirect triggers, and an end-of-day summary is a long-weekend build assuming the vision model is fine-tuned (transfer-learn from a base detector like YOLOv8 or a vision-language model running locally).

This is a good candidate for a real product, not just a workbook entry.