autonomous-ai Vision Tracking
Quick Start
Tracks and follows any object by name. YOLOWorld detects the object in the camera frame, TrackerVit follows it in real-time with servo movement.
Workflow
- User names an object to track.
- If the user also hints at a direction ("look at desk and follow cup", "point at the table and watch the phone"), fire
/servo/aimFIRST so the camera is pointing at the right region before YOLO detection runs. Then fire/servo/track. The aim should complete in ~2s before the track call takes over. - Prefix reply with
[HW:/servo/track:{"target":["<label1>","<label2>"]}]— the device detects and follows.targetaccepts a list of candidate labels. Pass 2-4 synonyms/variants to maximise the chance YOLOWorld finds the object on the first try.- A single string also works (
{"target":"cup"}) for backward compatibility, but the list form is preferred when the object could reasonably have multiple names.
- To stop, prefix with
[HW:/servo/track/stop:{}].
Examples
Input: "Look at the desk and follow the cup"
Output: [HW:/servo/aim:{"direction":"desk"}][HW:/servo/track:{"target":["cup","mug","coffee cup"]}] Looking at the desk and locking onto the cup.
Input: "Point at the table and track my phone"
Output: [HW:/servo/aim:{"direction":"desk"}][HW:/servo/track:{"target":["phone","smartphone","mobile phone"]}] Aimed at the table, tracking your phone.
Input: "Follow the cup"
Output: [HW:/servo/track:{"target":["cup","mug","coffee cup"]}] OK, following the cup!
Input: "Look at the bottle"
Output: [HW:/servo/track:{"target":["bottle","water bottle","plastic bottle"]}] Watching the bottle.
Input: "Track that person"
Output: [HW:/servo/track:{"target":["person","man","woman"]}] Following them now.
Input: "Watch my phone"
Output: [HW:/servo/track:{"target":["phone","smartphone","mobile phone"]}] Got it, tracking your phone.
Input: "Follow the teddy bear"
Output: [HW:/servo/track:{"target":["teddy bear","stuffed animal","plush toy"]}] Tracking the teddy bear!
Input: "Stop following" / "Stop tracking"
Output: [HW:/servo/track/stop:{}] Stopped tracking.
Input: "What can you track?" Output: I can track most common objects — cups, bottles, phones, laptops, books, people, bags, and more. Just tell me what to follow!
How to Control Tracking
No exec/curl needed. Inline markers at start of reply:
[HW:/servo/track:{"target":["cup","mug","coffee cup"]}] Following the cup.
[HW:/servo/track:{"target":["person"]}] Tracking you now.
[HW:/servo/track/stop:{}] Stopped tracking.
Target names
Prefer a list of 2-4 candidate labels in English. YOLOWorld evaluates all candidates and picks the highest-confidence detection across the set — synonyms increase the chance of a successful first-try detection when the user's wording doesn't exactly match COCO/training vocabulary.
Common objects: person, cup, bottle, glass, phone, laptop, keyboard, mouse, book, pen, notebook, bag, chair, monitor, remote control, plate, bowl, plant, vase, clock, lamp, speaker, headphones, watch, glasses, hat, shoe, toy, ball, teddy bear.
Any label works (open-vocabulary detection). Don't include too many unrelated items in the list — that risks matching a nearby but wrong object.
How it works internally
- Camera captures current frame
- YOLOWorld API evaluates all candidate labels (~1-2s)
- Highest-confidence detection becomes the initial bbox
- TrackerVit locks on and follows in a move-then-freeze loop (~7 FPS)
- Servo base_yaw + base_pitch nudges to keep object centered
- Auto-stops when object is lost, out of range, or after 5 minutes
Error Handling
- If the object is not found: "I can't see a {target} right now. Try pointing me toward it, or try a different name."
- If tracking stops unexpectedly: "I lost the {target}. Want me to try again?"
- If servo is not available: "Servo is not available right now."
Rules
- Only one object can be tracked at a time. Starting a new track stops the previous one.
- Tracking auto-stops when the object leaves the frame, gets occluded, or after 5 minutes.
- When tracking stops, the servo holds its last position (no snap back to idle).
- Do NOT use this for emotional reactions — use the Emotion skill instead.
- Prefer
targetas a list of 2-4 English synonyms for better first-try detection. Avoid packing unrelated labels into the list.