Reach

The problem. Without sight, finding something you just set down usually means sweeping your hands across the table. Apps can describe a scene or help you aim a phone; Reach closes the last half metre by guiding the hand itself, then tells you what you touched. How it works. You hold a phone over the table and speak. Muse Voice Transcribe runs as one continuous endpointing session with keyword biasing, and a local grammar turns "find the wallet", "what's on the table", "stop" or "louder" into a command about 0.5 s after you stop speaking; requests it can't parse go to Muse Spark as text. SAM 3.1 returns every match for the phrase with boxes, masks and confidences. Between its results (median 0.77 s), Reach carries the target by tracking the table itself with the hand masked out, so guidance never stalls and the lock holds while your hand covers the object. MediaPipe tracks the hand locally on every frame. A sound engine synthesises stereo in 5 ms blocks: pan for left and right, pitch for closer and farther, pulse rate for distance. Contact fires when the fingertip is inside the target, nearly still, and covering part of its mask; Muse Spark then compares a picture from before the reach with the touch and answers in strict JSON. A wrong object is excluded and guidance resumes. More. Two of the same object? Say "left" or "right". "What's on the table" lists what Muse Spark sees. Requests for sharp or hot things are refused before any model call. Hindi requests and replies, a high-contrast low-vision view, volume and verbosity by voice, and automatic camera reconnect. Engineering. One deterministic state machine makes every decision and logs its inputs and outputs, so reach replay rebuilds any session with no camera, models or network. CI replays a published session on Windows, macOS and Linux, alongside 452 tests. Results, on two laptops. 20/20 spoken commands understood on each; 10/10 sighted touches with a correct chime on each; with eyes covered, 5/5 reaches within 45 s on Windows, repeated in a second set; 30/30 saved same-object image pairs confirmed and 15/15 wrong objects caught; no audio dropouts in 10-minute sessions.