Proprio
Find it. Grab it. Eyes closed.
One sentence: a blind person asks for something on the table and spoken words walk their hand to it. Every model runs on the ASUS GX10.
#1
everyday challenge: finding a dropped or misplaced item
68%
of their requests for help needed an answer within 10 minutes
Lancet Global Health, 2020
Yi et al., Finding Objects for Assisting Blind People, 2013
Brady et al., CHI 2013: 40,748 questions from 5,329 blind users
43M blind worldwide, 61M by 2050 (Lancet GH 2020). #1: finding dropped or misplaced items is described as the most common challenge (Yi et al. 2013); interview studies say it touches nearly every household task. 68%: in VizWiz Social, most questions needed an answer within 10 minutes, so help has to be real time (Brady et al., CHI 2013).
An aid thatguides your hand.
Ask for any item, and one spoken word at a time walks your hand onto it. It remembers the room, so it can find what you can’t see.
Say it out loud: "Find my mug." Then let the scene talk: the hand moves one axis at a time and each bubble is the word in the earbud. Hands-free, any object by name, no setup. Memory: it keeps where items are in the room, so an item behind your hand or outside the camera's view can still be found. Today it holds a hidden item's place for 15 seconds using the phone's own motion tracking; the product keeps a map of everything it has seen.
What exists today
Meta glasses a volunteer talks you through it needs internet and a person Be My AI describes a photo never guides your hand Seeing AI leads you toward an item gets you near it, not onto it AIGuide guides your hand handheld, objects scanned first Proprio guides your hand by voice hands-free, offline, any object
Be My Eyes on Ray-Ban Meta, 2024; Microsoft Research, Find My Things; Lee et al., TACCESS 2022.
One Enter per tool: say what it does, then the gap. Meta glasses: hands-free, "Hey Meta, Be My Eyes" calls a sighted volunteer; needs internet and a person; 56% retrieval accuracy in an ASSETS 2025 study. Be My AI sends a photo to the cloud and describes it. Seeing AI Find My Things needs four videos per item and leads you until the item is within arm's reach, then stops: it never steers your hand onto it. AIGuide guides the hand but is handheld with pre-scanned objects. Last Enter: ours fills every gap.
Live demo
Or watch this video:
youtube.com/watch?v=wdUEhJDfMpk
The Ethernet cable is already out. "She's blindfolded, and the internet is unplugged." Wearer: "Find my mug." Narrate the screen; one grab, about 30 s, then back to the deck. Fallback: the demo video on YouTube, which this slide shows whenever the dashboard is not reachable.
Why not vibration?
74%
of overlapping buzzes were misread
Yeganeh, University of Iceland 2026: 26% read correctly when buzzes overlapped, 93% one at a time.
Vibration needs a motor per direction spread over the body (a wrist band, a belt, a face rig), and the buzzes overlap: when buzzes overlapped, 74% were misread (26% read correctly, vs 93% one at a time). A face rig has the same problem with more motors. Even single wrist buzzes were felt in the wrong spot 1 time in 10 (Trant et al., Sci Rep 2025, 41 blind participants), and spoken cues gave faster grasps. One word in an earbud needs no decoding.
How it works
See: iPhone color + lidar to the GX10. Find: Whisper hears the request and Nemotron turns it into an item name ("I'm thirsty" becomes water bottle); YOLO-World finds the mug, MediaPipe the fingertip. Measure: lidar turns both into 3D points. Speak: the biggest gap becomes one word in the earbuds.
The core loop
pixel and lidar depth to 3D where the hand must go one word: right, down, forward then “grab”
Median depth over a small patch, ignoring holes. Table plane by RANSAC. Words at most twice a second. About 35 ms per frame on the GX10.
Depth accuracy, 45 cm to 2 m
Depth to an item at six known distances from 45 cm to 2 m: a camera alone was off by 48% on average, lidar by 1%. Dots illustrate what that does to a reach.
First camera only: 52% accurate, the grey dots scatter. Then add lidar: 99%, the green dots land on the mug. Accuracy is 100% minus the average depth error over six known distances (45, 83, 100, 120, 150, 200 cm). Camera only read 37 to 72 cm at every one of them, so the farther the item, the worse it got: 64% off at 2 m. Lidar stayed within 2 cm.
Why local
Private
video of your home never leaves the room
128 GB
holds every model at once: YOLO-World, MediaPipe, Whisper, Nemotron 30B, Cosmos Reason 8B
35 ms
per frame, with no trip to a server
Offline
works with the internet unplugged
Private: a camera in someone's home never streams out. 128 GB unified memory keeps YOLO-World, MediaPipe, Whisper, the 30B Nemotron 3 Nano and the 8B Cosmos Reason model loaded together; most laptops can't hold the 30B model next to the rest. 35 ms: measured perception per frame on the GX10; guidance needs a steady loop and a cloud round trip adds varying delay. Offline: in the demo we unplug the GX10's internet and it keeps guiding, because nothing needs the network. Then thank the judges.
Proprio
Find it. Grab it. Eyes closed.
Thank the judges, then questions. Likely ones: how you measured the numbers (99/52 is depth accuracy at six known distances; published studies for the rest), why not run it on a laptop (privacy, all models together in 128 GB, offline), glass and shiny objects (lidar struggles; falls back to hand size).