health · digital · transport
reads well, struggles in crowds
Low-cost AI glasses made in India help blind users read and recognise faces, and cut navigation errors on a school walking course, but after a month at home only 22% of 85 users found the walking guidance useful, and an independent hospital study found it struggled in crowds
Source: Udayakumar et al., “Artificial intelligence-powered smart vision glasses for…” Open ↗source: primaryLast updated: 2026-11-17
Problem statement
India has about 4.8 to 5 million blind people (the 2025 paper gives 4.80 million in its abstract and 4.95 million in its introduction), and most AI assistive glasses sold abroad cost far more than they can pay. An Indian company, SHG Technologies of Bengaluru, makes "Smart Vision Glasses" (SVG): a camera, a LiDAR distance sensor and braille-coded touch keys clipped to a spectacle frame and paired with an Android phone, at about $360. They read text aloud, describe objects, recognise stored faces and offer "Walking Assistance", which announces obstacles and their distance. Three Indian studies tested them with blind and severely visually impaired people. Reading did well in both user studies, and the object and face functions did well in the Surat one. The walking function did worst in home use: after a month, 22.4% of 85 respondents in a five-centre study found it useful (against 72.9% for reading), and it was the function most often named as needing improvement. In a case series of 26 users at an eye hospital, run by optometrists from a Surat college with no stated link to the maker, the paper's table shows 11 of 26 rating the walking function poor or unsatisfactory (against 5 of 26 for reading), and participants said it "struggled with real-time obstacle detection in crowded environments"; in that study only currency recognition scored lower (19 of 26 poor or unsatisfactory). On a set school course, the same glasses used with a white cane cut navigation errors for 29 blind students and raised their confidence, though they walked more slowly, which the authors read as a more cautious strategy; one collision count fell, but a second collision row in the same tables rose outdoors. The 2025 authors and their users point to what needs to improve: detection that is not fast or accurate enough while walking, poor performance in low light, and speech cues that, the authors warn, may cover up the traffic sounds a blind pedestrian listens for. The unsolved problem is walking guidance from a low-cost wearable that blind people actually find worth using in streets and crowds, on top of the cane they already rely on.
Why this matters
The cane detects what is at ground level ahead; it gives limited information about distant hazards, spatial orientation or changing surroundings, and guide dogs are costly and "largely inaccessible in developing countries", as the trial paper notes. A cheap device that already earns its place for reading and faces could carry mobility help to many people who will never own a guide dog or a costly import, if its walking function worked well enough to use. Most published smart-glasses work, the trial authors write, is prototype or laboratory evaluation, with little evidence on commercially sold devices in real settings in India. Here reading is well used in both user studies, and faces and objects are well rated in the Surat one, so the gap is specific: it is the mobility function, the one where errors carry physical risk, that users set aside.
What’s been tried and why it hasn’t worked
(1) A month of home use (five centres; study run in 2022, published 2025). Ninety participants (mean age 23.5, range 6–56; 64% students; 57% rural, 51 of 90; 90% blind on the Indian scale, 98.9% on the WHO scale) from five vision-rehabilitation centres across India used the glasses for a month. All could operate every function, and 61% of respondents (52; the abstract says about two-thirds) used the device for an hour or more a day. Ratings of usefulness: Reading 72.9%, Things Around You 44.7%, Face Recognition 36.5%, Walking Assistance 22.4%; Walking Assistance topped the list of functions needing improvement (35.3%). Participants used the walking function mainly indoors, alongside their cane. The authors' overall conclusion is positive: the glasses are a "promising cost-effective device". No participant reported difficulty following the audio cues. The authors conclude that the glasses "should be used alongside tactile aids like a white cane for safety, as relying solely on obstacle detection and distance estimation is insufficient, especially on roads or in unfamiliar environments", that "the device may mask" auditory cues like traffic sounds, and that users asked for better speed and accuracy, better low-light performance and announcements of "distances to clear paths rather than obstacles". Limitations: all participants already used smartphones, follow-up was short, and the data are self-report. (2) A case series at an eye hospital, by optometrists from a Surat college (run September 2023–March 2024, published 2026). Twenty-six participants (mean age 29), recruited and studied by optometrists at a Surat college with no maker co-author, were trained in controlled and real-world settings and followed up at least two months later. Satisfaction with face recognition and reading rose (P < 0.05), and in the paper's table all 26 rated the device satisfactory or better overall (23 of them "very satisfactory"), but walking assistance was "rated less favorably": the table shows 11 of 26 rating it poor or unsatisfactory and 15 satisfactory or better (against 5 of 26 poor or unsatisfactory for reading and 1 for objects), and participants said it "struggled with real-time obstacle detection in crowded environments". Currency recognition fared worse still (19 of 26 poor or unsatisfactory). The paper's text gives different shares that its table does not bear out (38.5% satisfied with walking assistance, which matches the "unsatisfactory" ratings alone, 10 of 26; 88.5% rating it "satisfactory" or "outstanding" overall, which matches the 23 "very satisfactory" ratings), so the page uses the table's counts. Occasional disconnections were reported. The authors conclude that the glasses "show great potential" and that "improvements in walking assistance and currency identification are needed". (3) A crossover course test (two schools for the blind, Maharashtra, published 2026). Twenty-nine blind students aged 10–20 walked a 100 m indoor course and a 410 m outdoor loop with high kerbs, uneven pavement and pedestrians, each with the cane alone and with cane plus glasses, after 30–90 minutes of training. With the glasses, collisions fell (indoors 4.28 to 2.00; outdoors 1.62 to 0.62 per trial) and navigation errors roughly halved, and students reported more confidence; but completion time rose (indoors 229 s to 255 s; outdoors 578 s to 626 s; p < 0.001), which the authors read as "a more cautious and effective travel strategy". Their conclusion is that the glasses "can significantly improve mobility performance". They record audio dropouts and tracking delays, call for faster processing and better performance in low light and noisy outdoor settings, and say the device still needs testing at road crossings, with moving vehicles and in "crowded public spaces". The methods describe the cane-alone walks first and the glasses after training (they also say observers did not know the order, so the sequence is unclear), and the outdoor loop was walked in clear weather and daylight with observers close by. One row in the trial's tables points the other way, and the text does not discuss it for the outdoor loop: "number of obstacle collisions", which the methods define as "any physical contact with objects such as furniture, plants, walls, or door frames" and say was recorded as present or absent in each trial, went from 0.1 to 0.3 indoors (p = 0.096) and from 0.1 to 0.6 outdoors (p < 0.001) with the glasses. The trial authors also write that in both their study and the 2025 one the walking signals "were employed less frequently, indicating that dynamic guiding requires additional improvement". The authors of all three papers declare no conflicts of interest; the 2025 paper has an author from SHG Technologies' research department, and the trial an author from "Research, Smart Health Technologies Pvt. Ltd., Bengaluru" (and it thanks SHG staff for technical support). The Surat study has no maker-affiliated author. (4) Earlier tests of spoken against vibration cues (Canada 2019; United States 2017). With 12 blind adults walking indoor paths with their canes, a vibrating belt and spoken turn commands left recall of the simulated street sounds heard during each walk the same, and street noise slowed walking more under speech than under vibration, though the belt was slower in quiet and led to more missed or wrong turns (Bharadwaj et al. 2019); in a test where an investigator sent each cue by hand, 11 blind adults (10 of whom came back for the vibration test) were guided about equally well by speech through bone-conduction headphones and by a vibration vest and preferred speech, but the outdoor course was an empty parking lot with little ambient noise, and the authors say noisier settings such as street crossings still need testing (Adebiyi et al. 2017).
What would unlock progress
The gap is between a supervised course and everyday travel: on the course the walking function cut navigation errors, but in home use most users did not rate it useful, and the independent study found it struggling in crowds. Users' feedback and the 2025 authors' warning name the design levers: tell me where the clear path is, not every obstacle; do not talk over the traffic; respond fast enough for walking pace; work in poor light. Non-speech cues are one route: tones, spatialised sound or vibration (on the frame, a wristband or the cane) may carry less risk of masking than spoken announcements. A haptic "open path" mode already exists in research: Virtual Whiskers (Feng et al. 2025, New York) guides the user toward the most traversable direction with vibration, and in tests with 10 blind or low-vision participants it reduced hesitation and cane contacts. What has not been shown is such a cue on a device at this price, used by Indian users in crowds, or timed against speech for how quickly a listener notices an approaching vehicle (the 2019 belt study in attempt 4 measured hearing of street sounds only by what walkers recalled afterwards). A second route is to divide the work with the cane, as the 2025 authors suggest, so that the glasses add distance information beyond the cane's reach rather than duplicating it. Any change has to keep the price that makes the device reachable.
How we’d know it’s being solved
(a) In home-use studies like these, the share of users who rate walking assistance useful approaches that for reading (22.4% against 72.9% in the 2025 study; in the Surat study, poor or unsatisfactory ratings fall from 11 of 26 toward the 5 of 26 seen for reading). (b) Self-reported or logged use of the walking function outdoors and in crowds, not only indoors. (c) On course tests, the error gains of the 2026 trial are kept, neither of its two collision measures rises, and completion time is no longer slower than with the cane alone. (d) In listening tests, cues that do not reduce users' detection of approaching traffic or other environmental sounds compared with no device. (e) No rise in price beyond what users in India can pay (the device was about $360).
Where to start
Facility-free door (make a thing): obstacle-warning wearables and smart canes already exist (including these glasses), and a research haptic open-path aid that senses the way ahead itself (Virtual Whiskers) exists too. So the team does not build detection, AI or glasses, and does not remake Virtual Whiskers. The object is the cue alone, built from cheap parts: a wristband or headband with a few vibration motors, or short tones through open-ear (bone-conduction) headphones, that says "clear path this way" instead of naming each obstacle. A team member drives it by hand from a laptop or buttons (Wizard-of-Oz), so the cue can be judged apart from the sensing. Log the parts cost of each version against the device's own price (about $360). Then run a seated listening test with campus volunteers: play recorded street sound with an approaching vehicle and measure whether they detect the vehicle as fast while each cue type runs (vibration, tones, spoken alerts, no cue). Two earlier studies on the page come close and leave this open: one asked blind walkers afterwards which street sounds they had heard under a vibration belt or spoken turn commands (no difference), and the other sent speech and vibration cues by hand but not in street noise; neither timed how fast a listener notices an approaching vehicle while a cue runs, and neither tried tones. Everything fits in a makerspace; no partner or data is needed. User door: the real users are blind people in India, including rural users and school students, whom a campus team cannot reach. Blind and low-vision adults who travel with a cane near campus are a stand-in, reached through a campus disability office, a local blindness association or a vision-rehabilitation service, with an orientation-and-mobility instructor. Repeat the seated listening test with them, and in supervised indoor sessions on a marked course with a spotter beside them and their own cane always in use, compare the hand-driven clear-path cue with spoken obstacle announcements on how well they understand and trust each. No blind or blindfolded person walks near traffic or in a real crowd with the prototype: traffic is tested only as recorded sound played back, and outdoor or crowd trials are designed, not run, with the instructor's sign-off as an entry condition. What carries over from the stand-in: whether a cue is understood, trusted and masks traffic sound less than speech. What does not: Indian streets and crowds, the price users there can pay, and rural and under-18 users; those go into the trial design. Any session with blind students under 18 is design-the-test only. Testing with people needs the campus's ethics review. Relevant skills: interaction design, audio and haptic design, simple electronics, human factors.
Genes
Related problems
Stuck the same way, in other fields.
field: food & the living worldevery city counts its emissions differentlyUrban greenhouse gas source attribution gaptested in the wrong worldsource: primaryfield: food & the living worldcould be centuries, could be decadesThe timeline for West Antarctic ice sheet collapse is uncertain by orders of magnitudetested in the wrong worldsource: primaryfield: digital & spacewhen the guide robot freezesA suitcase-shaped guide robot for blind people at a Tokyo science museum often stops in crowds and queues; its builders have added spoken stop reasons and a robot "excuse me", but it still mixes up lines and crowds, and some bystanders do not understand its requeststested in the wrong worldsource: primaryfield: making & materialsbiology can't beat petroleum on priceBiological production of commodity molecules carries an inherent cost floor that chemical synthesis does nottested in the wrong worldsource: secondarySources
- Udayakumar D, Gopalakrishnan S, Raghuram A, Kartha A, Krishnan AK, Ramamirtham R, Muthangi R, Raju R. "Artificial intelligence-powered smart vision glasses for the visually impaired." Indian journal of ophthalmology 2025;73(Suppl 3):S492-S497, doi:10.4103/ijo.ijo_1621_24, accessed 2026-11-17. Link ↗
- Varshney AS, Chougle ME, Patel CV, Chauhan MD. "Evaluating usability of 'the smart vision glasses' for individuals who are visually impaired and totally blind." Saudi journal of ophthalmology 2026;40(1):42-47, doi:10.4103/sjopt.sjopt_241_24, accessed 2026-11-17. Link ↗
- Gogate P, Ghosh S, Naikodi M, Phadke S, Shitole A, Bajantri N, Banerjee S, Madan V. "Comparison of the Effectiveness of Smart Vision Glasses Combined With a White Cane Versus a White Cane Alone for Indoor and Outdoor Mobility in Blind Students." Cureus 2026;18(8):e114262, doi:10.7759/cureus.114262, accessed 2026-11-17. Link ↗
- Feng J, Hamilton-Fletcher G, Hudson TE, Beheshti M, Porfiri M, Rizzo JR. "Haptics-based, higher-order sensory substitution designed for object negotiation in blindness and low vision: Virtual Whiskers." Disability and Rehabilitation: Assistive Technology 2025;20(5):1433-1452, doi:10.1080/17483107.2025.2458112 (full text read in the NIH author manuscript, PMC12187558), accessed 2026-11-17. Link ↗
- Qin Y, Zhang Y, Desai N, Karimi HA, Huang KT, Du N. "Wayfinding and navigation approaches supporting mobility for people who are blind or visually Impaired: a scoping review." Disability and Rehabilitation: Assistive Technology 2026;21(7):3511-3536, doi:10.1080/17483107.2026.2685237 (abstract read via Europe PMC), accessed 2026-11-17. Link ↗
- Bharadwaj A, Shaw SB, Goldreich D. "Comparing Tactile to Auditory Guidance for Blind Individuals." Frontiers in human neuroscience 2019;13:443, doi:10.3389/fnhum.2019.00443, accessed 2026-11-17. Link ↗
- Adebiyi A, Sorrentino P, Bohlool S, Zhang C, Arditti M, Goodrich G, Weiland JD. "Assessment of feedback modalities for wearable visual aids in blind mobility." PloS one 2017;12(2):e0170531, doi:10.1371/journal.pone.0170531, accessed 2026-11-17. Link ↗
Revision notes
This brief gets checked against its sources and contains the following revisions. Future revisions get added here.
Worth knowing
- the school walking trial contradicts itself: one collision count fell, while a second measure of contact with objects rose outdoors; the page shows both
- one study has an author from the maker and a second has an author from a Bengaluru company that appears to be the maker; the hospital study states no such link
Full check
Sources read
- Udayakumar et al., home use of the AI glasses in India (Indian Journal of Ophthalmology, 2025): full text
- Varshney et al., hospital case series (Saudi Journal of Ophthalmology, 2026): full text
- Gogate et al., school walking-course trial (Cureus, 2026): full text
- Feng et al., Virtual Whiskers haptic guide (2025): full text
- Bharadwaj et al., vibration belt and spoken commands (2019): full text
- Adebiyi et al., spoken and vibration cues (2017): full text
- Qin et al., scoping review of wayfinding studies (2026): abstract only
What changed
- Stated more strongly than the source: reading did well in both user studies, but objects and faces only in the hospital study; in the home study faces and surroundings were rated useful by fewer than half; the home-study authors and users point to what needs to improve, not to the causes of the low walking rating; the hospital study's optometrists have no stated link to the maker, and the paper does not name the hospital's city; the trial's clear-daylight conditions applied to the outdoor loop only.
- Numbers: the trial's second collision measure, which rose outdoors, is now reported beside the collision counts that fell; daily use of an hour or more is 61% in the results, not about two-thirds as the abstract puts it.
- Credit: the warning that spoken cues may mask traffic sound is the authors', not a user suggestion.
- Dates: the trial is now dated by publication; the paper gives no dates for the walks.
- Title: the long title now says the glasses cut navigation errors on a school walking course, not collisions (display title unchanged).
- Entry points: the cue comparison starting point now says what two earlier cue studies left open: neither timed how fast a listener notices an approaching vehicle, and neither tried tones.
- Other: two earlier comparisons of spoken and vibration cues were added to the earlier attempts.
Left open
- that obstacle-detecting canes and other smart glasses are sold (rests on a search)
- that one trial author's company is the glasses' maker (no source says so outright)
Separate check of specific claims: India's count of blind people now gives both figures the 2025 paper prints (4.80 million and 4.95 million), and a heading no longer places the hospital study in a city the paper does not name.
Separate check of specific claims: A clause on what earlier cue studies measured was narrowed, the trial's maker link is now described as an inference, and two citations were added to the source list.
Earlier full check: The hospital study's satisfaction figures were restated from its own table, a cause-and-effect reading and an unsourced phrase about head-height obstacles were corrected, and two smaller wording fixes were made.
Separate check of specific claims: The make starting point now builds only the cue, driven by hand, and must not remake an existing haptic guide; the user starting point names stand-ins and keeps everyone indoors and away from traffic.
Earlier full check: The titles were corrected (display title 'Reads Well, Walks Badly' → 'Reads Well, Struggles in Crowds'), positive results are now stated first, the rural share and study dates were corrected, and an existing vibration-based path guide was added.