digital · transport · health
when the guide robot freezes
A 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 requests
Source: Kamikubo et al., “Beyond Omakase: Designing Shared Control for Navigation…” Open ↗source: primaryLast updated: 2026-11-17
Problem statement
The "AI Suitcase" is an autonomous navigation robot for blind people: it looks like a wheeled suitcase, the user holds its handle, and it leads them to a destination while avoiding people and obstacles. It is built by a consortium of Miraikan (Japan's National Museum of Emerging Science and Innovation, Tokyo) with IBM, Alps Alpine, Omron and Shimizu, on Carnegie Mellon's open-source CaBot platform, and has run in daily trial use at Miraikan since April 2024. Its default mode is what the developers call "omakase", a Japanese word for leaving decisions to others: the robot decides and the user follows. In a 2025 study in the museum during opening hours, 13 blind people walked routes with the robot while actors blocked the way or stood in a queue. The robot's safety-first design "often causes the robot to stop" in crowds, a known robotics difficulty called the freezing robot problem. When participants left follow mode in those moments, they mostly did so first to ask what was going on rather than to give a command (89% of such switches when the path was blocked, 93% in queues). Some were unsettled by stops and zigzags they could not explain, and a touch sensor on the handle sometimes switched off when they were not consciously gripping it. In a 2026 three-week follow-up, the same group added spoken, obstacle-specific stop reasons and a button that makes the robot say "Excuse me" to bystanders. Two problems were left: the robot "often struggled to distinguish between lines and crowds, leaving participants unsure whether they should wait or initiate an interaction", and some visitors did not understand the robot's spoken requests (participants pointed to the museum's many foreign visitors). The unsolved problem is how a guide robot tells its blind user, and the strangers around it, what is happening when it stops in a crowd or queue, and how the user chooses what happens next.
Why this matters
Crowds and queues are two of the main places where blind travellers say they struggle. In the 2025 study's first round of interviews, 14 blind cane users described "streams of people" at train stations, hospitals, festivals and shopping districts, being swept along, tripping others with the cane or damaging it, and not knowing where a queue starts, whether it is moving or whether they have cut in. A robot that can avoid obstacles but often stops when people bunch up fails in exactly those places. Guide dogs are scarce in Japan (the paper cites about 1/10 of the US number, and only 2 of the 13 participants used guide dogs). Miraikan's own account (our translation) is that users valued being able to "choose exhibits freely and move on their own", and that running the robot in the museum is bringing out "challenges other than technical ones" for putting it into society.
What’s been tried and why it hasn’t worked
Almost all the evidence here comes from the robot's own builders (Miraikan, IBM Research Tokyo and CMU authors), and Miraikan's museum report is written by its own staff. One independent study points the same way: in Microsoft Research's Glide study (Ranganeni et al., HRI 2023, abstract read), nine blind or low-vision users' preferred level of control "varied in different situations; no single mode 'won'".
(1) Full autonomy (omakase). The robot has a LiDAR scanner, depth cameras, motorised wheels and a handle with a touch sensor and buttons. In the builders' field-trials paper on the AI Suitcase (Takagi et al., CHI EA 2025), the handle also has vibration devices at its top and both sides that warn of a turn before it happens and signal before the robot starts and stops; the paper does not say the vibration tells the user why the robot stopped, and its authors write that the robot still needs "the ability to provide information when the robot stops". It can travel to a destination and wait in line on its own. In the 2025 museum tasks, actors blocked the path on every route and formed queues at some destinations. When participants switched out of follow mode, most switches went first to asking for information (89% when the path was blocked, 76% when people crossed in front, 93% when people formed lines, 62% in other situations); the switches to asking in "other" situations were partly caused by "feedback and interface disruptions" such as delayed arrival sounds and the handle's touch sensor switching off. One participant said: "When you can't see and the robot makes some strange movements, it can make you feel uneasy, like, 'What's happening?'" Nine of 13 agreed they would follow the robot given its current abilities, yet all also wanted the modes that let them ask about their surroundings and give commands; 10 of 13 preferred to ask bystanders questions themselves (such as where a line ends) rather than have the robot do it, and 7 of 13 preferred to alert bystanders themselves (such as saying "Excuse me") rather than have the robot do it. In that study the robot's answers were given by an experimenter walking with the participant (a Wizard-of-Oz set-up).
(2) Spoken stop reasons and robot-mediated requests (2026 follow-up). Six blind participants used an upgraded robot weekly for three weeks in a museum during opening hours (the paper does not name the museum; three of its five authors are at Miraikan). When it stopped, a camera image went to GPT-4o and a neck speaker told the user why ("There is a crowd ahead with about three people, facing away"); a handle button made the robot's external speaker ask people to move. This helped, but did not close the problem. The robot often could not tell a line from a crowd, so users did not know whether to wait or act. Some visitors did not understand the robot's spoken requests, and participants noted that the museum had many foreign visitors; they suggested switching languages or adding non-verbal signals such as beeps. Interactive questions were left out because language-model delays would have been noticeable while walking. The follow-up also changed the first study's picture of who should speak: over repeated use several participants handed bystander requests to the robot, especially in noisy crowds where their own voices did not carry, though one participant still preferred to act on his own when he could (all six handed some requests to the robot). A third study by the same group met the freeze again: in group tours of a science museum with eight blind participants and a tele-operated robot (Hu et al., CHI 2026), the robot "got stuck from time to time in several tours", "primarily caused by crowds in the museum", and could also take the user's own feet or cane for an obstacle.
(3) More robot capability. A 2025 model redesigned the body, added a wheel mechanism to climb steps, sensors for low obstacles, a disc-shaped direction indicator in the handle, and an AI voice that announces buildings, road conditions and pedestrians; it was planned for multi-unit trials at the 2025 Osaka Expo (Impress Watch, 22 January 2025, in Japanese). These address what the robot can sense and say to its user, not how strangers read it. The authors say research on shared control for blind navigation is "still neither active nor sufficient", and that the freezing robot problem itself is "an ongoing challenge". Two guides sold or announced elsewhere also talk mainly to their user. Glidance's Glide, a wheeled guide that steers while the user pushes, gives "tactile, haptic, and audio cues"; its home page describes no signal to bystanders. A 2026 talking robot guide dog from Binghamton University describes the scene to its handler, but its paper does not deal with stops, crowds or bystanders, and in the test the robot's movements were controlled remotely by an operator (a Wizard-of-Oz set-up).
(4) Sound aimed at pedestrians (2019). The same research group already built an outward sound. BBeep (Kayukawa et al., CHI 2019) is a suitcase that tracks pedestrians and beeps through a speaker only when it predicts a collision. With six blind users in crowded areas of Pittsburgh International Airport, it "reduces the number of situations of imminent collision risk when compared to notifying the blind user alone", and when the beep fired mattered more than which beep it was. It left three things open: it sometimes beeped at people who had already noticed the user; the authors did not study what sighted pedestrians thought of it; and users were hesitant to use it in quiet places such as hospitals or libraries. Its routes were straight lines through crowded areas near the gates, where pedestrians were often talking in groups or standing in line, and with the speaker on they cleared the path; but the study focused only on clearing a path, not on helping a user find or join a queue, and it did not test whether bystanders understood what the beep meant.
What would unlock progress
Treat the robot and its user like a guide dog team, a parallel the 2025 paper draws itself: the handler gives orientation and commands, and the dog handles the path. That moves the open design work from the robot's planner to two interfaces. The first faces the user: a fast, non-visual cue that separates "blocked by a standing group", "end of a queue" and "people crossing", that stays honest when the robot is unsure whether it is looking at a line or a crowd, and that lets the user answer "go forward", "wait", "go round" or "ask them" without speech that competes with the sounds blind travellers rely on. The second faces bystanders: a signal that strangers, including people who do not speak the local language or have their backs turned, understand as "a guide robot with a blind person is coming through" (a 2025 participant proposed the robot announce "AI suitcase is passing through", noting that a plain alert may be mistaken for a smartphone). Because users differ (some delegate to the robot from the start, some come to it over weeks, some mostly in noise or crowds), both interfaces need to offer a choice of who speaks, not fix one answer.
How we’d know it’s being solved
(a) Users can tell the reason for a stop (blocked, crossing, queue end, or "unsure") from the cue alone, measured as correct identification in trials, and they learn when the robot is uncertain rather than being told a wrong reason. (b) In crowd and queue scenarios like the studies', the time the user spends stopped and unsure falls. (c) Bystanders, including people who do not speak the local language, step aside or let the user join a queue after the robot's signal at a higher rate than after a spoken request alone, and can say afterwards what the signal meant. (d) Users can issue the commands they want (move forward, alert people, wait) without letting go of the handle, and these gains hold over weeks of use, not only at first contact.
Where to start
Facility-free door (bystander signal, no robot or blind participants needed): fit an ordinary wheeled suitcase or rolling cart, pushed by a team member, with a small speaker, light strip or flag, and design two or three candidate "guide robot coming through" signals (a spoken phrase in two languages, a distinctive tone, a light pattern, a printed sign). The builders' robot already speaks a fixed "Excuse me, please move", so a plain one-language spoken request is the baseline each new signal must beat, not a design to rebuild; switching languages and beeps were suggested by the 2026 participants but not built on that robot. A tone alone is not new either: BBeep (2019) already showed that a well-timed beep from a suitcase clears pedestrians from a blind user's path, including people talking in groups or standing in line at airport gates, so a tone candidate must add what BBeep did not test: that passers-by can say what it means, and that it helps the user find or join a queue rather than only clear the way. In a campus corridor or building lobby, test with sighted passers-by, including students who are not native speakers, whether they notice the signal, understand it, and move aside or make room in a queue, with backs turned and facing; record the response and a one-question follow-up. No one is deceived or obstructed: the cart moves at walking pace, stops rather than pushes through, is never left across a doorway or a real queue, and the follow-up tells each person it was a student mock-up with no blind traveller involved. Observing and approaching strangers needs the campus ethics office's view (review or a written exemption) before the first session. Facility-free door (user-side cue, bench first): build a 3D-printed handle with two or three vibration motors and a few large tactile buttons, driven by a microcontroller, and design a small vocabulary of haptic codes ("stopped: blocked", "stopped: crossing", "queue ahead", "not sure: line or crowd") and button replies ("forward", "wait", "you ask", "I'll ask"). Test whether sighted, blindfolded volunteers can learn and tell the codes apart while seated, before any walking. Blindfolded volunteers are stand-ins: they have no practice of travelling without sight and do not lean on sound and the cane the way blind travellers do, so the bench test can only weed out codes that feel alike; whether blind users read the codes quickly, trust an "unsure" code, or want touch at all is a question for the user door. CaBot and the 2025 AI Suitcase already have handles with touch sensors, buttons and a direction disc, and the AI Suitcase handle in the builders' field trials already vibrates before turns, starts and stops (Takagi et al. 2025), so the team starts from that vibration and adds what the builders' handles are not shown to do: give the stop reason, including uncertainty, by touch. User door: blind and low-vision cane users and guide-dog handlers can be reached through a campus disability office, a local blindness association or a vision-rehabilitation service, with an orientation-and-mobility instructor. Run supervised indoor sessions on a clear, flat corridor, with the cart pushed and steered by a team member (a Wizard-of-Oz set-up; the 2025 study used one for the robot's spoken answers), who stops it on cue so participants can practise reading the stop codes and replying, and can say which bystander signal they would want used on their behalf. No participant is led into crowds, stairs or traffic; crowd and queue trials in a public building are designed, not run, with the instructor's sign-off as an entry condition. Testing with people needs the campus's ethics review. Relevant skills: interaction and product design, sound and haptic design, embedded electronics, human-robot interaction.
Genes
Related problems
Stuck the same way, in other fields.
field: food & the living worldsame circuit, different organism, different outputEngineered genetic circuits behave differently in every organism they are transferred totested in the wrong worldsource: primaryfield: making & materialsevery truckload of fly ash is a surpriseLow-carbon cement replacements have unpredictable batch-to-batch qualitytested in the wrong worldsource: secondaryfield: health & medicineease the wrist, load the palmIn two small lab tests, white-cane redesigns each eased some hand or wrist load but raised it elsewhere: in Iran an angled handle eased the wrist but loaded the palm and scored lower on usability, and in Japan a half-weight cane cut load on four forearm muscles but raised a fifth by about 30%tested in the wrong worldsource: primaryfield: making & materialsstrong or charged, not yet bothStructural battery composites do not yet store energy at high density while bearing loadtested in the wrong worldsource: secondarySources
- Kamikubo R, Kayukawa S, Kaniwa Y, Wang A, Kacorri H, Takagi H, Asakawa C. "Beyond Omakase: Designing Shared Control for Navigation Robots with Blind People." Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), Yokohama, 2025, pp. 1-17, doi:10.1145/3706598.3714112 (full text read in the PMC author manuscript, and in arXiv v1 and v2), accessed 2026-11-17. Link ↗ Link 2 ↗
- Hata R, Kuribayashi M, Wang A, Takagi H, Asakawa C. "How Does Delegation in Social Interaction Evolve Over Time? Navigation with a Robot for Blind People." Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), doi:10.1145/3772318.3791439 (full text), accessed 2026-11-17. Link ↗
- 澤田拓実「人とテクノロジーの両輪で科学館体験のアクセシビリティを考える」『博物館研究』Vol.60 No.11(No.690), 2025, pp.17-20 (Sawada T. "Thinking about the accessibility of the science-museum experience with both people and technology." Hakubutsukan Kenkyu [Museum Studies]), Japanese original hosted by Miraikan, accessed 2026-11-17. Link ↗
- Hu Y, Kuribayashi M, Wang A, Kayukawa S, Sato D, Mutlu B, Takagi H, Asakawa C. "Robot-Assisted Group Tours for Blind People." Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), doi:10.1145/3772318.3790425 (full text), accessed 2026-11-17. Link ↗
- Kayukawa S, Higuchi K, Guerreiro J, et al. "BBeep: A Sonic Collision Avoidance System for Blind Travellers and Nearby Pedestrians." Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI '19), Glasgow, 2019, pp. 1-12, doi:10.1145/3290605.3300282 (full text read in the author version), accessed 2026-11-17. Link ↗
- Ranganeni V, Sinclair M, Ofek E, et al. "Exploring Levels of Control for a Navigation Assistant for Blind Travelers." Proceedings of the 2023 ACM/IEEE International Conference on Human-Robot Interaction (HRI '23), Stockholm, 2023, pp. 4-12, doi:10.1145/3568162.3578630 (arXiv:2301.02336; abstract), accessed 2026-11-17. Link ↗
- Hayamizu Y, DeFazio D, Mehta H, et al. "From Woofs to Words: Towards Intelligent Robotic Guide Dogs with Verbal Communication." arXiv preprint arXiv:2603.12574v1, 2026 (full text), accessed 2026-11-17. Link ↗
- Glidance, "AI Mobility Aid for Blind & Low Vision | Glidance" (maker's home page), accessed 2026-11-17. Link ↗
- Takagi H, Naito K, Sato D, Murata M, Kayukawa S, Asakawa C. "Field Trials of Autonomous Navigation Robot for Visually Impaired People." Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA '25), Yokohama, 2025, pp. 1-8, doi:10.1145/3706599.3706663 (author version), accessed 2026-11-17. Link ↗
- Impress Watch, 「視聴覚障がい者の移動を助ける"AIスーツケース" 大阪万博で実証へ」 (news article, in Japanese, 2025-01-22), 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 main studies of this robot come from its own builders; one independent study points the same way
Full check
Sources read
- Kamikubo et al., guide robot in crowds and queues (CHI, 2025): full text (two preprint versions and the author manuscript read; the publisher's version was not opened)
- Hata et al., spoken stop reasons and a robot 'excuse me' (CHI, 2026): full text
- Takagi et al., field trials of the guide robot (CHI Extended Abstracts, 2025): full text
- Sawada, museum report on the guide robot (2025, in Japanese): full text
- Impress Watch news article on the 2025 model (in Japanese): full text
- Kayukawa et al., BBeep collision-avoidance beep (CHI, 2019): full text
- Hayamizu et al., robotic guide dogs with speech (preprint): full text
- Glidance home page: full text (home page only)
- Ranganeni et al., levels of control for a navigation assistant (HRI, 2023): abstract only
- IPSJ magazine paper on the robot (2022, in Japanese): record only
What changed
- Not found in the sources, removed: that a participant's quotation was translated by the authors; the paper does not say so.
- Not found in the sources, replaced with what they say: the 2026 follow-up was said to run in the same museum; the paper does not name its museum, and the page now says so.
- Stated more strongly than the source: the robot 'often stops' when people bunch up, not whenever they do; the statement about guides that talk mainly to their user now rests on its two named examples; only Glidance's home page was read, and the page now says so.
- Numbers: 10 of 13 preferred to ask bystanders questions themselves; the page now adds that 7 of 13 preferred to alert bystanders themselves, the kind of request the newer robot makes.
- Entry points: the handle starting point now begins from the vibration the builders' handle already gives before turns, starts and stops, and adds what it is not shown to do: give the reason for a stop.
- Other: a sentence on that handle was added to the earlier attempts: a field-trial user felt uneasy when the robot stopped without notice.
Left open
None noted in this check.
Separate check of specific claims: A third study by the same group, in which the robot again got stuck in crowds during group tours, was added, and citations were added to the source list.
Earlier full check: Several statements were brought back to the papers: who misunderstood the robot, how all six users came to hand some requests to the robot, what the earlier beep study did and left out, and how the robot's answers were simulated in the earlier studies.
Separate check of specific claims: A search added an earlier outward beep for pedestrians from the same group, a two-wheeled guide and a robot guide dog, and the bystander starting point now builds on the beep.
Separate check of specific claims: The bystander starting point now names the builders' spoken request as the baseline, with rules so no passer-by is misled or blocked and an ethics check first; the bench starting point says what sighted stand-ins cannot show.
Earlier full check: The 89/76/93/62% figures were corrected to what they measure, the long title now names the builders' remaining problems after a newer follow-up, the source type now says the main studies come from the builders, and a citation venue was corrected.