A museum guide robot that will not guess — and a 2nd place nationwide.

Bangladesh’s Department of Archaeology manages 452 archaeological sites and operates roughly 17–21 museums. The Bangladesh National Museum alone reports approximately 15 million visitors a year, around 1.2 million of them foreign. Despite that, guided interpretation at regional and district-level institutions is minimal or absent — and the reason is structural rather than a lack of interest. A guide is a recurring salary a small institution cannot justify. Multilingual guides command higher pay, so the visitors least able to read a Bengali plaque get the least support. One guide serves one group at a time, and interpretation stops when a shift ends while the museum stays open.

When no guide is available, the visitor is left with a printed plaque, a pre-recorded audio track, or a phone app. All three share one limitation: they cannot be asked a question.

The knowledge is not scarce. Curators and historians already hold it, written down and verified. What is scarce is the mechanism that gets it to a visitor standing in front of an object at four o’clock on a Tuesday. This is a distribution problem, not a knowledge problem — and distribution problems respond well to engineering.

The entire vocabulary between the two boards

controller → computer   BOOTH_REACHED:<n>   a stop bar was confirmed; n = arrival count
controller → computer   CARD_UID:<hex>      a card was read, at any time
computer → controller   RESUME              begin driving — accepted only while stopped
computer → controller   STOP                unconditional halt, valid from any state

Four driven wheels, differential drive, two L298N dual H-bridges — one per side, one motor per channel to stay inside the roughly 2 A channel rating — and 12 V 300 RPM geared motors. Steering is a left/right speed difference, so there is no steering linkage at all. The motors are open-loop; the line is the feedback.

Why the economics follow from choosing RFID

Centralised inference is not a deployment detail. Each robot sends short audio and text rather than continuous video, so one modest GPU serves several units — and that is only possible because exhibits are identified by RFID rather than by camera. A vision-based guide would stream video per robot and could not be centralised. The identification decision and the economic model are the same decision.

Components for one robot come to about BDT 54,990 (roughly USD 458), excluding labour, assembly, testing, software and margin. Five robots plus the shared server come to about BDT 374,950 (roughly USD 3,125). The sixth robot costs a robot, not a robot plus an accelerator. Seven lines of that bill of materials are the team’s own estimates rather than checked retail prices, and the report flags each one individually.

Safety, wired ahead of the software

The main switch and the emergency stop are wired in series on the battery’s positive lead, ahead of every other component, so no software fault can keep the motors energised after the button is hit. The battery is charged under supervision; a protected pack with a BMS is treated as a prerequisite for unattended operation, not an optimisation.

Infrared reflectance sensors only read dependably within roughly 2–15 mm of the floor — a tolerance no deck plate holds by itself once a battery and four motors are bolted to it. All three LM393 modules therefore ride on one low, rigid, lightly sprung bracket ahead of the front axle, mechanically decoupled from the deck.

Two modules ride in the middle, one either side of the guide tape, so the tape passes between them and a drift darkens whichever module it reaches first. That geometry alone turns binary sensors into a signed error: the robot no longer asks whether it can see the line, it asks which of the two middle modules just found it, and corrects away from that one.

A third module sits outboard at the right-hand corner, clear of the tape. Its only job is stop detection. All three modules reading dark at once — and only all three — is the halt condition, because a full-width perpendicular bar is the only feature on the track that can darken the middle pair and the outboard module at the same instant. A narrow guide line structurally cannot. Following and stopping are separated by geometry rather than by threshold, so no separate stop sensor is needed.

on course       L ○   R ○   C ○   → tape between the pair → hold course
drifted right   L ●   R ○   C ○   → left module found tape → steer left
drifted left    L ○   R ●   C ○   → right module found tape → steer right
booth stop bar  L ●   R ●   C ●   → all three at once → confirm, then halt

The error is three-valued — left dark is −1, right dark is +1, neither dark is 0 — and the correction is proportional to it, with a ramp on persistence: the first dark read applies a small speed difference which grows the longer that module stays dark, so a brush against the tape is nudged while a real wander is turned. The ramp does an integral term’s job, clearing the standing offset four fixed wheels produce as they scrub to turn. The team does not claim a derivative term: on a signal that only steps between three values, its rate of change is an impulse, not a slope.

Each module is thresholded individually against the actual floor and tape under the actual gallery lighting. An LM393 compares its photodiode against a trimmer potentiometer, so this is a physical adjustment made in the venue rather than set at the factory. On Prototype 1 it was the difference between reliable and unreliable stop detection.

What this costs

Dropping the array costs resolution, and the team does not pretend otherwise: the robot knows which side it drifted to, not how far, so it corrects in steps and weaves marginally more on a long straight. If both middle modules read light it cannot tell centred from lost, and recovery falls back on the rotate-in-place ability the four-wheel base already has. What survives is what mattered: bar width is no longer readable, but bar count is — one bar means stop, two mean stop and expect a tap — so re-routing a gallery is still laying tape, not changing code.

The live caption bar was added so the robot could be understood in a noisy gallery. It is sized to be read at roughly two metres — large type, high contrast, two lines at most — because a caption nobody can read from where they are standing is decoration.

It also makes the robot usable by visitors who are deaf or hard of hearing, a group that pre-recorded audio guides exclude almost entirely. The same holds in reverse: spoken narration serves visitors with low vision or low literacy, for whom a dense wall panel is inaccessible. A single object explained simultaneously as speech and as on-screen text reaches substantially more people than either channel alone, and it costs nothing extra once the pipeline exists.

Accessibility and noise tolerance turned out to be the same feature — which is why captions survived a budget cut that removed almost everything else on the presence list.

MuseBot speaks English only today. Speech recognition and synthesis handle many languages natively, so hearing and speaking another language is nearly free. Being correct in another language is not. Machine-translating approved content at request time would silently reintroduce exactly the unverified-claim problem the grounding design exists to prevent, and a mistranslated cultural or religious term is precisely the kind of error a museum cannot afford. The retrieval and prompt layers are already keyed for per-language content; what is missing is a reviewed Bengali fact set for each exhibit, which is curatorial labour rather than code. It is deferred, and it is the deferred item the team most wants to close first.

And what is not built

The human-scale 150–170 cm body, the gesture arm, body lighting, a directional microphone, a protected battery pack with a BMS, and second-language content are all designed and costed but unfunded. None of them exists. MuseBot does not navigate freely, does not map a space, does not recognise objects by camera, and does not track individual visitors. Each of those is a deliberate choice this team defends rather than a gap it is hiding.

MuseBot AI took 2nd place — silver, Senior Future Innovators — at the WRO Bangladesh National Final 2026. It competes next at the WRO Asia-Pacific Open Championship in Hyderabad, India, 25–27 September 2026.

One thing worth stating plainly, because it is easy to imply otherwise: a WRO Open Championship is a regional international event, not a qualifying stage. WRO’s own rules are explicit that teams reach the International Final — San Juan, Puerto Rico, 8–10 December 2026 — only through the competition in their own country. Hyderabad is the next event on this team’s calendar, and the page will say more about Puerto Rico when there is something true to say.

Disclosure, per rule 6.5

AI appears in this project in two entirely different roles, and conflating them would be dishonest. Inside MuseBot, a language model turns museum-approved facts into natural spoken explanations and answers questions strictly within them — a presenter, never the source of truth. To build MuseBot, the team used Claude for design consultation, debugging assistance, code implementation under their direction, and language corrections to text they wrote. The team defined the behaviour, chose the architecture, performed the fault isolation, ran every test on hardware, and judged whether proposed solutions worked. Every hour of physical construction was theirs. It is the same “AI drafts, you verify” standard NebuCoders holds every project to.