Remove the driver and you have not removed a control input. You have removed the entire negotiation layer that dense traffic actually runs on.

Multimodal Interaction is a studio course at NID Bengaluru on modes, modalities and representation systems. I took the autonomous vehicle brief into the densest street I had access to and treated the eHMI as a language problem rather than a display problem. The constraint I set was that the vocabulary must cost nothing to learn: anything a road user would have to be taught before their first encounter with the vehicle is, on a street like this one, a safety defect.

What came out is a system of six colours, six animation behaviours and four hardware elements, from which every scenario in the project is assembled, plus one scenario worked all the way down to what each channel is doing in each of thirty one seconds.

The deeper problem is not the geometry. It is that all three speak in a language the road user must acquire before the encounter is safe. That acquisition cost is invisible in a lab study where participants are briefed, and it is the entire risk on a street where the first encounter is also the only one.

So the vocabulary had to be auditable for learning cost. I used Rogers's icon types to classify every signal element as resemblance, analogical, exemplar, symbolic or arbitrary. Resemblance and symbolic elements that borrow from traffic lights cost nothing, because the road user already holds the reference. Arbitrary elements cost something and must be taught. The rule I set is that no arbitrary element may be load bearing for a safety decision.

That single constraint is what forced the colour lock, and it is why the only arbitrary signal in the finished system, the white beltline sweep, carries a non-critical meaning: the vehicle is alive and scanning. If a road user never learns it, nobody is hurt.

Music brackets the journey and never enters the hazard window. Green is spoken for and cannot be borrowed for a confirmation. This is the rule most concept work breaks, because every individual addition looks reasonable in isolation. Each extra use of a channel is a small withdrawal from its legibility, and the balance is only checked at the moment it matters.

Sound is placed at the hazard, arriving from over the occupant's left shoulder and moving as the hazard moves. Light is cast on the ground where the foot goes. A chevron is projected at the oncoming rider and flows toward him, because peripheral vision detects motion long before it resolves shape. A haptic pulse fires on the side the danger is on.

The reasoning is a time budget. Language costs a syllable at a time and the system has 3.4 seconds. Direction costs nothing. A sentence that names a hazard takes longer to parse than a sound that is already coming from it.

A note on the boards below, because it changes how they should be read. The scenario frames are generated images, used as scenario illustration. They are not photographs of a built vehicle and no part of this system has been fabricated or road-tested. What is being claimed is the signal design: which channel fires, in what order, and why.

The matrix runs six interaction cycles against six temporal phases, with eleven interaction dimensions as rows: posture, modality favourability, biometric signals, proxemics, gaze, head pose, voice, mid-air gesture, surface touch, foot controls, and system output across visual, auditory and haptic. Each cell answers one question: what is this channel doing at this moment. A dash means the channel is deliberately silent, and silence is recorded as a decision rather than left as an omission.

The constraint that generates the whole thing is that her hands are full and she has no picture of the street. The vehicle's modes exist to compensate for the modalities the street has taken away. Any cell that demands a free hand or sustained visual attention outside the vehicle is a defect, and writing it as a grid is what makes those defects visible.

Buckle press, voice and door control all express the same intent and any one is sufficient to request. None is sufficient to open. Only the corridor scan releases, and the scan is mandatory on every opening including one she explicitly asked for, because the door is never opened by the person who cannot see behind them.

The belt stays latched for a second reason. A released belt retracts into the reel and can no longer communicate, and it is needed as a channel eighteen seconds later, at the one moment in the journey when it matters.

Four channels fire at once and no single one is treated as sufficient for a safety commit. The voice says “Hold, bike coming up the wrong side. Three seconds.” Two things in that sentence are rules. Name the hazard, because “please wait” gets overridden and “bike coming up the wrong side” is checkable, and a checkable instruction is obeyed the next time. Name the duration, because amber with no number reads as a fault. The voice stays flat and unhurried: three channels already carry the alarm, and a fourth makes the cabin a panic.

The vehicle positions its body to deny the near-side filtering gap before any door logic runs, ranks candidate stop points with door side outranking walking distance, and shifts the landing zone 600 millimetres past a puddle before the occupant knows there was a decision to make. A gap that does not exist needs no warning, no light and no trust.

At release the door slides rearward to a measured 740 millimetres, the aperture sized to actual lateral clearance rather than defaulted. It tells her how much room she has by being how much room she has. Green and the mechanical click land on the same beat, the only mechanical event in thirty one seconds, and then the interface stops talking.

Most autonomous vehicle interface work is designed for streets where the rules are already doing the negotiating, and then assumes it will generalise downward into denser traffic. It does not. Starting from a street where negotiation is constant and faces are the medium produces a different system: fewer signals, locked meanings, direction instead of description, and geometry as the first line of safety rather than the last. That is a harder starting condition, and it is the one most of the world's traffic is actually in.