Perception is
complementary.

One simulated scene, read by many sensor models. Blur, photon noise, event thresholds, phase wrapping and heat — applied to a world whose annotations never move.

01 Conventional RGB 40 ms exposure · motion blur

Unified perception requires diverse sensing

The same room and the same instant can be perceived differently depending on the sensor. Which one is right?

A single exposure is a compromise

A conventional camera selects a single exposure for the entire frame, which is challenging in this room: the lamp is about four stops brighter than the unlit corners. As the lamp turns on, the camera must re-meter. If it exposes for the table, the lamp becomes overexposed; if it exposes for the lamp, the table appears noisy and underexposed.

conventional · **0 EV** · metered on the dark frame

The unlit table, grain through the shadows — before · +3.2 EV · noisy

The lit table, lampshade washed out — after · −0.8 EV · saturated

One bit per pixel captures the highlights and shadows

A single-photon camera detects the presence or absence of one or more photons per pixel, so each frame carries one bit per pixel and, on its own, looks like noise. Summing eight frames, however, lets the room emerge. With more summed frames and by inverting the camera's response function, both the unlit table and the illuminated lampshade become visible.

Cushions recovered from the count — before · pillow · +1.9 EV

The lampshade, held open — after · lamp · −9.1 EV

When sparse updates saturate the bandwidth

In contrast, an event camera transmits only changes: each pixel reports when the light falling on it varies beyond a certain threshold. A still room therefore produces almost no signal, while the sensor responds rapidly to motion and edges. When the lamp switches on, that change occurs everywhere at once, causing the entire array to flood in a single instant.

The lamp region saturated with events — at the switch

Seeing through darkness and through time

A thermal camera captures long-wave infrared radiation instead of reflected visible light, so turning on a lamp does not instantly illuminate the scene. This also means the camera retains a kind of thermal memory: the floor remains about 0.2 K warmer where a foot landed, even seconds after the robot has moved on.

Two prints behind the feet — the prints, thermal radiance

Every annotation, for every view.

A single scene is rendered once with dense, pixel-perfect, temporally synchronised ground truth, so any downstream task learns from exact labels.

Batteries included.

It comes with a large-scale dataset rendered through this pipeline: 100M raw data points from eight sensors, across 100 indoor scenes, with 480k synchronized ground-truth views. Coming soon. The prerelease is available in the meantime.