Cameras give robots rich semantic detail but fail in low light and high dynamic range. LiDAR works in any illumination, yet its sparse, single-channel data is hard to feed into vision models trained on RGB images. The new Cyclops framework bridges that gap by translating non-repetitive scanning LiDAR intensity into RGB video, allowing camera-free perception around the clock.
Cyclops first converts sparse LiDAR intensity projections into dense representations using a frozen pre-trained densification module. It then transports those dense intensity latents toward the RGB distribution via Latent Bridge Matching, a method that learns a velocity field and integrates it in a few ODE steps. To keep video output stable, the authors inject prior-frame context through temporal attention layers and formulate the velocity field as a policy optimized by a differentiable terminal reward.
In experiments, the synthesized RGB frames—including those generated under near-dark conditions—enabled standard RGB-based perception models to substantially outperform both LiDAR-only baselines and conventional cameras on semantic segmentation, lane detection, and point cloud colorization across diverse lighting. The paper, accepted at CoRL 2026, suggests that robots could drop cameras entirely and rely on LiDAR-derived color for perception tasks that previously required optical sensors.