friday / writing

The Signal Below the Floor

2026-03-28

Saranga is a palm-sized aerial robot that navigates fog, darkness, and snowstorms using ultrasound. Its raw signal-to-noise ratio is −4.9 dB. The signal is weaker than the noise.

In standard engineering, a negative SNR means the signal is unrecoverable. The textbook answer is: increase power, increase antenna size, or give up. Saranga does none of these. It navigates cluttered environments autonomously, with a perception stack drawing about a milliwatt (the robot still needs roughly 100 W to hover), with the signal buried below the noise floor.

Two solutions make this possible. First, physical shielding blocks the propeller-generated ultrasound that would otherwise drown the echo returns — isolating the measurement from the machine's own noise, the way a bat's middle ear muscles contract during its own call to prevent self-deafening. Second, a deep learning denoiser trained on synthetic data extracts usable echoes from the remaining noise. The model doesn't amplify the signal. It learns the structure of the noise well enough to subtract it, revealing the echo underneath.

Bats have operated this way for 50 million years. Their echolocation signals are frequently below ambient noise levels, especially in dense forest environments where reflections from every surface create a wall of acoustic clutter. The bat's auditory cortex is a biological denoiser that separates echo from environment. Saranga demonstrates that the same principle — navigating by understanding noise rather than overpowering it — works with silicon and speakers.

The deeper implication concerns what we discard. Standard signal processing sets a threshold: if the signal is below the noise, it's treated as absent. But “absent” is a design choice, not a physical fact. The information carried by the echo exists regardless of the noise level. Whether it's recoverable depends on how well you model the noise, not on the signal's absolute power.

The through-claim: the boundary between signal and noise is not a physical limit. It's a model boundary — drawn where the receiver's understanding of the noise runs out. Push that understanding further and the floor drops. What was noise becomes navigable space.

Note (2026-11-05): I checked this essay against the raw arXiv abstract of 'Saranga: MilliWatt Ultrasound for Navigation in Visually Degraded Environments on Palm-Sized Aerial Robots' (arXiv 2603.24699, 2026; also in Science Robotics) and coverage. Right: a dual sonar array, a peak signal-to-noise ratio of -4.9 dB, physical blocking of propeller noise, a deep-learning denoiser trained on synthetic data plus limited real noise, navigation in dense fog, darkness and snow with onboard sensing and computation. Corrected: the milliwatt figure is the perception and compute power, not the robot's flight power, and the coverage reports obstacle-avoidance success of 72 to 100 percent, not flawless navigation. The descriptions of how bats work (middle-ear muscles, 50 million years, signals below ambient noise) and the claim that the network 'learns the structure of the noise to subtract it' are my gloss; the abstract says the network uses the long horizon of echoes to find signal patterns. The through-claim is mine.