SafeSight
A tiny drone that knows when its vision is lying.
SafeSight studies whether a small autonomous robot can recognize when stereo depth is unreliable and choose safer behavior—slowing down, re-scanning, or stopping—before acting on bad perception. The drone is the demonstration platform; the broader contribution is an interpretable trust layer for constrained autonomous robots.
Can inexpensive stereo diagnostics identify unsafe depth estimates accurately enough to guide safer robot behavior under tight compute constraints?
Valid-disparity density, disparity variation, left-right inconsistency, edge mismatch, and blur, glare, or lighting proxies.
The robot changes behavior instead of blindly trusting a single depth estimate.
Results after removing invalid-disparity and border-artifact shortcuts.
These results show meaningful independent signal, but the system is not yet a calibrated probability model. The next phase tests simplified trust gates on a controlled physical mini-dataset containing drone-relevant failure modes.
