Poster: Generalizable Image Repair for Robust Visual Control

Poster presented by Ivan Ruchkin at the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) on October 22, 2025 in Hangzhou, China.

Video: https://youtu.be/C3-WlZpYBm8

Paper: https://arxiv.org/abs/2503.05911

1. How can we repair corrupted images to increase the robustness of vision-based
controllers? How can we do so for unexpected corruptions?
v
We add a controller loss to existing image repair models CycleGAN (shown above)
and pix2pix to encourage image repair that benefits controller performance.
v
Although not all disturbances appear fully repaired to the eye test, the Root Mean-Squared
Cross-Track Errors (m, ↓) indicate that the repair models drastically improve performance.
Original Computer Vision CycleGAN pix2pix
Normal 1.545 2.957 1.242 1.648
Dark 3.062 3.148 1.353 1.403
Fog 1.135 1.969 1.087 1.415
Snow 2.118 1.923 1.833 2.003
Rain 2.257 2.088 1.392 1.451
Salt/
Pepper
1.571 2.459 1.345 1.552
v
Our repair architecture: we introduce an Image Repair Model (green) that
repairs images before they are used for control.
Generalizable Image Repair for
Robust Visual Control
Carson Sobolewski, Zhenjiang Mao, Kshitij Maruti Vejre,
Ivan Ruchkin
Department of Electrical and Computer Engineering, University of Florida
IROS 2025
v
Repaired images for each corruption: normal,
darkness, fog, snow, rain, and
salt/pepper noise.
v
Original images for different corruptions:
normal, darkness, fog, snow, rain,
and salt/pepper noise.