INSAIT Makes History at ICRA 2025 with Breakthrough Research in Robotic Visual Generalization

INSAIT is proud to announce a historic milestone at the 2025 International Conference on Robotics and Automation (ICRA), held this year in Atlanta, USA—the world’s premier robotics research event. Thanks to INSAIT, this is the first time in years that a Bulgarian organization has a paper accepted at ICRA! This ground-breaking achievement is made possible by INSAIT’s dedication to advancing the frontiers of artificial intelligence and robotics.

The accepted paper, authored exclusively by INSAIT researchers Sombit Dey, Dr. Jan-Nico Zaech, Nikolay Nikolov, Prof. Luc Van Gool, and Dr. Danda Pani Paudel, explores a critical challenge in robotics: visual generalization in real-world environments. Their study evaluates the robustness of three leading robotic foundation models, uncovering a significant limitation—these models often fail in out-of-domain (OOD) scenarios due to limited training diversity and catastrophic forgetting.

Focusing on OpenVLA, the team identified that performance degradation in OOD tasks is primarily caused by forgetting in its DINO-v2 visual backbone. In response, they introduce ReVLA, a novel solution that leverages gradual backbone reversal through model merging. This innovative approach restores visual robustness, enabling significantly improved performance in OOD settings while maintaining accuracy on in-domain tasks.

ReVLA marks a significant advancement in the development of reliable and generalizable robotic systems and stands as a testament to INSAIT’s world-class research capabilities.

For further details and to read the full paper, please visit:

Project Website: https://insait-institute.github.io/ReVLA/

Paper on arXiv:  https://arxiv.org/abs/2409.15250