Self-Supervised Active Failure Discovery in Collaborative Human-Robot Vision

Most collaborative robotics systems (Cohen et al., 2025) focus on improving recognition in standard scenarios, but rarely do they “go looking” for their own blind spots. Drawing from cognitive psychology (Prahara et al., 2020; curiosity-driven attention) and active learning, this idea proposes a vision model that tracks its own uncertainty and failure cases (e.g., misclassifications, low-confidence predictions), actively seeks out situations or environments where it’s likely to fail (e.g., rare object poses, novel lighting), and uses self-supervision and human-in-the-loop feedback to efficiently learn from these “hard” cases. This approach not only improves robustness but also provides a framework for life-long learning in robotic vision. The novelty is in flipping the paradigm: rather than passively “fixing” failures after deployment, the system proactively hunts for and learns from them, much like a curious human or animal.

References:

1. Fusion of Computer Vision and AI in Collaborative Robotics: A Review and Future Prospects. Yuval Cohen, Amir Biton, S. Shoval (2025). Applied Sciences.
2. Bottom-up visual attention model for still image: a preliminary study. A. Prahara, Murinto Murinto, Dewi Pramudi Ismi (2020).