ZurichRobotics #1

After two years and over 25 events of NLP and CV we are finally launching our Robotics series! We have many exciting topics and speakers in the pipeline here, thank you for all your patience and support. For the first one:

Barnabas Gavin Cangan (ETHZ): Why Robot Hands are Hard: A Hands-On Perspective on Dexterous Manipulation Dexterous manipulation is a central challenge in robotics and embodied AI. Despite decades of progress, replicating the adaptability and sensory richness of the human hand remains elusive. In this talk, I will discuss the current state of the art in dexterous manipulation from two perspectives: the design and durability of robotic hands, and the use of learning-based methods for control. Drawing on our experiences in surgical robotics, I will share insights from building hardware for dexterous manipulation as well as lessons learned from applying learning-based approaches in practice. These perspectives highlight both the promise and the limits of current methods, and point toward where future progress may lie.

Caterina Caccavella (ZHAW, ETHZ): Bio-Inspired Active Sensing for Scene Understanding Conventional robotic perception often depends on large deep neural networks (DNNs), which are costly to deploy, slow to adapt, and prone to overconfidence. Most approaches also rely on passive datasets, overlooking the crucial perception–action coupling needed for real-world adaptability. We provide examples of how biology can inspire artificial systems that are not only more efficient but also more adaptable and reliable in complex, ever-changing environments. Inspired by biology, we propose an active perception strategy that maintains a coarse, continuously updated map of the environment and selectively focuses on the most relevant regions. Attention arises through bottom-up cues and competition via Dynamic Neural Fields (DNFs), while top-down mechanisms prevent repeated fixations and guide exploration based on task or uncertainty. This approach reduces reliance on exhaustive scene processing and large passive networks, leading to more efficient, adaptive, and reliable systems for dynamic environments.