Robot Learning 2026 – Lecture 11: Frontier & Open Problems | ETH Zürich

Robot learning currently faces a significant generalization gap, as impressive social media demos often fail when deployed outside narrow, controlled training distributions. Achieving true embodied intelligence requires moving beyond passive internet-scale data toward models that ground reasoning in physical, multimodal, and temporal experiences. Key open problems include identifying the optimal model backbone, developing scalable data collection interfaces, and enabling dexterous, mobile manipulation. Furthermore, reinforcement learning remains essential for stitching together suboptimal trajectories and fueling autonomous improvement. Research success depends on balancing method-driven and problem-driven approaches, prioritizing simple, scalable ideas, and maintaining rigorous introspection to identify model limitations. Ultimately, the field must transition from static, frozen policies to lifelong learning systems capable of rapid, in-context adaptation to novel, real-world tasks.