Robot Learning 2026 – Lecture 9: Generalist Robot Policies | ETH Zürich

Generalist robot policies aim to create versatile models capable of controlling diverse robot embodiments across various tasks and environments. Achieving this requires addressing three primary challenges: data scarcity, heterogeneous action spaces, and scalable evaluation. Leveraging large-scale, cross-embodied datasets like OpenX Embodiment allows for training transformer-based policies that outperform specialist models. Architectural innovations, such as tokenizing observations and actions, enable the application of multi-modal sequence modeling techniques from natural language processing to robotics. Furthermore, training Vision-Language-Action (VLA) models using internet-scale priors helps preserve semantic capabilities while facilitating real-time control. To mitigate the high costs and logistical difficulties of real-world testing, frameworks like Simpler provide correlated simulation-based evaluations, offering a reliable method to assess policy performance and generalization before physical deployment.