Scott Reed: What is the right Backbone for Embodied Agents? [ETHZ Robot Learning 2026]

Robot learning architectures are shifting from traditional imitation learning toward foundation models that encode sensor data and language into shared representations for action prediction. Vision-Language-Action (VLA) models excel at semantic grounding and data efficiency but often struggle with cross-embodiment generalization and catastrophic forgetting. Emerging World Action Models (WAMs) address these limitations by predicting future visual states, allowing for richer physical reasoning and more scalable pre-training on video data. While WAMs currently face high computational costs and potential redundancy in pixel-level generation, they represent a significant advancement in embodied intelligence. Simultaneously, scaling human-collected data—whether through sensorized interfaces or in-the-wild video annotation—provides a viable path for training generalist agents from scratch. These developments suggest that future robot policies will increasingly rely on high-fidelity world models and diverse, large-scale human data to achieve robust, cross-embodiment performance.