Robotics Requires More Than Policy Scaling for General Intelligence

The prevailing paradigm in generalist robot intelligence research, which largely frames the problem as one of policy scaling—collecting more demonstrations and training larger Vision-Language-Action (VLA) models—is fundamentally incomplete. This position paper argues that the central bottleneck is not merely the scale of policy learning but the absence of robust mechanisms to convert the world's abundant unstructured behavioral data into grounded robot supervision. While human motion, internet videos, and simulation rollouts contain rich information about tasks, goals, and physical constraints, this data is largely unusable by current robot policies due to a lack of embodiment-specific action labels, task semantics, and reward structures. This insight shifts the focus from simply increasing model size to developing foundational data infrastructure.

The context for this re-evaluation stems from the limitations observed in current robot learning approaches, where even large VLA models struggle to generalize effectively across diverse real-world scenarios. The paper identifies four critical missing components necessary for the next generation of robotics: data interfaces for autolabelling unstructured behavior, embodiment interfaces for retargeting human motion to robot actions, world-model interfaces for physics-grounded 3D reasoning, and reward interfaces for inferring task progress and success from multimodal inputs like video and language. These components are not incremental improvements but represent architectural necessities for robots to effectively leverage the vast, messy data streams available in the real world, moving beyond curated datasets to truly learn from observation and interaction.

The forward implications are profound for the trajectory of robotics research and development. By explicitly outlining these missing interfaces, the paper provides a clear research agenda that could unlock significant advancements in robot autonomy and generalization. Successfully developing these components would enable robots to acquire skills from a much broader range of sources, reducing the need for costly and time-consuming manual data collection and labeling. This would accelerate the development of more versatile and adaptable robots capable of performing complex tasks in unstructured environments, ultimately bringing generalist robot intelligence closer to reality by addressing the foundational challenge of data utilization rather than just model capacity.