Enabling the Robotic Revolution: Bridging Performance Gap Between Present and Future

Abstract

The robotics industry has reached an inflection point due to recent advancements in robotics foundation models, which enable general-purpose action generation in physical environments. However, a significant performance gap remains between these models and the constraints of edge hardware. This paper characterizes foundation models for robotics and analyzes hardware bottlenecks across three critical dimensions: model architecture diversity, real-time latency constraints, and power and thermal limits. By analyzing state-of-the-art models such as Physical Intelligence’s pi0, NVIDIA GR00T N1.6, and NVIDIA DreamZero on modern platforms, we show that smooth and responsive humanoid dexterity under stringent conditions requires substantially more compute throughput, memory bandwidth, and power efficiency than current systems provide. We conclude by calling on the VLSI community to help provide the innovations needed to move from research prototypes to ubiquitous, reliable humanoid deployment.