Should we pursue a perfect multi-modal sensor fusion for maximum redundancy, as seen in systems with LiDAR, radar, cameras, etc.? Or should we champion the development of an advanced, vision-centric AI that can overcome the inherent ambiguities of the physical world through superior reasoning, an approach famously adopted by Tesla FSD? This question about the fundamental trade-off between hardware robustness and algorithmic intelligence is central to the future of autonomy of robotics perception and navigation in general, or what we called, "Spatial Intelligence".

As robots gain the ability to understand tasks through language and vision -- thanks to models like RT-2, PaLM-E, and LM-Nav -- we are now facing the deeper challenge: how do we translate this semantic understanding into reliable, executable control strategies? I believe what robotics lacks is a system-level abstraction that unifies perception, high-level policy reasoning, and low-level control execution across hardware platforms. My long-term goal is to build a modular, uncertainty-aware semantic control interface — akin to CUDA for GPUs — that enables developers to define goals like pick("red mug") or go_to("coffee machine") and deploy them on real mobile manipulation systems with minimal retuning.