Inertial sensors play a crucial role in robot state estimation because they can provide high-frequency 6-DOF motion measurements that are independent of environmental conditions. However, inertial navigation systems (INS) are prone to rapid drift due to inherent biases and noise, necessitating external aiding to constrain error accumulation. This presentation introduces several aided-INS approaches using low-cost IMUs for different types of robots, including a wheel-mounted IMU-based localization system (Wheel-INS) and SLAM system (Wheel-SLAM) for wheeled robots, a multiple leg-mounted IMUs-based state estimator (DogLegs) for legged robot, and an efficient LiDAR-inertial odometry (LIO-EKF) for generic robots.

The deployment of perception algorithms in real-world systems is inherently constrained by the capabilities of available hardware. As a result, their design has often evolved reactively, adapting to advances in sensing and computing platforms. This has led to a predominant reliance on frame-based visual sensors and centralized processing using CPUs and GPUs. While this approach has shaped much of the current algorithmic landscape, emerging sensing and processing paradigms are now challenging these assumptions and opening up exciting new directions in Spatial AI. In this talk, we will examine the reactive processing paradigm enabled by event cameras, whose sparse and asynchronous outputs give rise to low-latency, high-efficiency perception algorithms suited for real-time robotic applications. We will then delve into the distributed processing paradigm through Decentralized SLAM for multi-agent systems, discussing not only the locus of computation but also memory and communication patterns that support scalable, robust spatial understanding across agent networks. Finally, we will highlight an emerging sensor-processor paradigm that brings computation closer to the pixel level, reducing latency and energy consumption while eliminating the need for full-frame transmission.