AMD EdgeAI4R Spotlight Awards

Keywords: Rat robot, SLAM, Low-texture environments, Hierarchical framework

TL;DR: This paper proposes a hierarchical SLAM framework that runs monocular ORB-SLAM remotely and combines coarse and fine matching strategies to enable stable mapping for rat robots in low-texture environments.

Abstract: Simultaneous Localization and Mapping (SLAM) on small quadruped robots, such as rat robots, is challenged by body vibrations, limited onboard computation, and low-texture environments. These factors cause sparse features, unstable viewpoints, and frame loss, which hinder conventional SLAM. To address these challenges, we redesign the rat robot with an external antenna to improve communication and a monocular camera to reduce bandwidth, while running monocular ORB-SLAM on a remote computer. In addition, we introduce a hierarchical SLAM framework that switches between coarse frame matching under sparse features and fine keyframe matching when features are sufficient. Experiments in a low-texture drainage channel show that our system maintains continuous mapping despite frequent interruptions, demonstrating feasibility for real-world deployment of rat-sized quadruped robots

Xueyang Bai [1], Bo Li [1], Wei Mao [1], Guoyong Shi [2], Genquan Han [1]

[1] Xi'an University of Electronic Science and Technology [2] Shanghai Jiao Tong University

Keywords: FeFET, motion control, robot computing, hardware-software co-design, computing-in-memory (CIM), autonomous system.

Abstract: In this paper, we present a Computing-in-Memory (CIM) linear equation solver circuit implemented in a ferroelectric field-effect transistor (FeFET) array which is used for robot arm motion control. Robotic motion control requires solving a forward/inverse problem described by a matrix-form second order differential equation. The acceleration vector can be obtained in real-time by solving a linear matrix equation. For real-time operation, non-CPU-based computation can reduce hardware cost and realize customized implementation. The proposed FeFET controller enables complex O(1) computation with a low-cost design. A software-hardware co-design method is presented, covering control problem formulation, solver array construction, circuit implementation, and system-level co-optimization. Simulation-based experimental results have demonstrated the feasibility of conceptual design, validating improved tolerance to solution error and power consumption comparing to the implementation based on other devices.

OpenReview Links, Talk Recordings, and Posters will be listed after the workshop