Computer Vision focuses on techniques to enable computers to extract and interpret information from the environment based on data from sensors, primarily through cameras. Currently, there is an abundant application of this field in autonomous or assistive mobile robotics. With a great practical appeal, robotics presents potential to aid in educational fields, including Computer Vision, by providing engaging hands-on learning. However, the use of robotics as a didactic tool is hindered by the price of robotic kits and the need of using proprietary software that requires a long learning curve. Additionally, it is necessary to have knowledge of low-level concepts such as motor control techniques, hardware assembly, and low-level coding. Therefore, this study aims to present a robotic system intended to serve as a tool that allows teachers and students to apply concepts in a practical, engaging and, most importantly, cost-effective manner. The solution features a camera-based robot platform and a high-level library, which encapsulates the communication with the robot and provides the user with the ability to control it through simple commands. The data, 3D models, and codes are publicly available at: https: //github.com/MaVILab-UFV/FPVCar_LARS-2025.
Large Language Model (LLM) Augmentation methods have introduced the capability of employing Agentic AI systems in Autonomous Vehicles (AVs) for autonomous navigation purposes. Such systems typically are implemented using cloud-based computing and are not exposed to the hardware limitations that embodied robotics suffers from. For this reason, we present a comparative analysis on augmentation methods for a real-time Agentic AI systems for AVs. In particular we: (1) evaluated the real-time feasibility of employing a single-board AI embedded computer for LLMs against consumer-graded hardware and (2) compared the accuracy of multiple LLM responses on spatial queries following two different augmentation methods, i.e., Retrieval-Augmented Memory for Embodied Robots (ReMEmbR) and Cache Augmented Generation (CAG). Our tests indicated that the Jetson Orin AGX embedded computer was incapable of generating real-time performance for position related queries, and the consumer-graded hardware demonstrated that models with a higher number of parameters seemed to improve precision when using CAG as the augmentation method.
This work presents a compact testbench that emulates tibiofemoral contact using a sphere-on-sphere knee surrogate. A primary electromagnet applies a constant 50 newton magnetic preload to maintain contact, and a magnetic release mechanism standardizes initial conditions with a synchronized release. A multi-camera system records six-degree-of-freedom motion at 100 hertz with submillimeter accuracy. A single holonomic surface contact point constrains the motion. The sole objective is data acquisition: register trajectories across repeated experiments and report the measured motion. Across trials, contact persisted, transients were repeatable, and trajectories converged to equilibrium. The motion is neither purely translational nor purely rotational; it is a combined trajectory with translation coupled to rotation, consistent with friction-induced tangential slip. Internal forces are not measured; however, the normal reaction and the tangential friction response act throughout the motion and are encoded in the kinematics. The dataset captures signatures of internal contact reactions and provides a basis for subsequent inverse-reconstruction studies under controlled, repeatable conditions.
This paper presents the design and implementation of a digital twin architecture for a four-wheeled omnidirectional robot competing in the RoboCup Small Size League, developed in the MATLAB/Simulink environment using Simscape Multibody. The framework integrates kinematic modeling, multibody dynamics, and realistic contact forces, while importing CAD models through the Simscape Multibody Link to preserve geometric and inertial accuracy. Experimental validation was performed using in-game situation trajectories from a public SSL dataset. Results show that the digital twin achieves trajectory root-mean-square errors on the order of 150-240 mm and maximum positional errors below 350 mm across all scenarios, while maintaining near realtime simulation performance. These findings confirm that the twin can reproduce real robot behavior even under dynamic conditions. This work presents an architecture for modeling omnidirectional robot dynamics and contact interactions within Simscape Multibody, providing a robust foundation for control validation, performance analysis, and parameter optimization.
Central Pattern Generators (CPG) are networks of oscillators that can generate rhythmic patterns inspired by biological systems. Their simplicity and adaptability have been demonstrated to perform well in the control of locomotion in robotic systems. Recently, fractional calculus has been explored as an alternative for modeling central pattern generators. In this work, we present the design and hardware implementation on a Field-Programmable Gate Array (FPGA) of a new fractional-order CPG model based on a modification of a Kuramoto-type phase oscillator with Grunwald-Letnikov-type fractional derivatives. For the validation of this proposal, the CPG was applied to the control of a robotic limb with three degrees of freedom (hip, knee, and ankle). The results obtained demonstrate that incorporation of a fractional order alpha enables more control over the generation of rhythmic patterns by directly influencing the amplitude and phase of the generated signals.
