This paper addresses the problem of a synthesis of model predictive control for quadrupedal robots with adaptive change of weight matrices using a neural network. The robot model in the predictive control algorithm is considered as a single rigid body, which is affected by forces in the contact spots. The control method is a combination of a swing leg and ground force controllers with the use of whole-body impulse control for the latter. A fully connected neural network is applied for automatic tuning of weight coefficients used in convex model predictive control. The effectiveness of the proposed approach is demonstrated using numerical simulations, which provide a significant reduction in errors for various robot speeds.

Reinforcement Learning (RL) is increasingly transitioning from controlled simulation environments to real-world robotic applications. However, this shift presents a major challenge: designing effective reward functions and training procedures tailored to specific robotic tasks remains time-consuming and largely empirical. In this work-in-progress study, we investigate the impact of reward shaping and algorithm selection on robotic manipulation, focusing on two foundational skills: reaching and pushing. We benchmark five RL algorithms PPO, SAC, SAC-HER, DDPG, and DDPG-HER using a UR10 robotic arm in the PyBullet simulator. We compare sparse and dense reward formulations and analyze their influence on training dynamics and preliminary success rates. Our results demonstrate the effectiveness of hindsight-based approaches in sparse-reward scenarios and highlight the importance of proper reward design. Ongoing work includes extending the benchmark to more complex tasks such as sliding and pick-and-move, exploring hierarchical learning methods, and validating sim-toreal transfer on a physical UR10 platform.

In this paper, we assess the detection of the graspable edge of parts printed with powder-based methods, particularly HP Multi Jet Fusion (MJF). When extracted from the printer, the parts remain coated with unprocessed powder, and their unconventional shapes and white color make it challenging to identify pre-existing grasp points compared to everyday items. We introduce a pipeline that combines the YOLOv7 Convolutional Neural Network (CNN) with over-segmentation to detect a graspable edge on such printed parts. The Point Cloud Data (PCD) captured from a depth camera is used to validate the graspable edge, employing methods such as RANSAC and nearest neighbor search. Our method demonstrates the feasibility of edge pose detection for powder-based parts in an industrial processing line, where the excess unprocessed powder is removed automatically. The purpose of finding a graspable edge is to facilitate material handling and avoid obscuring the part while the gripper holds it during the inspection process. An image dataset from 12 different powder-printed parts was collected to train the CNN model. The dataset and scripts for this pipeline are accessible at https://github.com/thd-research/edge-grasp-pose-detection.

This paper is addressed to the problem of state estimation for permanent magnet synchronous motors (PMSMs) with uncertain parameters. The proposed observer of flux, rotor position and speed uses measurements of stator currents and control voltages only. Moreover, it is assumed that all the motor parameters are unknown. The designed estimation algorithm generates estimates of several parameters that are used by the state observer. Presented approach is based on motor model transformation that results in a linear regression model with unknown parameters and application of Kreisselmeier’s dynamic extension of the regressor with suitable mixing. It is shown that parameter and state estimates converge in a fixed time under some reasonable assumptions. Simulation results demonstrate efficiency of the proposed solution for a typical scenario of motor operation.

This paper presents experimental results of mobile robot navigation using two predictive controllers – a conventional model-predictive control and a Q-learning predictive controller. The latter essentially substitutes the running objective roll-outs with predicted action-value (Q-function) estimates. The idea behind such an approach is to integrate capabilities of reinforcement learning agents into the setting of model-predictive control while retaining the safety guarantees of the latter. Noteworthy the action sequence calculation step in both algorithms is of the same computational complexity. Yet, as we observed in our experiments, the learning predictive controller was able to outperform the model-predictive baseline. The code for the environment simulation may be found under https://github.com/thd-research/RL-autonomous-navigation.

Autonomous robots are becoming part of people’s daily lives in service applications such as vacuum cleaners, grass mowers, and food serving. One of the most recurrent challenges is to find a fitting path to move in an unknown environment. This student project aims to simulate this problem using the Turtlebot robot platform, inside an unknown maze. The main goal is the derivation of an autonomous exploration algorithm that allows the robot to eventually exit the maze, based on the assumption, that the robot has no preexisting knowledge about the maze. Additionally, to perceive his environment the robot only uses its integrated sensors, like LiDAR. The derived algorithms are implemented using two simulation environments, firstly a pure mathematical simulation implemented in Python and secondly as a Gazebo simulation, using ROS. In its final status, the robot can explore the maze probability-driven based on a type of occupancy grid. After finding the exit, the suboptimal path is determined using a wavefront algorithm in combination with a gradient descent method.

Traditional control approaches for robotic systems, such as linear quadratic regulator (LQR) or model predictive control (MPC), often rely on a known model of the environment. However, in the real world, uncertainty is a common feature of control problems hence models have imperfections. In this work, we address reward engineering for underactuated robotic systems with parameter uncertainty. We introduce a novel reinforcement learning (RL) method to plan the reward function, specifically designed for underactuated robotic systems with parameter uncertainty. We present and validate a new algorithm called Greedy-Divide and Conquer. We implement this algorithm with a single RL agent to address the challenge of swinging up and balancing a Pendubot system with uncertain parameters and give another example with a 2D-Drone with body mass uncertainty. Our ultimate objective is to enhance the system’s ability to adapt and perform reliably in the face of varying uncertainties.

The article aims to research the application of neural networks in signal adaptive filtering. The problem of filtering signals from noise and distortion is relevant in control systems. In this paper, white noise filtering using adaptive filters and neural networks is reviewed. Neural network algorithms were chosen to solve the problem of signal filtering. Neural networks and classical adaptive filtering algorithms, such as the least mean squares and the recursive least squares, were compared considering their efficiency for additive white noise filtering tasks. A multi harmonic signal was filtered from the additive white Gaussian noise using these approaches. As a result, classical adaptive filtering algorithms demonstrated better performance in signal filtering tasks.

Although the Adaptive Cruise Control (ACC) system is a safe and beneficial driving aid, it faces several performance challenges, primarily attributable to response delays in¬curred during acceleration command calculation and execution, as well as those inherent to the controller employed. This paper focuses on investigating the response of an ACC-based Model Predictive Control (MPC) system to a lane-changing vehicle while adjusting its velocity in the presence of another vehicle in the same lane. This scenario presents a significant challenge for the system and may lead to collisions. Therefore, this paper explores methods to enhance the performance of the ACC-based MPC system in such driving scenario, employing lane change predictors such as Fine K-Nearest Neighbor (FKNN), Optimizable k-Nearest Neighbor (OKNN), Fine Gaussian Support Vector Machine (FGSVM), and Fine Decision Tree (FT).

While the Adaptive Cruise Control (ACC) system is a useful and safe driving aid, it has a number of performance issues. These are mostly caused by response delays that occur during the calculation and execution of acceleration commands, as well as issues that are specific to the controller that is being used. This paper focuses on investigating the response of an ACC-based Model Predictive Control (MPC) system to a lane-changing vehicle while adjusting its velocity in the presence of another vehicle in the same lane. This scenario presents a significant challenge for the system and may lead to collisions. Therefore, this paper explores methods to enhance the performance of the ACC-based MPC system in this driving scenario, employing lane change predictors such as Fine K-Nearest Neighbor (Fine KNN), Wide Neural Network (WNN), Fine Gaussian Support Vector Machine (Fine Gaussian SVM), and Fine Decision Tree, in both straight and curved road configurations.

In this paper, the design of a lower limb active exoskeleton for rehabilitation purposes is discussed. Active exoskeletons provide the additional required force for motion and gait correction, especially for patients who have suffered from limb impairment. The development was carried out taking into account sensor input for the control system. A Finite-State Machine (FSM) enables the different motions which rely on the data from the torque and acceleration sensors. To replicate the anatomical knee rotation, a four-bar linkage was modelled and integrated in the actuator drive. For this project, a prototype was created using Additive Manufacturing Processes (AMP). The prototype is part of the ‘ForCEs' project at Deggendorf Institute of Technology (DIT) and was tested at an experimental stage for an initial set of results.

The control design problem for finite-time stabilization of linear descriptor systems is considered. The scheme of control parameters selection is presented by a linear matrix inequalities. Obtained stabilizability conditions are more reliable (linear matrix inequalities are not fragile) than the existing results. The settling time estimate is obtained. The theoretical results are supported by numerical simulations.

This Research to Practice Paper aims to illustrate how real-life projects can be used to teach students engaged in an academic course in robotic science the basics skills in control theory and computer vision. To design of modern control systems, engineers must be familiar with a range of technical fields. The range of theoretical knowledge and skills provided to the students include trajectory control, adaptive control, computer vision, as well as the capacity to build of complex mechatronic systems, all of which form part of an integrated curriculum. The paper describes how, using a LEGO educational set, the students are assigned a real-time project consisting in the development of a control system for a mobile car-typed robot. The stated objective is for the assembled robot to be able to move from a starting point to a goal pose while bypassing potential obstacles. At the end of the course, students must present their results, with a demonstration of the robot completing a given task together with a report. The project entails monitoring on the students progression and technical mastery as well as how the work process helps increase students interest in their studies and translate their theoretical knowledge into practice.