Dynamic Neural Networks for Robot Systems: Data-Driven and Model-Based Applications

Navigating safely and efficiently in dense crowds remains a challenging problem for mobile robots. The interaction mechanisms involved in collision avoidance require robots to exhibit active and foresighted behaviors while understanding the crowd dynamics. Deep reinforcement learning methods have shown superior performance compared to model-based approaches. However, existing methods lack an intuitive and quantitative safety evaluation for agents, and they may potentially trap agents in local optima during training, hindering their ability to learn optimal strategies. In addition, sparse reward problems further compound these limitations. To address these challenges, we propose SafeCrowdNav, a comprehensive crowd navigation algorithm that emphasizes obstacle avoidance in complex environments. Our approach incorporates a safety evaluation function to quantitatively assess the current safety score and an intrinsic exploration reward to balance exploration and exploitation based on scene constraints. By combining prioritized experience replay and hindsight experience replay techniques, our model effectively learns the optimal navigation policy in crowded environments. Experimental outcomes reveal that our approach enables robots to improve crowd comprehension during navigation, resulting in reduced collision probabilities and shorter navigation times compared to state-of-the-art algorithms. Our code is available at https://github.com/Janet-xujing-1216/SafeCrowdNav.

Contact-rich robotic manipulation tasks such as assembly are widely studied due to their close relevance with social and manufacturing industries. Although the task is highly related to vision and force, current methods lack a unified mechanism to effectively fuse the two sensors. We consider coordinating multimodality from perception to control and propose a vision-force curriculum policy learning scheme to effectively fuse the features and generate policy. Experiments in simulations indicate the priorities of our method, which could insert pegs with 0.1 mm clearance. Furthermore, the system is generalizable to various initial configurations and unseen shapes, and it can be robustly transferred from simulation to reality without fine-tuning, showing the effectiveness and generalization of our proposed method. The experiment videos and code will be available at https://sites.google.com/view/vf-assembly.

Introduction: Since tracked mobile robot is a typical non-linear system, it has been a challenge to achieve the trajectory tracking of tracked mobile robots. A zeroing neural network is employed to control a tracked mobile robot to track the desired trajectory.

Methods: A new fractional exponential activation function is designed in this study, and the implicit derivative dynamic model of the tracked mobile robot is presented, termed finite-time convergence zeroing neural network. The proposed model is analyzed based on the Lyapunov stability theory, and the upper bound of the convergence time is given. In addition, the robustness of the finite-time convergence zeroing neural network model is investigated under different error disturbances.

Results and discussion: Numerical experiments of tracking an eight-shaped trajectory are conducted successfully, validating the proposed model for the trajectory tracking problem of tracked mobile robots. Comparative results validate the effectiveness and superiority of the proposed model for the kinematical resolution of tracked mobile robots even in a disturbance environment.