Achieving Human-Like Movements with Neural Network-Based Planner in Collaborative Robotics

Achieving Human-Like Movements with Neural Network-Based Planner in Collaborative Robotics Federico Lozer University of Udine Udine, Italy [email protected] Lorenzo Scalera University of Udine Udine, Italy [email protected] Alessandro Gasparetto University of Udine Udine, Italy [email protected] Mathias Brandst¨ otter ADMiRE Research Center Villach, Austria [email protected] Abstract—This work presents a neural network-based motion planning approach designed to allow redundant robotic manipulators to emulate the arm movements of a human subject, by selecting the optimal kinematic configuration. The proposed approach is developed with the aim of improving humanrobot collaboration in terms of acceptability and trustability. Experimental results on a robot with 7 degrees of freedom demonstrate the feasibility and the effectiveness of the proposed approach, in correctly replicating the motion of a human arm. Index Terms—trajectory planning, neural network, artificial intelligence, collaborative robotics, redundancy I. INTRODUCTION Collaborative robotics is increasingly relevant in Industry 5.0, aiming to integrate robotic manipulators with human workers to enhance productivity [1]. A key challenge is motion planning, especially for tasks requiring human-level dexterity or complex trajectories [2]. Traditional methods like optimization fall short in robustness and stability. To address this, recent approaches in collaborative robotics integrate advanced trajectory planning techniques, such as neural networks, valued for their adaptability and learning capabilities. Neural networks (NNs) are powerful tools for robot control and motion planning, especially in human-robot collaboration tasks. They enable adaptive behavior based on human imitation abilities [3]. NNs are also applied in action recognition and future motion prediction using recurrent architectures [4]. In motion planning, NNs emulate human behavior for manipulators and robotic hands, using methods like deep residual networks and learning from demonstration [5]. Despite these advances, no current work fully exploits robotic arm redundancy to generate optimal human-like motion. In this work, we propose an approach based on a self-trained artificial neural network, addressing the challenge of planning human-like trajectories. Human-friendly motion improves robot acceptability and trustability [6], making movements easier for people to predict. Since the human arm could be seen as a 7-DOF system (Fig. 1), it shares kinematic configuration with certain manipulators, making it feasible for robots to mimic human arm configurations and behavior. Traditional planning methods have limited effectiveness, so, in our approach, we use a neural network to obtain the desired kinematic configuration for a certain end-effector pose and solve the robot’s redundancy in an optimal way. Redundancy Hand J7 q7 EE y x z y x z y x z y x z q7 J7 L M N O Fig. 1: Robot end-effector and human hand kinematics. enhances dexterity but introduces complexity due to infinite IK solutions for a given pose. Analytical IK solutions are often not feasible without added constraints. A common approach is to optimize a specific redundant joint value, reducing degrees of freedom and limiting the number of solutions. The proposed approach focuses on using a redundant manipulator to emulate human motion, and investigates robot redundancy optimization [7], selecting the most suitable inverse kinematics (IK) solutions of a seven-degree-of-freedom (DOF) manipulator to emulate human movements. This method is tested on a 7-DOF Franka Emika Panda robot both in simulation and experimentally. II. PROPOSED APPROACH The strategy proposed in this work combines the analytical method for the inverse kinematics computation presented in [8] with an artificial neural network, trained on a custom-built dataset, to determine the optimal value of the redundant joint (the 7th) of a robotic manipulator with 7 DOFs to reproduce the motion of a human arm. More in detail, the input of the proposed neural network is the end-effector pose, whereas the output is the value of the robot 7th joint. Eventually, alternative robot joints can also be considered [7]. Accurate neural network training requires a complete dataset of human arm movements, captured via a tracking system with five markers (Fig. 2a). These markers provide 3D coordinates, which are converted into pose matrices to define target positions and 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629638 55 orientations. To train the neural network, a custom algorithm processes the marker data to build a dataset including the end-effector poses and the 7th joint values. Optimal robot base placement is crucial for reachability; a vertical ceilingmounted is selected, to better emulate human kinematics (Fig. 2b). (a) (b) Fig. 2: Test recording (a), robot simulation (b). The neural network is trained using 80% of the data for training and 20% for evaluation. Training is performed via backpropagation to minimize prediction error over several epochs. Model accuracy is measured by how closely predicted values for the 7th joint match the reference value within 1% of its motion range. Multiple training loops with different hyperparameters are tested, and only the best-performing model is finally retained and used for predictions. III. EXPERIMENTAL RESULTS The dataset and the tests are collected using Optitrack tracking system (Fig. 2a). During data collection, a human subject performs natural arm movements, while test recordings involve performing pick-and-place tasks. Dataset recordings train the NN; test data is reserved for final evaluation. The developed neural network predicts the redundant joint value from Cartesian trajectories and computes full joint motion via IK. The system runs on Ubuntu 18.04, with Python (using PyTorch 2.4.1+cpu) for NN and motion planning, and C++ for IK and data processing. To evaluate its performance, the model is tested experimentally on a Franka Emika Panda robot. Robot control is handled via ROS (Melodic Morenia), whereas the simulations are run in Gazebo. Task executed by the human subject are replicated in simulation, and the predictions given by the neural network are compared with the reference data. The test was conducted using recorded pick-and-place tasks, in Tab. I. The model used for the experimental evaluation has 93% of accuracy. TABLE I: Results of the experimental test. NN time indicates the inference time, whereas IK time indicates the time to solve the inverse kinematics for each of the considered waypoints. RMSE [rad] Error [%] NN time [s] IK time [s] waypoints 0.039 0.673 0.299 0.041 177 The results of the experimental test show a good performance of the neural network in terms of prediction accuracy and demonstrate the effectiveness of neural networks Fig. 3: Comparison between reference and experimental q7(a), desired and measured OEE trajectory (b). Fig. 4: Exemplary test executed by a human subject and robot simulation. in emulating human arm movements on a redundant robotic manipulator. The results indicate a high degree of accuracy in trajectory reproduction, and the ability of the neural network to generalize across varying pick-and-place tasks highlights its adaptability. Future works will improve the training dataset by leveraging large-scale human motion capture datasets, and explore different manipulation tasks. ACKNOWLEDGMENT We thank the ADMiRE Research Center (Austria) and the Robotics and Intelligent Machines PhD Program of the University of Genova (Italy). REFERENCES [1] A. Brunello, G. Fabris, A. Gasparetto, A. Montanari, N. Saccomanno, and L. Scalera. A survey on recent trends in robotics and artificial intelligence in the furniture industry. Robotics and Computer-Integrated Manufacturing, 93:102920, 2025. [2] L. Scalera, F. Lozer, A. Giusti, and A. Gasparetto. An experimental evaluation of robot-stopping approaches for improving fluency in collaborative robotics. Robotica, 42(5):1386–1402, 2024. [3] W. Huang, C. Wang, Y. Li, R. Zhang, and L. Fei-Fei. Rekep: Spatiotemporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. [4] L. Qiyuan. Behavior imitation for manipulator control and grasping with deep reinforcement learning. arXiv preprint arXiv:2405.01284, 2024. [5] L. Coelho, S. M. Cerqueira, V. Martins, J. Andr´ e, and C. P. Santos. A DMPs-based approach for human-like robotic movements. In Int. Conf. on Aut. Rob. Syst. and Comp., pages 201–206. IEEE, 2024. [6] E. Asali and P. Doshi. Visual IRL for Human-Like Robotic Manipulation. arXiv preprint arXiv:2412.11360, 2024. [7] F. Lozer, L. Scalera, P. Boscariol, and A. Gasparetto. Planning optimal minimum-jerk trajectories for redundant robots. Robotics and Autonomous Systems, page 105049, 2025. [8] Y. He and S. Liu. Analytical inverse kinematics for Franka Emika Panda–a geometrical solver for 7-DOF manipulators with unconventional design. In 2021 9th International Conference on Control, Mechatronics and Automation (ICCMA), pages 194–199. IEEE, 2021. 56