Manuel Beschi

This paper addresses motion replanning in human-robot collaborative scenarios, with an emphasis on reactivity and safety-compliant efficiency. While existing human-aware motion planners perform well in structured environments, they often struggle with unpredictable human behavior. This can result in safety measures that hinder the robot’s performance and overall throughput. This study combines reactive path replanning and a safety-aware cost function, enabling the robot to adapt its path to the changes in the scene in real-time. This solution reduces the execution time and trajectory slowdowns while ensuring safety. Simulations and real-world experiments show the method’s effectiveness compared to standard human-robot cooperation approaches, with efficiency enhancements of up to 60%.

The rapid evolution of industrial robot hardware has created a technological gap with software, limiting its adoption. The software solutions proposed in recent years have yet to meet the industrial sector’s requirements, as they focus more on the definition of task structure than the definition and tuning of its execution parameters. A framework for task parameter optimization was developed to address this gap. It breaks down the task using a modular structure, allowing the task optimization piece by piece. The optimization is performed with a dedicated hill-climbing algorithm. This paper revisits the framework by proposing an alternative approach that replaces the algorithmic component with reinforcement learning (RL) models. Five RL models are proposed with increasing complexity and efficiency. A comparative analysis of the traditional algorithm and RL models is presented, highlighting efficiency, flexibility, and usability. The results demonstrate that although RL models improve task optimization efficiency by 95%, they still need more flexibility. However, the nature of these models provides significant opportunities for future advancements.

This article aims to design an assistive controller for physical Human-Robot Interaction (pHRI) based on Dynamic Cooperative Game Theory (DCGT). In particular, a distributed Model Predictive Control (dMPC) is formulated based on the DCGT principles (GT-dMPC). For proper implementation, one crucial piece of information regards human intention, which is defined as the desired trajectory that a human wants to follow over a finite rolling prediction horizon. To predict the desired human trajectory, a learning model is composed of cascaded Long-Short Term Memory (LSTM) and Fully Connected (FC) layers (RNN $+$ FC). Iterative training and Transfer Learning (TL) techniques are proposed to adapt the model to different users. The behavior of the proposed GT-dMPC framework is thoroughly analyzed with simulations to understand its applicability and the tuning of its parameters for a pHRI assistive controller. Moreover, real-world experiments were carried out on a UR5 robotic arm equipped with a force sensor was installed. First, a brief validation of the RNN $+$ FC model integrated with the GT-dMPC is proposed for the iterative procedure and the TL. Finally, an application scenario is proposed for co-manipulating two objects and comparing the obtained results with other controllers typically used in the pHRI. Results show that the proposed controller reduces the required force of the human in completing tasks, even in the presence of unknown and different loads and inertia. Moreover, the proposed controller allows for precise reaching of the target point and does not introduce any undesirable oscillations. Finally, a subjective questionnaire shows that the proposed controller is, in general, preferred by different users. Note to Practitioners —This work presents a method to design an assistive controller to help a human perform physically coupled shared tasks with a robot. The target applications of this work are co-handling tasks of large or heavy objects. Such tasks require two agents to be performed easily, and the proposed work aims to make the robot a companion for the human partner. The proposed approach also quickly adapts to new users or tasks, making it feasible for real production systems or daily scenarios. Another possible target application is the co-manipulating large flexible components such as carbon fiber plies. This application would require small modifications, particularly in how the force is exchanged. Some additional/different sensors should be used, such as vision to map object deformations with virtual forces. The present work does not directly consider these kinds of applications. Indeed, this work strictly relies on force measurements that are not reliable when dealing with flexible materials, at least in a compression state. Such an issue will be investigated in future works by using vision systems to measure a virtual force that allows this method to be applicable even in the case of flexible components.

This paper improves the performance of RRT$^*$-like sampling-based path planners by combining admissible informed sampling and local sampling (i.e., sampling the neighborhood of the current solution). An adaptive strategy regulates the trade-off between exploration (admissible informed sampling) and exploitation (local sampling) based on online rewards from previous samples. The paper demonstrates that the algorithm is asymptotically optimal and has a better convergence rate than state-of-the-art path planners (e.g., Informed-RRT*) in several simulated and real-world scenarios. An open-source, ROS-compatible implementation of the algorithm is publicly available.

The industry needs controllers that allow smooth and natural physical Human-Robot Interaction (pHRI) to make production scenarios more flexible and user-friendly. Within this context, particularly interesting is Role Arbitration, which is the mechanism that assigns the role of the leader to either the human or the robot. This paper investigates Game-Theory (GT) to model pHRI, and specifically, Cooperative Game Theory (CGT) and Non-Cooperative Game Theory (NCGT) are considered. This work proposes a possible solution to the Role Arbitration problem and defines a Role Arbitration framework based on differential game theory to allow pHRI. The proposed method can allow trajectory deformation according to human will, avoiding reaching dangerous situations such as collisions with environmental features, robot joints and workspace limits, and possibly safety constraints. Three sets of experiments are proposed to evaluate different situations and compared with two other standard methods for pHRI, the Impedance Control, and the Manual Guidance. Experiments show that with our Role Arbitration method, different situations can be handled safely and smoothly with a low human effort. In particular, the performances of the IMP and MG vary according to the task. In some cases, MG performs well, and IMP does not. In some others, IMP performs excellently, and MG does not. The proposed Role Arbitration controller performs well in all the cases, showing its superiority and generality. The proposed method generally requires less force and ensures better accuracy in performing all tasks than standard controllers. Note to Practitioners —This work presents a method that allows role arbitration for physical Human-Robot Interaction, motivated by the need to adjust the role of leader/follower in a shared task according to the specific phase of the task or the knowledge of one of the two agents. This method suits applications such as object co-transportation, which requires final precise positioning but allows some trajectory deformation on the fly. It can also handle situations where the carried obstacle occludes human sight, and the robot helps the human to avoid possible environmental obstacles and position the objects at the target pose precisely. Currently, this method does not consider external contact, which is likely to arise in many situations. Future studies will investigate the modeling and detection of external contacts to include them in the interaction models this work addresses.

This work addresses human intention identification during physical Human-Robot Interaction (pHRI) tasks to include this information in an assistive controller. To this purpose, human intention is defined as the desired trajectory that the human wants to follow over a finite rolling prediction horizon so that the robot can assist in pursuing it. This work investigates a Recurrent Neural Network (RNN), specifically, Long-Short Term Memory (LSTM) cascaded with a Fully Connected layer. In particular, we propose an iterative training procedure to adapt the model. Such an iterative procedure is powerful in reducing the prediction error. Still, it has the drawback that it is time-consuming and does not generalize to different users or different co-manipulated objects. To overcome this issue, Transfer Learning (TL) adapts the pre-trained model to new trajectories, users, and co-manipulated objects by freezing the LSTM layer and fine-tuning the last FC layer, which makes the procedure faster. Experiments show that the iterative procedure adapts the model and reduces prediction error. Experiments also show that TL adapts to different users and to the co-manipulation of a large object. Finally, to check the utility of adopting the proposed method, we compare the proposed controller enhanced by the intention prediction with the other two standard controllers of pHRI.