Although reinforcement learning (RL) has demonstrated large applicability in the robotics field, the deployment of models trained in simulation on an actual device is still a challenge. This happens due to the differences between simulation and real environments, which is known as simulation to reality (sim-to-real) gap. Approaches to solve this problem typically involve domain randomization (DR) in which simulation parameters are repeatedly varied during the training stage, or transfer learning (TL) in which real data is introduced to the RL model pretrained in simulation. This problem can also be seen through the safe RL perspective in which constraints are incorporated into the policy by aiming at the robustness of the trained agent. Each of these approaches may bridge the sim-to-real gap with limited capability to deal with unforeseen situations when used separately. For increased robustness and adaptability of the RL agent, this paper proposes a novel solution that combines safe RL, DR, and TL throughout the sim-to-real transfer process. This pipeline starts by designing an agent in a safe RL formulation based on Trust Region Conditional Value at Risk (TRC). This agent is then trained on simulation through DR. Finally, the agent is adapted to the real environment by means of a short online training via fine-tuning (i.e. TL). The proposed solution was validated on an actual mobile robot assigned to navigate toward a goal position by avoiding varied obstacles. Results showed that our solution presented superior performance to accomplish tasks in situations not presented before, compared to the separate usage of sim-to-real approaches.

A main challenge of using 5G network slices is to meet all the quality of service requirements of the slices (which are agreed with the customer in a service level agreement (SLA)), throughout the network slices' lifecycle. To avoid the penalty for violation, a proactive solution is presented, including predicting the SLA violation, explaining the violation cause, and then providing an adaptation to traffic. This work uses counterfactual (CF) explanations to 1) explain the main factors affecting the identified model's SLA violation prediction and 2) present modifications in the input values, which are required to configure the network traffic to avoid such a violation. We evaluate the CF explanation at two different levels where the generated CF instance is fed to the predictive model, and then actuation data are generated to evaluate the result in the real network. Our solution minimizes the violation up to 98%. This information can be utilized to reconfigure the system, either by humans or by the system automatically, to make it fully autonomous on the one hand and comply with the 'right to explanation' introduced by the European Union's General Data Protection Regulation on the other hand.

Robots with constrained hardware resources usually rely on Multi-access Edge Computing infrastructures to offload computationally expensive tasks to meet real-time and safety requirements. Offloading every task might not be the best option due to dynamic changes in the network conditions and can result in network congestion or failures. This work proposes a task offloading strategy for mobile robots in a Human-Robot Collaboration scenario that optimizes the edge resource usage and reduces network delays, leading to safety enhancement. The solution utilizes a Deep Reinforcement Learning (DRL) agent that observes safety and network metrics to dynamically decide at runtime if (i) a less accurate model should run on the robot; (ii) a more complex model should run on the edge; or (iii) the previous output should be reused through temporal coherence verification. Experiments are performed in a simulated warehouse where humans and robots have close interactions and safety needs are high. Our results show that the proposed DRL solution outperforms the baselines in several aspects. The edge is used only when the network performance is reliable, reducing the number of failures (up to 47 %). The latency is not only decreased (up to 68 %) but also adapted to the safety requirements (risk × latency reduced up to 48 %), avoiding unnecessary network congestion in safe situations and letting other devices use the network. Overall, the safety metrics get improved, such as the increased time in the safe zone by up to 3.1%.

Artificial Intelligence (AI) is implemented in various applications of telecommunication domain, ranging from managing the network, controlling a specific hardware function, preventing a failure, or troubleshooting a problem till automating the network slice management in 5G. The greater levels of autonomy increase the need for explainability of the decisions made by AI so that humans can understand them (e.g. the underlying data evidence and causal reasoning) consequently enabling trust. This paper presents first, the application of multiple global and local explainability methods with the main purpose to analyze the root-cause of Service Level Agreement violation prediction in a 5G network slicing setup by identifying important features contributing to the decision. Second, it performs a comparative analysis of the applied methods to analyze explainability of the predicted violation. Further, the global explainability results are validated using statistical Causal Dataframe method in order to improve the identified cause of the problem and thus validating the explainability.