We introduce a novel human-centric deep reinforcement learning recommender system designed to co-optimize energy consumption, thermal comfort, and air quality in commercial buildings. Existing approaches typically optimize these objectives separately or focus solely on controlling energy-consuming building resources without directly engaging occupants. We develop a deep reinforcement learning architecture based on multitask learning with humans-in-the-loop and demonstrate how it can jointly learn energy savings, comfort, and air quality improvements for different building and occupant actions. In addition to controlling typical building resources (e.g., thermostat setpoint), our system provides real-time actionable recommendations that occupants can take (e.g., move to a new location) to co-optimize energy, comfort, and air quality. Through real deployments across multiple commercial buildings, we show that our multitask deep reinforcement learning recommender system has the potential to reduce energy consumption by up to 8% in energy-focused optimization, improve all objectives by 5–10% in joint optimization, and improve thermal comfort by up to 21% in comfort and air quality-focused optimization compared to existing solutions.

Thanks to its real-time computation efficiency, deep reinforcement learning (DRL) has been widely applied in motion planning for mobile robots. In DRL-based methods, a DRL model computes an action for a robot based on the states of its surrounding obstacles, including other robots that may communicate with it. These methods always assume that the environment is attack-free and the obtained obstacles’ states are reliable. However, in the real world, a robot may suffer from obstacle localization attacks (OLAs), such as sensor attacks, communication attacks, and remote-control attacks, which cause the robot to retrieve inaccurate positions of the surrounding obstacles. In this paper, we propose a robust motion planning method ObsGAN-DRL, integrating a generative adversarial network (GAN) into DRL models to mitigate OLAs in the environment. First, ObsGAN-DRL learns a generator based on the GAN model to compute the approximation of obstacles’ accurate positions in benign and attack scenarios. Therefore, no detectors are required for ObsGAN-DRL. Second, by using the approximation positions of the surrounding obstacles, ObsGAN-DRL can leverage the state-of-the-art DRL methods to compute collision-free motion commands (e.g., velocity) efficiently. Comprehensive experiments show that ObsGAN-DRL can mitigate OLAs effectively and guarantee safety. We also demonstrate the generalization of ObsGAN-DRL.

Safe and socially compliant navigation in a crowded environment is essential for social robots. Numerous research efforts have shown the advantages of deep reinforcement learning techniques in training efficient policies, while most of them ignore fast-moving pedestrians in the crowd. In this paper, we present a novel design of safety measure, named Risk-Area, considering collision theory and motion characteristics of different robots and humans. The geometry of Risk-Area is formed based on the real-time relative positions and velocities of the agents in the environment. Our approach perceives risk in the environment and encourages the robot to take safe and socially compliant navigation behaviors. The proposed method is verified with three existing well-known deep reinforcement learning models in densely populated environments. Experiment results demonstrate that our approach combined with the reinforcement learning techniques can efficiently perceive risk in the environment and navigate the robot with high safety in the crowds with fast-moving pedestrians.

Deep reinforcement learning (DRL) is promising for solving control problems in fluid mechanics, but it is a new field with many open questions. Possibilities are numerous and guidelines are rare concerning the choice of algorithms or best formulations for a given problem. Besides, DRL algorithms learn a control policy by collecting samples from an environment, which may be very costly when used with Computational Fluid Dynamics (CFD) solvers. Algorithms must therefore minimize the number of samples required for learning (sample efficiency) and generate a usable policy from each training (reliability). This paper aims to (a) evaluate three existing algorithms (DDPG, TD3, and SAC) on a fluid mechanics problem with respect to reliability and sample efficiency across a range of training configurations, (b) establish a fluid mechanics benchmark of increasing data collection cost, and (c) provide practical guidelines and insights for the fluid dynamics practitioner. The benchmark consists in controlling an airfoil to reach a target. The problem is solved with either a low-cost low-order model or with a high-fidelity CFD approach. The study found that DDPG and TD3 have learning stability issues highly dependent on DRL hyperparameters and reward formulation, requiring therefore significant tuning. In contrast, SAC is shown to be both reliable and sample efficient across a wide range of parameter setups, making it well suited to solve fluid mechanics problems and set up new cases without tremendous effort. In particular, SAC is resistant to small replay buffers, which could be critical if full-flow fields were to be stored.

Abstractive summarization is an approach to document summarization that is not limited to selecting sentences from the document but can generate new sentences as well. We address the two main challenges in abstractive summarization: how to evaluate the performance of a summarization model and what is a good training objective. We first introduce new evaluation measures based on the semantic similarity of the input and corresponding summary. The similarity scores are obtained by the fine-tuned BERTurk model using either the cross-encoder or a bi-encoder architecture. The fine-tuning is done on the Turkish Natural Language Inference and Semantic Textual Similarity benchmark datasets. We show that these measures have better correlations with human evaluations compared to Recall-Oriented Understudy for Gisting Evaluation (ROUGE) scores and BERTScore. We then introduce a deep reinforcement learning algorithm that uses the proposed semantic similarity measures as rewards, together with a mixed training objective, in order to generate more natural summaries in terms of human readability. We show that training with a mixed training objective function compared to only the maximum-likelihood objective improves similarity scores.

The control of a wing-in-ground craft (WIG) usually allows for many needs, like cruising, speed, survival and stealth. Various degrees of emphasis on these requirements result in different trajectories, but there has not been a way of integrating and quantifying them yet. Moreover, most previous studies on other vehicles’ multi-objective trajectory is planned globally, lacking for local planning. For the multi-objective trajectory planning of WIGs, this paper proposes a multi-objective function in a polynomial form, in which each item represents an independent requirement and is adjusted by a linear or exponential weight. It uses the magnitude of weights to demonstrate how much attention is paid relatively to the corresponding demand. Trajectories of a virtual WIG model above the wave trough terrain are planned using reward shaping based on the introduced multi-objective function and deep reinforcement learning (DRL). Two conditions are considered globally and locally: a single scheme of weights is assigned to the whole environment, and two different schemes of weights are assigned to the two parts of the environment. Effectiveness of the multi-object reward function is analysed from the local and global perspectives. The reward function provides WIGs with a universal framework for adjusting the magnitude of weights, to meet different degrees of requirements on cruising, speed, stealth and survival, and helps WIGs guide an expected trajectory in engineering.

An onboard three-dimensional (3D) trajectory generation approach based on the reinforcement learning (RL) algorithm and deep neural network (DNN) is proposed for hypersonic vehicles in glide phase. Multiple trajectory samples are generated offline through the convex optimisation method. The deep learning (DL) is employed to pre-train the DNN for initialising the actor network and accelerating the RL process. Based on the offline deep policy deterministic actor-critic algorithm, a flight target-oriented reward function with path constraints is designed. The actor network is optimised by the end-to-end RL and policy gradients of the critic network until the reward function converges to the maximum. The actor network is considered as the onboard trajectory generator to compute optimal control values online based on the real-time motion states. The simulation results show that the single-step online planning time meets the real-time requirements of onboard trajectory generation. The significant improvement in terminal accuracy of the online trajectory and the better generalisation under biased initial states for hypersonic vehicles in glide phase is observed.

In this paper, we propose a set of robust training methods for deep reinforcement learning to transfer learning acquired in one control task to a set of previously unseen control tasks. We improve generalization in commonly used transfer learning benchmarks by a novel sample elimination technique, early stopping, and maximum entropy adversarial reinforcement learning. To generate robust policies, we use sample elimination during training via a method we call strict clipping. We apply early stopping, a method previously used in supervised learning, to deep reinforcement learning. Subsequently, we introduce maximum entropy adversarial reinforcement learning to increase the domain randomization during training for a better target task performance. Finally, we evaluate the robustness of these methods compared to previous work on simulated robots in target environments where the gravity, the morphology of the robot, and the tangential friction coefficient of the environment are altered.

Robotic systems are usually controlled to repetitively perform specific actions for manufacturing tasks. The traditional control methods are domain-dependent and model-dependent with cost of much human efforts. They cannot meet the new requirements of generality and flexibility in many areas such as intelligent manufacturing and customized production. This paper develops a general model-free approach to enable robots to perform multi-step object sorting tasks through deep reinforcement learning. Taking projected heightmap images from different time steps as input without extra high-level image analysis and understanding, critic models are designed to produce a pixel-wise Q value map for each type of action. It is a new trial to apply pixel-wise Q value-based critic networks to solve multi-step sorting tasks that involve many types of actions and complex action constraints. The experimental validations on simulated and realistic object sorting tasks demonstrate the effectiveness of the proposed approach. Qualitative results (videos), code for simulated and realistic experiments, and pre-trained models are available at https://github.com/JiatongBao/DRLSorting

Lower-body exoskeleton control that adapts to users and provides assistance-as-needed can increase user participation and motor learning and allow for more effective gait rehabilitation. Adaptive model-based control methods have previously been developed to consider a user’s interaction with an exoskeleton; however, the predefined dynamics models required are challenging to define accurately, due to the complex dynamics and nonlinearities of the human-exoskeleton interaction. Model-free deep reinforcement learning (DRL) approaches can provide accurate and robust control in robotics applications and have shown potential for lower-body exoskeletons. In this paper, we present a new model-free DRL method for end-to-end learning of desired gait patterns for over-ground gait rehabilitation with an exoskeleton. This control technique is the first to accurately track any gait pattern desired in physiotherapy without requiring a predefined dynamics model and is robust to varying post-stroke individuals’ baseline gait patterns and their interactions and perturbations. Simulated experiments of an exoskeleton paired to a musculoskeletal model show that the DRL method is robust to different post-stroke users and is able to accurately track desired gait pattern trajectories both seen and unseen in training.

Operation of tools has long been studied in robotics. Although appropriate hold of the tool by robots is the base of successful tool operation, it is not with ease especially for tools with complicated shape. In this paper, an assist system for a four-limbed robot is proposed for remote operation of reaching and grasping electric drills using deep reinforcement learning. Through comparative evaluation experiments, the increase of success rate for reaching and grasping is verified and the decrease in both physical and mental workload of the operator is also validated by the index of NASA-TLX.

Collision avoidance for robots and vehicles in unpredictable environments is a challenging task. Various control strategies have been developed for the agent (i.e., robots or vehicles) to sense the environment, assess the situation, and select the optimal actions to avoid collision and accomplish its mission. In our research on autonomous ships, we take a machine learning approach to collision avoidance. The lack of available ship steering data of human ship masters has made it necessary to acquire collision avoidance knowledge through reinforcement learning (RL). Given that the learned neural network tends to be a black box, it is desirable that a method is available which can be used to design an agent's behavior so that the desired knowledge can be captured. Furthermore, RL with complex tasks can be either time consuming or unfeasible. A multi-stage learning method is needed in which agents can learn from simple tasks and then transfer their learned knowledge to closely related but more complex tasks. In this paper, we explore the ways of designing agent behaviors through tuning reward functions and devise a transfer RL method for multi-stage knowledge acquisition. The computer simulation-based agent training results have shown that it is important to understand the roles of each component in a reward function and the various design parameters in transfer RL. The settings of these parameters are all dependent on the complexity of the tasks and the similarities between them.

Autonomous landing on the deck of a boat or an unmanned surface vehicle (USV) is the minimum requirement for increasing the autonomy of water monitoring missions. This paper introduces an end-to-end control technique based on deep reinforcement learning for landing an unmanned aerial vehicle on a visual marker located on the deck of a USV. The solution proposed consists of a hierarchy of Deep Q-Networks (DQNs) used as high-level navigation policies that address the two phases of the flight: the marker detection and the descending manoeuvre. Few technical improvements have been proposed to stabilize the learning process, such as the combination of vanilla and double DQNs, and a partitioned buffer replay. Simulated studies proved the robustness of the proposed algorithm against different perturbations acting on the marine vessel. The performances obtained are comparable with a state-of-the-art method based on template matching.