Most of the state-of-the-art visual simultaneous localization and mapping (V-SLAM) algorithms perform well in structured environments. However, the assumption of a static world limits further practical application. In this paper, we present RS-SLAM, a robust semantic RGB-D SLAM system. RS-SLAM is able to detect both moving and movable objects for high accuracy localization. A semantic segmentation model is applied to recognize the dynamic objects. Specifically, the segmentation result is refined using the context information based on Bayesian update, which provides the more accurate extraction of the interested region. With the refined results, a kind of movable object detection approach is developed. In addition, an updating strategy during mapping is introduced and a 3D semantic OctoMap only with a static world is built in real-time. Our SLAM system is evaluated both on public datasets and real-world environments. The results show that RS-SLAM outperforms the accuracy of standard visual SLAM baselines and is capable of constructing a clean static semantic OctoMap in a dynamic environment.

Learning to map the images acquired by a moving agent equipped with a camera sensor to motion commands for multigoal navigation is challenging. Most existing approaches are still struggling against collision avoidance, faster convergence, and generalization. In this article, a novel actor–critic architecture is presented to learn the optimal navigation policy. We introduce single-step reward observation and collision penalty to reshape the reinforcement learning (RL) reward function. The collision perception can be obtained by the reshaped reward function and treated as measurement information from the visual observation to avoid obstacles. Besides, expert trajectories are used to generate subgoals. A subgoal reward shaping is then proposed to accelerate policy learning with the expert knowledge of subgoals. In order to generate human-aware navigation policies, an observation-action consistency (OAC) model is introduced to ensure that the agent reaches the subgoals in turn, and moves toward the target. The whole training process is performed on a self-supervised RL approach, accompanied by an expert supervision signal. This method balances the exploration and exploitation, helping the proposed model to generalize to unseen goals. The training experiments on AI2-THOR show better performance and faster convergence speed, compared with the existing approaches. For the generalization capacity to unseen goals, the proposed method achieves the state-of-the-art success rate, with at least a 30% improvement of average episode collision.

It is still a challenge to realize safe and fast autonomous driving through deep reinforcement learning. Most autonomous driving reinforcement learning models are subject to a single experience replay approach for training agents and how to improve the driving speed and safety of agent has become the focus of research. Therefore, we present an improved Double Bias Experience Replay (DBER) approach, which enables the agent to choose its own driving learning tendency. A new loss function is proposed to ameliorate the relationship between negative loss and positive loss. The proposed approach has been applied to three algorithms to verify: Deep Q Network (DQN), Dueling Double DQN (DD-DQN) and Quantile Regression DQN (QR-DQN). Compared with the existing approaches, the proposed approach show better performance and robustness of driving policy on the driving simulator, which is implemented by the Unity ML-agents. The approach makes the vehicle agent obtain better performance, such as higher reward, faster driving speed, less lane changing and more in the same training time.

In semantic visual simultaneous localization and mapping (VSLAM), accurate data association of semantic measurement from visual sensor is very crucial for robot state estimation and scene reconstruction. However, most of the related works assume a simple world for semantic association. It is still a challenge to deal with the ambiguity of data association in a cluttered environment. In this article, we propose a novel approach to reduce the uncertainty of data association via multiple hypothesis Dirichlet process (MHDP). The posterior distribution of data association is inferred by Dirichlet process (DP) first. Ambiguous associations from the distribution are tackled by a hypothesis tree, and hypothesis testing-based ambiguity judgment is then proposed for each object measurement to provide a strategy for branch growing of the hypothesis tree. Moreover, the proposed data association approach is integrated with a geometric featured-based simultaneous localization and mapping (SLAM) system in a tightly coupled way. The qualitative and quantitative evaluation on simulated and real-world public data sets demonstrates the robustness and effectiveness of our approach compared to other data association methods and the state-of-the-art SLAM system.