In this paper, we propose a novel Deep Reinforcement Learning (DRL) algorithm which can navigate non-holonomic robots with continuous control in an unknown dynamic environment with moving obstacles. We call the approach MK-A3C (Memory and Knowledge-based Asynchronous Advantage Actor-Critic) for short. As its first component, MK-A3C builds a GRU-based memory neural network to enhance the robot’s capability for temporal reasoning. Robots without it tend to suffer from a lack of rationality in face of incomplete and noisy estimations for complex environments. Additionally, robots with certain memory ability endowed by MK-A3C can avoid local minima traps by estimating the environmental model. Secondly, MK-A3C combines the domain knowledge-based reward function and the transfer learning-based training task architecture, which can solve the non-convergence policies problems caused by sparse reward. These improvements of MK-A3C can efficiently navigate robots in unknown dynamic environments, and satisfy kinetic constraints while handling moving objects. Simulation experiments show that compared with existing methods, MK-A3C can realize successful robotic navigation in unknown and challenging environments by outputting continuous acceleration commands.

Referring expression comprehension (REC) is a vision-language task to locate a target object in an image based on a language expression. Fully fine-tuning general-purpose pre-trained vision-language foundation models for REC yields impressive performance but becomes increasingly costly. Parameter-efficient transfer learning (PETL) methods have shown strong performance with fewer tunable parameters. However, directly applying PETL to REC faces two challenges: (1) insufficient multi-modal interaction between pre-trained vision-language foundation models, and (2) high GPU memory usage due to gradients passing through the heavy vision-language foundation models. To this end, we present M2IST: Multi-Modal Interactive Side-Tuning with M3ISAs: Mixture of Multi-Modal Interactive Side-Adapters. During fine-tuning, we fix the pre-trained uni-modal encoders and update M3ISAs to enable efficient vision-language alignment for REC. Empirical results reveal that M2IST achieves better performance-efficiency trade-off than full fine-tuning and other PETL methods, requiring only 2.11% tunable parameters, 39.61% GPU memory, and 63.46% training time while maintaining competitive performance. Our code is released at https://github.com/xuyang-liu16/M2IST.

Spatial mapping and navigation are critical cognitive functions of autonomous agents, enabling one to learn an internal representation of an environment and move through space with real-time sensory inputs, such as visual observations. Existing models for vision-based mapping and navigation, however, suffer from memory requirements that increase linearly with exploration duration and indirect path following behaviors. This article presents e-TM, a self-organizing neural network-based framework for incremental topological mapping and navigation. e-TM models the exploration trajectories explicitly as episodic memory, wherein salient landmarks are sequentially extracted as ``events'' from streaming observations. A memory consolidation procedure then performs a playback mechanism and transfers the embedded knowledge of the environmental layout into spatial memory, encoding topological relations between landmarks. Fusion adaptive resonance theory (ART) networks, as the building block of the two memory modules, can generalize multiple input patterns into memory templates and, therefore, provide a compact spatial representation and support the discovery of novel shortcuts through inferences. For navigation, e-TM applies a transfer learning paradigm to integrate human demonstrations into a pretrained locomotion network for smoother movements. Experimental results based on VizDoom, a simulated 3-D environment, have shown that, compared to semiparametric topological memory (SPTM), a state-of-the-art model, e-TM reduces the time costs of navigation significantly while learning much sparser topological graphs.

Since an individual approach can hardly navigate robots through complex environments, we present a novel two-level hierarchical framework called JPS-IA3C (Jump Point Search improved Asynchronous Advantage Actor-Critic) in this paper for robot navigation in dynamic environments through continuous controlling signals. Its global planner JPS+ (P) is a variant of JPS (Jump Point Search), which efficiently computes an abstract path of neighboring jump points. These nodes, which are seen as subgoals, completely rid Deep Reinforcement Learning (DRL)-based controllers of notorious local minima. To satisfy the kinetic constraints and be adaptive to changing environments, we propose an improved A3C (IA3C) algorithm to learn the control policies of the robots’ local motion. Moreover, the combination of modified curriculum learning and reward shaping helps IA3C build a novel reward function framework to avoid learning inefficiency because of sparse reward. We additionally strengthen the robots’ temporal reasoning of the environments by a memory-based network. These improvements make the IA3C controller converge faster and become more adaptive to incomplete, noisy information caused by partial observability. Simulated experiments show that compared with existing methods, this JPS-IA3C hierarchy successfully outputs continuous commands to accomplish large-range navigation tasks at shorter paths and less time through reasonable subgoal selection and rational motions.

With the gradual expansion of high buildings and underground spaces, deep foundation pits have been widely used in these engineering projects, but if they are not well-designed, safety problems occur. Proper deep foundation pit design requires proper exit distribution. However, calculating an adequate number of exit distributions for evaluation is difficult due to the numerous influential factors existing in the deep foundation pit environment. To this end, this paper presents a prototype of a decision-making system that uses agent-based modeling to simulate deep foundation pit evacuation in the presence of collapse disaster. By modeling the collapse occurrence process and agent escape process, an agent-based evacuation model is built, and a modified simulation-based particle swarm optimization algorithm is used to solve the optimization problem of exit distribution. Extensive experiments are conducted to verify the system, and the results show that the system provides a feasible framework for deep foundation pit evacuation.