Navigation is a fundamental capability for robots to operate in expansive spaces. Reliable navigation in unknown environments is crucial for deploying robots in areas such as disaster rescue and industrial inspection. In such scenarios, it is essential for robots to construct memories based on historical data to support long-term, optimized decision-making. However, many existing techniques focus on memorizing direct features from raw perceptions, often resulting in redundancy due to irrelevant textures and areas. This approach leads to inefficiencies in computation and storage, and produces a memory structure that lacks general applicability. We suggest that it may not be necessary to store specific scene features. Instead, recalling the robot’s episodic movements could provide sufficient cognitive cues for navigation. To address this, we introduce Memory Enhanced Navigation with Embedded Odometry (MENEO), a framework consisting of three steps: ego-motion estimation, memory aggregation, and adaptive policy generation. MENEO offers two main advantages: its streamlined architecture significantly boosts computational and storage efficiency, and its universal design supports various sensor modalities, adapts to multiple navigation tasks, and accommodates different scenarios. We test MENEO in two different environments: maze exploration using a lidar-IMU sensor, and image-goal visual navigation in photorealistic indoor scenes. In both cases, MENEO demonstrates competitive navigation performance, outperforming existing methods by reducing storage and computational requirements. Additionally, MENEO’s compact memory representation not only enhances adaptability across diverse environments but also shows flexibility in real-world applications.Note to Practitioners—In learning-based navigation, the memory mechanism is essential for long-term optimized policies. It allows intelligent robots to make informed decisions by utilizing a wide range of temporal and spatial cues derived from historical data. Traditional methods use various types of memory (such as internal, unstructured, or structured), but these often result in computational and storage inefficiencies due to the direct inclusion of complex and redundant raw scene features. Furthermore, because these memory systems are closely linked to specific scene features, they lack general applicability across different sensor configurations, scene types, and tasks. In this paper, we aim to eliminate the need to directly manage the redundant environmental features found in previous memory structures. Instead, we propose focusing solely on memorizing a robot’s self-movements. Since the pose sequence is streamlined and compact, our approach not only enhances computational and storage efficiency but also improves the interpretability and universality of the navigation system. This advantage enables MENEO to be seamlessly integrated into a wide variety of intelligent navigation systems. It is especially beneficial for small robots with limited computing power, such as those used in search and rescue operations, where enhanced memory can greatly enhance their autonomous navigation capabilities.

Autonomous navigation substantially depends on memory mechanisms for enhancing path optimality. However, the dynamic nature of environments can cause inconsistencies between the stored scene memory and real-time perception, leading to potentially catastrophic navigation errors. Existing memory structures often fall short in addressing these challenges, as they typically only account for object-level dynamics and falter when faced with long-term alterations in environmental structure. To counter these constraints, this paper presents a learning-based framework designed to construct and maintain a geometric memory, thereby facilitating improved visuomotor navigation under structural dynamics. This proposed framework employs visual inputs to construct a geometric representation of the environment, and identifies structural changes by assessing the consistency between the established memory and instantaneous perception. To update the geometric memory efficiently, we introduce a memory updater grounded in a structure storage pool. Furthermore, a two-phase hierarchical planner is proposed to decompose navigation tasks and formulate smooth navigation strategies. Experimental results from photorealistic simulations underscore the efficacy of the proposed system in managing long-term dynamics and navigation control. The effectiveness of the proposed system is further corroborated through deployment and testing in real-world environments.Note to Practitioners—Classic geometry-based navigation methods hinge on the construction of a global map for spatial reasoning and optimized robot control. Within a learning-based pipeline, the dependence on memory information becomes critical for enabling robots to develop spatial and temporal awareness. While the maintenance of memory structures can significantly expand the field of the robot’s perception in time and space, discrepancies between historical memory and real-time perception in dynamic environments can lead to misguided decisions. Unlike most existing research that addresses short-term dynamic issues at the object level, this paper centers on long-term, large-scale dynamics precipitated by changes in environmental structure. We put forward a learning-based framework that incorporates dynamic perception, map maintenance, and hierarchical navigation. The experimental results highlight the efficiency and real-time processing capability of this method in handling structural dynamics, hence enhancing navigation efficiency.

Visual navigation is fundamental for embodied agents operating in expansive workspaces. The cognitive abilities of these agents form the essential basis for creating intelligent behavioral patterns. Memory and reasoning are vital components among these abilities. The former enhances decision-making by preserving a wide array of episodic spatio-temporal perception cues, while the latter allows proactive and advanced probabilistic inference of task distributions based on long-term experiences. Despite individual studies on these two cognitive modalities, their integration for enhanced decision-making presents a considerable challenge due to their substantial differences in representation and behavioral characteristics. In this paper, we introduce Semantic-based Multi-modal Cognitive Graph (SMCG) for intelligent visual navigation. This framework is distinguished by its unified semantic-level representation of both memory and reasoning capabilities. Specifically, SMCG, rather than directly memorizing perceptual features as per previous methods, records observed object sequences. Simultaneously, reasoning is based on a semantic relation graph that represents correlations among objects. We additionally develop a hierarchical cognition extraction (HCE) pipeline and employ it to decode cognitive cues within SMCG and situation-aware subgraphs, thereby enhancing intelligent navigation behavior. Experimental results in image-goal navigation show pronounced performance improvements, credited to the effective induction and rational application of heterogeneous cognitive modalities.

The fundamental prerequisite for embodied agents to make intelligent decisions lies in autonomous cognition. Typically, agents optimize decision-making by leveraging extensive spatiotemporal information from episodic memory. Concurrently, they utilize long-term experience for task reasoning and foster conscious behavioral tendencies. However, due to the significant disparities in the heterogeneous modalities of these two cognitive abilities, existing literature falls short in designing effective coupling mechanisms, thus failing to endow robots with comprehensive intelligence. This article introduces a navigation framework, the hierarchical topology-semantic cognitive navigation (HTSCN), which seamlessly integrates both memory and reasoning abilities within a singular end-to-end system. Specifically, we represent memory and reasoning abilities with a topological map and a semantic relation graph, respectively, within a unified dual-layer graph structure. Additionally, we incorporate a neural-based cognition extraction process to capture cross-modal relationships between hierarchical graphs. HTSCN forges a link between two different cognitive modalities, thus further enhancing decision-making performance and the overall level of intelligence. Experimental results demonstrate that in comparison to existing cognitive structures, HTSCN significantly enhances the performance and path efficiency of image-goal navigation. Visualization and interpretability experiments further corroborate the promoting role of memory, reasoning, as well as their online learned relationships, on intelligent behavioral patterns. Furthermore, we deploy HTSCN in real-world scenarios to further verify its feasibility and adaptability.

Robotic systems with hundreds or even thousands of robots are widely implemented in logistic and industrial applications. In such systems, cooperative path planning is of great importance, as local congestion and motion conflict may greatly degrade system performance, especially in the presence of uncertainties. Our idea is to consider traffic flow equilibrium in path planning to relieve any potential congestion and increase efficiency. In this paper, we propose a hierarchical framework, which includes a traffic flow prediction layer, a sector-level planning layer, and a road-level coordination layer. In traffic flow prediction, we propose a spatio-temporal graph neural network that integrates local information to predict the evolution of future robot density distribution. In sector-level planning, we generate sector-level paths that consider travel distance and traffic flow equilibrium simultaneously. In road-level coordination, we implement the conflict-based search algorithm within each sector to ensure conflict-free local paths. In addition, we also explicitly consider motion/communication uncertainties that are unavoided in practical systems. We validate our effectiveness in simulations with over 1000 robots, what’s more, real experiments are provided.

In the context of visual navigation in unknown scenes, both “exploration” and “exploitation” are equally crucial. Robots must first establish environmental cognition through exploration and then utilize the cognitive information to accomplish target searches. However, most existing methods for image-goal navigation prioritize target search over the generation of exploratory behavior. To address this, we propose the Navigation with Uncertainty-driven Exploration (NUE) pipeline, which uses an implicit and compact scene representation, NeRF, as a cognitive structure. We estimate the uncertainty of NeRF and augment the exploratory ability by the uncertainty to in turn facilitate the construction of implicit representation. Simultaneously, we extract memory information from NeRF to enhance the robot’s reasoning ability for determining the location of the target. Ultimately, we seamlessly combine the two generated abilities to produce navigational actions. Our pipeline is end-to-end, with the environmental cognitive structure being constructed online. Extensive experimental results on image-goal navigation demonstrate the capability of our pipeline to enhance exploratory behaviors, while also enabling a natural transition from the exploration to exploitation phase. This enables our model to outperform existing memory-based cognitive navigation structures in terms of navigation performance. Project page: https://github.com/IRMVLab/NUE-NeRF-nav

We propose an end-to-end visuomotor navigation framework that leverages Neural Radiance Fields (NeRF) for spatial cognition. To the best of our knowledge, this is the first effort to integrate such implicit spatial representation with embodied policy end-to-end for cognitive decision-making. Consequently, our system does not necessitate modularized designs nor transformations into explicit scene representations for downstream control. The NeRF-based memory is constructed online during navigation, without relying on any environmental priors. To enhance the extraction of decision-critical historical insights from the rigid and implicit structure of NeRF, we introduce a spatial information extraction mechanism named Structural Radiance Attention (SRA). SRA empowers the agent to grasp complex scene structures and task objectives, thus paving the way for the development of intelligent behavioral patterns. Our comprehensive testing in image-goal navigation tasks demonstrates that our approach significantly outperforms existing navigation models. We demonstrate that SRA markedly improves the agent's understanding of both the scene and the task by retrieving historical information stored in NeRF memory. The agent also learns exploratory awareness from our pipeline to better adapt to low signal-to-noise memory signals in unknown scenes. We deploy our navigation system on a mobile robot in real-world scenarios, where it exhibits evident cognitive capabilities while ensuring real-time performance.

Creating memory representations is essential for developing viable navigation strategies for intelligent agents. Although neural radiance fields (NeRFs) have shown great promise as a novel method for spatial representation, their potential for integration into learning-based navigation as a memory structure has been largely overlooked in the existing literature. In this article, we introduce a navigation pipeline that incorporates NeRF into visuomotor navigation. Initially, we present a derivative radiance field that facilitates one-shot pose and depth estimation from a single query image. By assuming equivalence between density and space occupancy, we generate a geometric accessibility map based on an offline-constructed NeRF prior. Utilizing the above information, we design a global planner that decomposes long-term tasks by performing waypoint estimation and rendering. Finally, we employ an imitation-learned local controller to achieve a reliable navigation policy. Our pipeline effectively utilizes NeRF's compact spatial representation for task decomposition and action generation, enabling efficient navigation. Experimental results highlight its advantages over recent implicit and explicit memory approaches in image-goal navigation tasks. Moreover, we conduct interpretability studies and apply our algorithm in real-world scenarios to further attest to its practicality and effectiveness.

This article investigates the multirobot cooperative navigation problem based on raw visual observations. A fully end-to-end learning framework is presented, which leverages graph neural networks to learn local motion coordination and utilizes deep reinforcement learning to generate visuomotor policy that enables each robot to move to its goal without the need of environment map and global positioning information. Experimental results show that, with a few tens of robots, our approach achieves comparable performance with the state-of-the-art imitation learning-based approaches with bird-view state inputs. We also illustrate our generalizability to crowded and large environments and our scalability to ten times number of the training robots. In addition, we demonstrate that our model trained for multirobot case can also improve the success rate in the single-robot navigation task in unseen environments. Note to Practitioners—With the development of intelligent industrial and logistic systems, robotic transportation systems are widely implemented. However, existing multirobot path coordination and navigation approaches are basically under some unreasonable assumptions, which are very hard to be implemented in practical scenarios. This article aims to greatly promote the real application of learning-based multirobot cooperative navigation approach, in order to achieve the following. First, we introduce an end-to-end reinforcement learning framework instead of the commonly used imitation learning strategy, as the latter one needs exhaustive training data to cover all the scenarios and does not have the required generalizability. Second, we directly use the raw sensor data instead of the commonly used bird-eye-view semantic observations, as the latter one is generally not representative of practical application scenario from the robot perspective and cannot solve the occlusion issue. Third, we interpret our learned model to illustrate which parts of the input and shared observations contribute most to the robots’ final actions. The above interpretability ensures predictability (thus safety) of our visuomotor policy in practical applications. Our learned visuomotor policy has the ability to coordinate dozens of robots by only using raw visual observations in unknown environments without map nor global localization information, this is the first time in the literature. Our future work includes solving the sim-to-real issue and conducting physical experiments.