World-model-based reinforcement learning achieves high sample efficiency by learning from imagined rollouts. However, its success critically depends on the accuracy of the learned world model, which is prone to producing unrealistic or hallucinated rollouts when queried beyond its domain of competence. These flawed predictions can trap the agent in a vicious cycle: by misleading exploration toward implausible or uninformative regions, they degrade the quality of collected data, which in turn corrupts policy learning with inaccurate rollouts. To break this cycle, we introduce the notion of a knowledge boundary—the region within which the world model provides reliable predictions—and propose a unified framework that both identifies and leverages this boundary. Concretely, we approximate the boundary using model uncertainty, quantified via disagreement across an ensemble of lightweight predictors, which serves as a practical proxy. This uncertainty signal is used in two complementary ways: as an intrinsic reward to guide exploration toward under-explored yet learnable regions, and as a dynamic filter to exclude unreliable imagined rollouts from policy optimization. Extensive experiments across diverse benchmarks—including CARLA, DeepMind Control Suite, Atari, and MemoryMaze—demonstrate that our approach consistently outperforms prior state-of-the-art methods.

Model-based reinforcement learning (MBRL) enables efficient decision-making by learning predictive world modelsof environment dynamics. Despite recent advances, existingmodels often struggle to reconcile accurate short-term transitions with coherent long-term planning, especially in partially observable or long-horizon settings. We argue that thislimitation often stems from modeling all transitions at a single temporal resolution, which makes it challenging to simultaneously capture fine-grained local dynamics and abstractglobal structures. To this end, we propose SF-RSSM (Slow-Fast Recurrent State-Space Model), a novel method that decouples short-term and long-term dynamics via a dualbranchdesign. The fast branch captures short-horizon transitions using residual prediction, while the slow branch models long-range dependencies with a GRU-based recurrent pathway.A distillation mechanism is developed to enable cooperationacross timescales, with the slow model providing soft targetsto guide the fast model. Additionally, a curiosity module encourages exploration by promoting learning in regions wherethe fast and slow branches exhibit divergent dynamics. Experiments on CARLA, DMControl and Atari benchmarks showthat SF-RSSM outperforms strong baselines in policy performance.

Anomaly detection is essential for the safety and reliability of autonomous driving systems. Current methods often focus on detection accuracy but neglect response time, which is critical in time-sensitive driving scenarios. In this paper, we introduce real-time anomaly detection for autonomous driving, prioritizing both minimal response time and high accuracy. We propose a novel multimodal asynchronous hybrid network that combines event streams from event cameras with image data from RGB cameras. Our network utilizes the high temporal resolution of event cameras through an asynchronous Graph Neural Network and integrates it with spatial features extracted by a CNN from RGB images. This combination effectively captures both the temporal dynamics and spatial details of the driving environment, enabling swift and precise anomaly detection. Extensive experiments on benchmark datasets show that our approach outperforms existing methods in both accuracy and response time, achieving millisecond-level real-time performance.

Continual Generalized Category Discovery (C-GCD) aims to incrementally identify both known and novel classes from unlabeled data streams while preserving previously acquired knowledge. However, current approaches face a critical limitation we term unstructured knowledge interference, a critical issue that arises when unconstrained parameter updates entangle discriminative representations across classes, severely contaminating the feature space and introducing significant transfer and bias risks. To address these challenges, we propose the Tree of Prompts (ToP), a novel hierarchical prompting framework that facilitates structured knowledge adaptation through multi-granular parameter regulation. ToP hierarchically integrates three synergistic components: (1) Stage-level prompts preserve historical knowledge by isolating task-specific parameters, thereby mitigating conflicts between incremental tasks; (2) Centroid-level prompts disentangle category semantics through learnable prototype calibration, sharpening decision boundaries in the feature space; and (3) Context-level prompts dynamically capture discriminative local features to suppress contamination from superficial similarities. Experimental results demonstrate that ToP markedly outperforms existing methods and provides a comprehensive and efficient solution for C-GCD.

One important desideratum of lifelong learning aims to discover novel classes from unlabelled data in a continuous manner. The central challenge is twofold: discovering and learning novel classes while mitigating the issue of catastrophic forgetting of established knowledge. To this end, we introduce a new paradigm called Adaptive Discovering and Merging (ADM) to discover novel categories adaptively in the incremental stage and integrate novel knowledge into the model without affecting the original knowledge. To discover novel classes adaptively, we decouple representation learning and novel class discovery, and use Triple Comparison (TC) and Probability Regularization (PR) to constrain the probability discrepancy and diversity for adaptive category assignment. To merge the learned novel knowledge adaptively, we propose a hybrid structure with base and novel branches named Adaptive Model Merging (AMM), which reduces the interference of the novel branch on the old classes to preserve the previous knowledge, and merges the novel branch to the base model without performance loss and parameter growth. Extensive experiments on several datasets show that ADM significantly outperforms existing class-incremental Novel Class Discovery (class-iNCD) approaches. Moreover, our AMM also benefits the class-incremental Learning (class-IL) task by alleviating the catastrophic forgetting problem. The source code is included in the supplementary materials.

Accurate environment dynamics modeling is crucial for obtaining effective state representations in visual reinforcement learning (RL) applications. However, when facing multiple input modalities, existing dynamics modeling methods (e.g., DeepMDP) usually stumble in addressing the complex and volatile relationship between different modalities. In this paper, we study the problem of efficient dynamics modeling for multi-modal visual RL. We find that under the existence of modality heterogeneity, modality-correlated and distinct features are equally important but play different roles in reflecting the evolution of environmental dynamics. Motivated by this fact, we propose Dissected Dynamics Modeling (DDM), a novel multi-modal dynamics modeling method for visual RL. Unlike existing methods, DDM explicitly distinguishes consistent and inconsistent information across modalities and treats them separately with a divide-and-conquer strategy. This is done by dispatching the features carrying different information into distinct dynamics modeling pathways, which naturally form a series of implicit regularizations along the learning trajectories. In addition, a reward predictive function is further introduced to filter task-irrelevant information in both modality-consistent and inconsistent features, ensuring information integrity while avoiding potential distractions. Extensive experiments show that DDM consistently achieves competitive performance in challenging multi-modal visual environments.

Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on Deep-Mind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods.

Vision-based reinforcement learning (RL) depends on discriminative representation encoders to abstract the observation states. Despite the great success of increasing CNN parameters for many supervised computer vision tasks, reinforcement learning with temporal-difference (TD) losses cannot benefit from it in most complex environments. In this paper, we analyze that the training instability arises from the oscillating self-overfitting of the heavy-optimizable encoder. We argue that serious oscillation will occur to the parameters when enforced to fit the sensitive TD targets, causing uncertain drifting of the latent state space and thus transmitting these perturbations to the policy learning. To alleviate this phenomenon, we propose a novel asymmetric interactive cooperation approach with the interaction between a heavy-optimizable encoder and a supportive light-optimizable encoder, in which both their advantages are integrated including the highly discriminative capability as well as the training stability. We also present a greedy bootstrapping optimization to isolate the visual perturbations from policy learning, where representation and policy are trained sufficiently by turns. Finally, we demonstrate the effectiveness of our method in utilizing larger visual models by first-person highway driving task CARLA and Vizdoom environments.

Integrating RGB frames with alternative modality inputs is gaining increasing traction in many vision-based reinforcement learning (RL) applications. Existing multi-modal vision-based RL methods usually follow a Global Value Estimation (GVE) pipeline, which uses a fused modality feature to obtain a unified global environmental description. However, such a feature-level fusion paradigm with a single critic may fall short in policy learning as it tends to overlook the distinct values of each modality. To remedy this, this paper proposes a Local modality-customized Value Estimation (LVE) paradigm, which dynamically estimates the contribution and adjusts the importance weight of each modality from a value-level perspective. Furthermore, a task-contextual re-fusion process is developed to achieve a task-level re-balance of estimations from both feature and value levels. To this end, a Hierarchical Adaptive Value Estimation (HAVE) framework is formed, which adaptively coordinates the contributions of individual modalities as well as their collective efficacy. Agents trained by HAVE are able to exploit the unique characteristics of various modalities while capturing their intricate interactions, achieving substantially improved performance. We specifically highlight the potency of our approach within the challenging landscape of autonomous driving, utilizing the CARLA benchmark with neuromorphic event and depth data to demonstrate HAVE's capability and the effectiveness of its distinct components.

Generalization in image-based reinforcement learning (RL) aims to learn a robust policy that could be applied directly on unseen visual environments, which is a challenging task since agents usually tend to overfit to their training environment. To handle this problem, a natural approach is to increase the data diversity by image based augmentations. However, different with most vision tasks such as classification and detection, RL tasks are not always invariant to spatial based augmentations due to the entanglement of environment dynamics and visual appearance. In this paper, we argue with two principles for augmentations in RL: First, the augmented observations should facilitate learning a universal policy, which is robust to various distribution shifts. Second, the augmented data should be invariant to the learning signals such as action and reward. Following these rules, we revisit image-based RL tasks from the view of frequency domain and propose a novel augmentation method, namely Spectrum Random Masking (SRM),which is able to help agents to learn the whole frequency spectrum of observation for coping with various distributions and compatible with the pre-collected action and reward corresponding to original observation. Extensive experiments conducted on DMControl Generalization Benchmark demonstrate the proposed SRM achieves the state-of-the-art performance with strong generalization potentials.

Person re-identification (Re-ID) across multiple datasets is a challenging task due to two main reasons: the presence of large cross-dataset distinctions and the absence of annotated target instances. To address these two issues, this paper proposes a domain adaptive attention learning approach to reliably transfer discriminative representation from the labeled source domain to the unlabeled target domain. In this approach, a domain adaptive attention model is learned to separate the feature map into domain-shared part and domain-specific part. In this manner, the domain-shared part is used to capture transferable cues that can compensate cross-dataset distinctions and give positive contributions to the target task, while the domain-specific part aims to model the noisy information to avoid the negative transfer caused by domain diversity. A soft label loss is further employed to take full use of unlabeled target data by estimating pseudo labels. Extensive experiments on the Market-1501, DukeMTMC-reID and MSMT17 benchmarks demonstrate the proposed approach outperforms the state-of-the-arts.

This paper focuses on a novel task named masked faces recognition (MFR), which aims to match masked faces with common faces and is important especially during the global outbreak of COVID-19. It is challenging to identify masked faces for two main reasons. Firstly, there is no large-scale training data and test data with ground truth for MFR. Collecting and annotating millions of masked faces is labor-consuming. Secondly, since most facial cues are occluded by mask, it is necessary to learn representations which are both discriminative and robust to mask wearing. To handle the first challenge, this paper collects two datasets designed for MFR: MFV with 400 pairs of 200 identities for verification, and MFI which contains 4,916 images of 669 identities for identification. As is known, a robust face recognition model needs images of millions of identities to train, and hundreds of identities is far from enough. Hence, MFV and MFI are only considered as test datasets to evaluate algorithms. Besides, a data augmentation method for training data is introduced to automatically generate synthetic masked face images from existing common face datasets. In addition, a novel latent part detection (LPD) model is proposed to locate the latent facial part which is robust to mask wearing, and the latent part is further used to extract discriminative features. The proposed LPD model is trained in an end-to-end manner and only utilizes the original and synthetic training data. Experimental results on MFV, MFI and synthetic masked LFW demonstrate that LPD model generalizes well on both realistic and synthetic masked data and outperforms other methods by a large margin.

Masked faces recognition (MFR) aims to match a masked face with its corresponding full face, which is an important task especially during the global outbreak of COVID-19. However, most existing face recognition models generalize poorly in this case, and it is hard to train a robust MFR model due to two main reasons: 1) the absence of large scale training data as well as ground truth testing data, and 2) the presence of large intra-class variation between masked faces and full faces. To address the first challenge, this paper firstly contributes a new dataset denoted as MFSR, which consists of two parts. The first part contains 9,742 masked face images with mask region segmentation annotation. The second part contains 11,615 images of 1,004 identities, and each identity has masked and full face images with various orientations, lighting conditions and mask types. However, it is still not enough for training MFR models with deep learning. To obtain sufficient training data, based on the MFSR, we introduce a novel Identity Aware Mask GAN (IAMGAN) with segmentation guided multi-level identity preserve module to generate the synthetic masked face images from the full face images. In addition, to tackle the second challenge, a Domain Constrained Ranking (DCR) loss is proposed by adopting a center-based cross-domain ranking strategy. For each identity, two centers are designed which correspond to the full face images and the masked face images respectively. The DCR forces the feature of masked faces getting closer to its corresponding full face center and vice-versa. Experimental results on the MFSR dataset demonstrate the effectiveness of the proposed approaches.

Person re-identification (ReID) aims to match detected pedestrian images from multiple non-overlapping cameras. Most existing methods employ a backbone CNN to extract a vectorized feature representation by performing some global pooling operations (such as global average pooling and global max pooling) on the 3D feature map (i.e., the output of the backbone CNN). Although simple and effective in some situations, the global pooling operation only focuses on the statistical properties and ignores the spatial distribution of the feature map. Hence, it can not distinguish two feature maps when they have similar response values located in totally different positions. To handle this challenge, a novel method is proposed to learn the discriminative spatial features. Firstly, a self-constrained spatial transformer network (SC-STN) is introduced to handle the misalignments caused by detection errors. Then, based on the prior knowledge that the spatial structure of a pedestrian often keeps robust in vertical orientation of images, a novel vertical convolution network (VCN) is proposed to extract the spatial feature in vertical. Extensive experimental evaluations on several benchmarks demonstrate that the proposed method achieves state-of-the-art performances by introducing only a few parameters to the backbone.