Huijing Zhao

Ziyang Yao, Zeyu Zhu, YunCheng Jiang et al.

Discrete visual tokens should provide a compact representation for both token-based world modeling and planning in autonomous driving. However, most tokenizers are inherited from image generation and are optimized mainly for pixel reconstruction, which may leave a gap between what is easy to generate and what is useful to decode for driving decisions. We present a representation-guided and geometry-enhanced tokenizer that learns discrete tokens under joint supervision. The tokenizer aligns its discrete bottleneck with a frozen DINO feature space through feature decoding, while preserving appearance via RGB reconstruction with perceptual and adversarial losses. To inject geometric state-related cues, we add adjacent-frame depth and relative-pose supervision during training and stabilize joint objectives with multi-codebook quantization. We evaluate the same learned tokens with a lightweight planning readout and a GPT-style next-token world model. Experiments on NAVSIM show improved reconstruction fidelity and representation consistency, competitive planning performance under a fixed decoder, and better generative quality under matched settings.

Yancheng Pan, Fan Xie, Huijing Zhao

3D semantic segmentation (3DSS) is an essential process in the creation of a safe autonomous driving system. However, deep learning models for 3D semantic segmentation often suffer from the class imbalance problem and out-of-distribution (OOD) data. In this study, we explore how the class imbalance problem affects 3DSS performance and whether the model can detect the category prediction correctness, or whether data is ID (in-distribution) or OOD. For these purposes, we conduct two experiments using three representative 3DSS models and five trust scoring methods, and conduct both a confusion and feature analysis of each class. Furthermore, a data augmentation method for the 3D LiDAR dataset is proposed to create a new dataset based on SemanticKITTI and SemanticPOSS, called AugKITTI. We propose the wPre metric and TSD for a more in-depth analysis of the results, and follow are proposals with an insightful discussion. Based on the experimental results, we find that: (1) the classes are not only imbalanced in their data size but also in the basic properties of each semantic category. (2) The intraclass diversity and interclass ambiguity make class learning difficult and greatly limit the models' performance, creating the challenges of semantic and data gaps. (3) The trust scores are unreliable for classes whose features are confused with other classes. For 3DSS models, those misclassified ID classes and OODs may also be given high trust scores, making the 3DSS predictions unreliable, and leading to the challenges in judging 3DSS result trustworthiness. All of these outcomes point to several research directions for improving the performance and reliability of the 3DSS models used for real-world applications.

Zhezhang Ding, Huijing Zhao

Precise trajectory prediction in complex driving scenarios is essential for autonomous vehicles. In practice, different driving scenarios present varying levels of difficulty for trajectory prediction models. However, most existing research focuses on the average precision of prediction results, while ignoring the underlying distribution of the input scenarios. This paper proposes a critical example mining method that utilizes a data-driven approach to estimate the rareness of the trajectories. By combining the rareness estimation of observations with whole trajectories, the proposed method effectively identifies a subset of data that is relatively hard to predict BEFORE feeding them to a specific prediction model. The experimental results show that the mined subset has higher prediction error when applied to different downstream prediction models, which reaches +108.1% error (greater than two times compared to the average on dataset) when mining 5% samples. Further analysis indicates that the mined critical examples include uncommon cases such as sudden brake and cancelled lane-change, which helps to better understand and improve the performance of prediction models.

Haotian Zhou, Xiaole Wang, He Li et al.

Navigating to a designated goal using visual information is a fundamental capability for intelligent robots. Most classical visual navigation methods are restricted to single-goal, single-modality, and closed set goal settings. To address the practical demands of multi-modal, open-vocabulary goal queries and multi-goal visual navigation, we propose LagMemo, a navigation system that leverages a language 3D Gaussian Splatting memory. During exploration, LagMemo constructs a unified 3D language memory. With incoming task goals, the system queries the memory, predicts candidate goal locations, and integrates a local perception-based verification mechanism to dynamically match and validate goals during navigation. For fair and rigorous evaluation, we curate GOAT-Core, a high-quality core split distilled from GOAT-Bench tailored to multi-modal open-vocabulary multi-goal visual navigation. Experimental results show that LagMemo's memory module enables effective multi-modal open-vocabulary goal localization, and that LagMemo outperforms state-of-the-art methods in multi-goal visual navigation. Project page: https://weekgoodday.github.io/lagmemo

Zeyu Zhu, Huijing Zhao

In recent years, great efforts have been devoted to deep imitation learning for autonomous driving control, where raw sensory inputs are directly mapped to control actions. However, navigating through densely populated intersections remains a challenging task due to uncertainty caused by uncertain traffic participants. We focus on autonomous navigation at crowded intersections that require interaction with pedestrians. A multi-task conditional imitation learning framework is proposed to adapt both lateral and longitudinal control tasks for safe and efficient interaction. A new benchmark called IntersectNav is developed and human demonstrations are provided. Empirical results show that the proposed method can achieve a success rate gain of up to 30% compared to the state-of-the-art.

Biao Gao, Xijun Zhao, Huijing Zhao

Off-road semantic segmentation with fine-grained labels is necessary for autonomous vehicles to understand driving scenes, as the coarse-grained road detection can not satisfy off-road vehicles with various mechanical properties. Fine-grained semantic segmentation in off-road scenes usually has no unified category definition due to ambiguous nature environments, and the cost of pixel-wise labeling is extremely high. Furthermore, semantic properties of off-road scenes can be very changeable due to various precipitations, temperature, defoliation, etc. To address these challenges, this research proposes an active and contrastive learning-based method that does not rely on pixel-wise labels, but only on patch-based weak annotations for model learning. There is no need for predefined semantic categories, the contrastive learning-based feature representation and adaptive clustering will discover the category model from scene data. In order to actively adapt to new scenes, a risk evaluation method is proposed to discover and select hard frames with high-risk predictions for supplemental labeling, so as to update the model efficiently. Experiments conducted on our self-developed off-road dataset and DeepScene dataset demonstrate that fine-grained semantic segmentation can be learned with only dozens of weakly labeled frames, and the model can efficiently adapt across scenes by weak supervision, while achieving almost the same level of performance as typical fully supervised baselines.

Shaochi Hu, Hanwei Fan, Biao Gao et al.

Categorizing driving scenes via visual perception is a key technology for safe driving and the downstream tasks of autonomous vehicles. Traditional methods infer scene category by detecting scene-related objects or using a classifier that is trained on large datasets of fine-labeled scene images. Whereas at cluttered dynamic scenes such as campus or park, human activities are not strongly confined by rules, and the functional attributes of places are not strongly correlated with objects. So how to define, model and infer scene categories is crucial to make the technique really helpful in assisting a robot to pass through the scene. This paper proposes a method of task-driven driving scene categorization using weakly supervised data. Given a front-view video of a driving scene, a set of anchor points is marked by following the decision making of a human driver, where an anchor point is not a semantic label but an indicator meaning the semantic attribute of the scene is different from that of the previous one. A measure is learned to discriminate the scenes of different semantic attributes via contrastive learning, and a driving scene profiling and categorization method is developed based on that measure. Experiments are conducted on a front-view video that is recorded when a vehicle passed through the cluttered dynamic campus of Peking University. The scenes are categorized into straight road, turn road and alerting traffic. The results of semantic scene similarity learning and driving scene categorization are extensively studied, and positive result of scene categorization is 97.17 \% on the learning video and 85.44\% on the video of new scenes.

Biao Gao, Shaochi Hu, Xijun Zhao et al.

Road detection or traversability analysis has been a key technique for a mobile robot to traverse complex off-road scenes. The problem has been mainly formulated in early works as a binary classification one, e.g. associating pixels with road or non-road labels. Whereas understanding scenes with fine-grained labels are needed for off-road robots, as scenes are very diverse, and the various mechanical performance of off-road robots may lead to different definitions of safe regions to traverse. How to define and annotate fine-grained labels to achieve meaningful scene understanding for a robot to traverse off-road is still an open question. This research proposes a contrastive learning based method. With a set of human-annotated anchor patches, a feature representation is learned to discriminate regions with different traversability, a method of fine-grained semantic segmentation and mapping is subsequently developed for off-road scene understanding. Experiments are conducted on a dataset of three driving segments that represent very diverse off-road scenes. An anchor accuracy of 89.8% is achieved by evaluating the matching with human-annotated image patches in cross-scene validation. Examined by associated 3D LiDAR data, the fine-grained segments of visual images are demonstrated to have different levels of toughness and terrain elevation, which represents their semantical meaningfulness. The resultant maps contain both fine-grained labels and confidence values, providing rich information to support a robot traversing complex off-road scenes.

Zeyu Zhu, Huijing Zhao

Autonomous driving (AD) agents generate driving policies based on online perception results, which are obtained at multiple levels of abstraction, e.g., behavior planning, motion planning and control. Driving policies are crucial to the realization of safe, efficient and harmonious driving behaviors, where AD agents still face substantial challenges in complex scenarios. Due to their successful application in fields such as robotics and video games, the use of deep reinforcement learning (DRL) and deep imitation learning (DIL) techniques to derive AD policies have witnessed vast research efforts in recent years. This paper is a comprehensive survey of this body of work, which is conducted at three levels: First, a taxonomy of the literature studies is constructed from the system perspective, among which five modes of integration of DRL/DIL models into an AD architecture are identified. Second, the formulations of DRL/DIL models for conducting specified AD tasks are comprehensively reviewed, where various designs on the model state and action spaces and the reinforcement learning rewards are covered. Finally, an in-depth review is conducted on how the critical issues of AD applications regarding driving safety, interaction with other traffic participants and uncertainty of the environment are addressed by the DRL/DIL models. To the best of our knowledge, this is the first survey to focus on AD policy learning using DRL/DIL, which is addressed simultaneously from the system, task-driven and problem-driven perspectives. We share and discuss findings, which may lead to the investigation of various topics in the future.

Biao Gao, Yancheng Pan, Chengkun Li et al.

3D semantic segmentation is a fundamental task for robotic and autonomous driving applications. Recent works have been focused on using deep learning techniques, whereas developing fine-annotated 3D LiDAR datasets is extremely labor intensive and requires professional skills. The performance limitation caused by insufficient datasets is called data hunger problem. This research provides a comprehensive survey and experimental study on the question: are we hungry for 3D LiDAR data for semantic segmentation? The studies are conducted at three levels. First, a broad review to the main 3D LiDAR datasets is conducted, followed by a statistical analysis on three representative datasets to gain an in-depth view on the datasets' size and diversity, which are the critical factors in learning deep models. Second, a systematic review to the state-of-the-art 3D semantic segmentation is conducted, followed by experiments and cross examinations of three representative deep learning methods to find out how the size and diversity of the datasets affect deep models' performance. Finally, a systematic survey to the existing efforts to solve the data hunger problem is conducted on both methodological and dataset's viewpoints, followed by an insightful discussion of remaining problems and open questions To the best of our knowledge, this is the first work to analyze the data hunger problem for 3D semantic segmentation using deep learning techniques that are addressed in the literature review, statistical analysis, and cross-dataset and cross-algorithm experiments. We share findings and discussions, which may lead to potential topics in future works.

Donghao Xu, Zhezhang Ding, Xu He et al.

Driving in a human-like manner is important for an autonomous vehicle to be a smart and predictable traffic participant. To achieve this goal, parameters of the motion planning module should be carefully tuned, which needs great effort and expert knowledge. In this study, a method of learning cost parameters of a motion planner from naturalistic driving data is proposed. The learning is achieved by encouraging the selected trajectory to approximate the human driving trajectory under the same traffic situation. The employed motion planner follows a widely accepted methodology that first samples candidate trajectories in the trajectory space, then select the one with minimal cost as the planned trajectory. Moreover, in addition to traditional factors such as comfort, efficiency and safety, the cost function is proposed to incorporate incentive of behavior decision like a human driver, so that both lane change decision and motion planning are coupled into one framework. Two types of lane incentive cost -- heuristic and learning based -- are proposed and implemented. To verify the validity of the proposed method, a data set is developed by using the naturalistic trajectory data of human drivers collected on the motorways in Beijing, containing samples of lane changes to the left and right lanes, and car followings. Experiments are conducted with respect to both lane change decision and motion planning, and promising results are achieved.

Donghao Xu, Zhezhang Ding, Chenfeng Tu et al.

Intra-driver and inter-driver heterogeneity has been confirmed to exist in human driving behaviors by many studies. In this study, a joint model of the two types of heterogeneity in car-following behavior is proposed as an approach of driver profiling and identification. It is assumed that all drivers share a pool of driver states; under each state a car-following data sequence obeys a specific probability distribution in feature space; each driver has his/her own probability distribution over the states, called driver profile, which characterize the intradriver heterogeneity, while the difference between the driver profile of different drivers depict the inter-driver heterogeneity. Thus, the driver profile can be used to distinguish a driver from others. Based on the assumption, a stochastic car-following model is proposed to take both intra-driver and inter-driver heterogeneity into consideration, and a method is proposed to jointly learn parameters in behavioral feature extractor, driver states and driver profiles. Experiments demonstrate the performance of the proposed method in driver identification on naturalistic car-following data: accuracy of 82.3% is achieved in an 8-driver experiment using 10 car-following sequences of duration 15 seconds for online inference. The potential of fast registration of new drivers are demonstrated and discussed.

Shaochi Hu, Donghao Xu, Huijing Zhao

This work addresses on the following problem: given a set of unsynchronized history observations of two scenes that are correlative on their dynamic changes, the purpose is to learn a cross-scene predictor, so that with the observation of one scene, a robot can onlinely predict the dynamic state of another. A method is proposed to solve the problem via modeling dynamic correlation using latent space shared auto-encoders. Assuming that the inherent correlation of scene dynamics can be represented by shared latent space, where a common latent state is reached if the observations of both scenes are at an approximate time, a learning model is developed by connecting two auto-encoders through the latent space, and a prediction model is built by concatenating the encoder of the input scene with the decoder of the target one. Simulation datasets are generated imitating the dynamic flows at two adjacent gates of a campus, where the dynamic changes are triggered by a common working and teaching schedule. Similar scenarios can also be found at successive intersections on a single road, gates of a subway station, etc. Accuracy of cross-scene prediction is examined at various conditions of scene correlation and pairwise observations. Potentials of the proposed method are demonstrated by comparing with conventional end-to-end methods and linear predictions.

Jilin Mei, Huijing Zhao

We propose a graph neural network(GNN) based method to incorporate scene context for the semantic segmentation of 3D LiDAR data. The problem is defined as building a graph to represent the topology of a center segment with its neighborhoods, then inferring the segment label. The node of graph is generated from the segment on range image, which is suitable for both sparse and dense point cloud. Edge weights that evaluate the correlations of center node and its neighborhoods are automatically encoded by a neural network, therefore the number of neighbor nodes is no longer a sensitive parameter. A system consists of segment generation, graph building, edge weight estimation, node updating, and node prediction is designed. Quantitative evaluation on a dataset of dynamic scene shows that our method has better performance than unary CNN with 8% improvement, as well as normal GNN with 17% improvement.

Xiaoliang Ju, Donghao Xu, Huijing Zhao

Localization is an essential technique in mobile robotics. In a complex environment, it is necessary to fuse different localization modules to obtain more robust results, in which the error model plays a paramount role. However, exteroceptive sensor-based odometries (ESOs), such as LiDAR/visual odometry, often deliver results with scene-related error, which is difficult to model accurately. To address this problem, this research designs a scene-aware error model for ESO, based on which a multimodal localization fusion framework is developed. In addition, an end-to-end learning method is proposed to train this error model using sparse global poses such as GPS/IMU results. The proposed method is realized for error modeling of LiDAR/visual odometry, and the results are fused with dead reckoning to examine the performance of vehicle localization. Experiments are conducted using both simulation and real-world data of experienced and unexperienced environments, and the experimental results demonstrate that with the learned scene-aware error models, vehicle localization accuracy can be largely improved and shows adaptiveness in unexperienced scenes.

Biao Gao, Anran Xu, Yancheng Pan et al.

We propose a method for off-road drivable area extraction using 3D LiDAR data with the goal of autonomous driving application. A specific deep learning framework is designed to deal with the ambiguous area, which is one of the main challenges in the off-road environment. To reduce the considerable demand for human-annotated data for network training, we utilize the information from vast quantities of vehicle paths and auto-generated obstacle labels. Using these autogenerated annotations, the proposed network can be trained using weakly supervised or semi-supervised methods, which can achieve better performance with fewer human annotations. The experiments on our dataset illustrate the reasonability of our framework and the validity of our weakly and semi-supervised methods.

Yancheng Pan, Biao Gao, Jilin Mei et al.

3D semantic segmentation is one of the key tasks for autonomous driving system. Recently, deep learning models for 3D semantic segmentation task have been widely researched, but they usually require large amounts of training data. However, the present datasets for 3D semantic segmentation are lack of point-wise annotation, diversiform scenes and dynamic objects. In this paper, we propose the SemanticPOSS dataset, which contains 2988 various and complicated LiDAR scans with large quantity of dynamic instances. The data is collected in Peking University and uses the same data format as SemanticKITTI. In addition, we evaluate several typical 3D semantic segmentation models on our SemanticPOSS dataset. Experimental results show that SemanticPOSS can help to improve the prediction accuracy of dynamic objects as people, car in some degree. SemanticPOSS will be published at \url{www.poss.pku.edu.cn}.

Zeyu Zhu, Nan Li, Ruoyu Sun et al.

Terrain traversability analysis is a fundamental issue to achieve the autonomy of a robot at off-road environments. Geometry-based and appearance-based methods have been studied in decades, while behavior-based methods exploiting learning from demonstration (LfD) are new trends. Behavior-based methods learn cost functions that guide trajectory planning in compliance with experts' demonstrations, which can be more scalable to various scenes and driving behaviors. This research proposes a method of off-road traversability analysis and trajectory planning using Deep Maximum Entropy Inverse Reinforcement Learning. To incorporate vehicle's kinematics while solving the problem of exponential increase of state-space complexity, two convolutional neural networks, i.e., RL ConvNet and Svf ConvNet, are developed to encode kinematics into convolution kernels and achieve efficient forward reinforcement learning. We conduct experiments in off-road environments. Scene maps are generated using 3D LiDAR data, and expert demonstrations are either the vehicle's real driving trajectories at the scene or synthesized ones to represent specific behaviors such as crossing negative obstacles. Different cost functions of traversability analysis are learned and tested at various scenes of capability in guiding the trajectory planning of different behaviors. We also demonstrate the performance and computation efficiency of the proposed method.

Jilin Mei, Huijing Zhao

This work studies semantic segmentation using 3D LiDAR data. Popular deep learning methods applied for this task require a large number of manual annotations to train the parameters. We propose a new method that makes full use of the advantages of traditional methods and deep learning methods via incorporating human domain knowledge into the neural network model to reduce the demand for large numbers of manual annotations and improve the training efficiency. We first pretrain a model with autogenerated samples from a rule-based classifier so that human knowledge can be propagated into the network. Based on the pretrained model, only a small set of annotations is required for further fine-tuning. Quantitative experiments show that the pretrained model achieves better performance than random initialization in almost all cases; furthermore, our method can achieve similar performance with fewer manual annotations.

Jilin Mei, Biao Gao, Donghao Xu et al.

This work studies the semantic segmentation of 3D LiDAR data in dynamic scenes for autonomous driving applications. A system of semantic segmentation using 3D LiDAR data, including range image segmentation, sample generation, inter-frame data association, track-level annotation and semi-supervised learning, is developed. To reduce the considerable requirement of fine annotations, a CNN-based classifier is trained by considering both supervised samples with manually labeled object classes and pairwise constraints, where a data sample is composed of a segment as the foreground and neighborhood points as the background. A special loss function is designed to account for both annotations and constraints, where the constraint data are encouraged to be assigned to the same semantic class. A dataset containing 1838 frames of LiDAR data, 39934 pairwise constraints and 57927 human annotations is developed. The performance of the method is examined extensively. Qualitative and quantitative experiments show that the combination of a few annotations and large amount of constraint data significantly enhances the effectiveness and scene adaptability, resulting in greater than 10% improvement