Large language models (LLMs) have demonstrated remarkable capabilities in reasoning and code generation, raising the prospect that they could assist in developing and optimizing the very infrastructure that powers them. However, existing benchmarks mainly focus on isolated kernels, predefined operators, or pre-specified optimization targets, and therefore fail to evaluate the ability of LLMs to perform open-ended, long-horizon LLM infrastructure engineering. To address this gap, we present Φ-Bench, a benchmark for systematically evaluating LLMs on engineering the LLM infrastructure stack. Derived from optimization problems studied in frontier research and grounded in real-world code repositories, Φ-Bench provides broad coverage of the LLM infrastructure stack and spans tasks of varying complexity, ranging from localized kernel-level function completion to long-horizon implementation and end-to-end system optimization. Extensive experiments on frontier LLMs reveal their current capabilities and limitations in engineering complex LLM infrastructure, offering insights into the challenges that remain on the path toward autonomous optimization of future AI infrastructure.

Monocular depth estimation infers per-pixel depth from a single RGB image. It remains particularly challenging in far-range regions, where sparse observations and long-tailed depth distributions bias learning toward near-range content. Diffusion models offer a promising alternative to discriminative foundation models by leveraging rich generative priors for zero-shot generalization. However, existing methods still struggle to recover fine distant structures under weak supervision and limited image evidence. To address this, we propose VistaDepth, a diffusion framework for far-range depth estimation. It improves far-range prediction through two complementary mechanisms: Latent Frequency Modulation (LFM), which refines latent features via dynamic, content-aware spectral filtering, and BiasMap, which adaptively reweights the diffusion objective toward distant, structurally informative regions while remaining aligned with denoising. Extensive experiments on five benchmarks show that VistaDepth achieves state-of-the-art performance among diffusion-based methods. Despite training on orders of magnitude less data than discriminative baselines, it remains competitive globally and excels in challenging far-range regions.

Visual-inertial odometry (VIO) plays a pivotal role in providing accurate 6-DOF pose estimation for robotics and autonomous systems. However, both classical geometric pipelines and recent learning-based approaches often struggle to preserve critical visual information and to represent complex motion dynamics, particularly under nonlinear or nonstationary trajectories. To address these challenges, we propose MixVIO, an adaptive visual-inertial odometry framework that integrates refined attention mechanisms with expressive kernel representations.A policy network is first employed to determine whether visual processing is necessary at each time step, enabling conditional execution of the visual branch. When activated, visual features are enhanced by a Refined Phase Attention (RPA) module, which injects spatial context and performs channel-wise recalibration to preserve salient structures. To model complex motion dynamics, a MixCore-KAN module augments LSTM hidden states using a hybrid basis of Gaussian and Lorentzian functions, improving flexibility in representing diverse motion patterns.Extensive evaluations on outdoor and indoor benchmarks demonstrate that MixVIO generalizes robustly under diverse sensing conditions. In particular, On the KITTI odometry benchmark, it achieves an average relative translation error of 1.69% while maintaining controlled visual usage, indicating that high-precision pose estimation can be achieved without uniformly invoking dense visual computation.

When planning for autonomous driving, it is crucial to consider essential traffic elements such as lanes, intersections, traffic regulations, and dynamic agents. However, they are often overlooked by the traditional end-to-end planning methods, likely leading to inefficiencies and non-compliance with traffic regulations. In this work, we endeavor to integrate the perception of these elements into the planning task. To this end, we propose P erception H elps P lanning (PHP), a novel framework that reconciles lane-level planning with perception. This integration ensures that planning is inherently aligned with traffic constraints, facilitating safe and efficient driving. Specifically, PHP focuses on both edges of a lane for planning and perception purposes, taking into account the positions of both lane edges in Bird's Eye View (BEV), along with attributes related to lane intersections, lane directions, and lane occupancy. In the algorithmic design, the process begins with the transformer encoding multi-camera images to extract the above features and predict lane-level perception results. Next, the hierarchical feature early fusion module refines the features for predicting planning attributes. Finally, a specific interpreter utilizes a late-fusion process designed to integrate lane-level perception and planning information, culminating in generating vehicle control signals. Experiments on three Carla benchmarks show significant improvements in driving scores of 27.20%, 33.47%, and 15.54% over existing algorithms, respectively, achieving state-of-the-art performance, with the system operating up to 22.57 FPS.

SLAM is a fundamental capability of unmanned systems, with LiDAR-based SLAM gaining widespread adoption due to its high precision. Current SLAM systems can achieve centimeter-level accuracy within a short period. However, there are still several challenges when dealing with largescale mapping tasks including significant storage requirements and difficulty of reusing the constructed maps. To address this, we first design an elastic and lightweight map representation called CELLmap, composed of several CELLS, each representing the local map at the corresponding location. Then, we design a general backend including CELL-based bidirectional registration module and loop closure detection module to improve global map consistency. Our experiments have demonstrated that CELLmap can represent the precise geometric structure of large-scale maps of KITTI dataset using only about 60 MB. Additionally, our general backend achieves up to a 26.88% improvement over various LiDAR odometry methods.

Recently, 3D Gaussian Splatting (3DGS) has reshaped the field of photorealistic 3D reconstruction, achieving impressive rendering quality and speed. However, when applied to large-scale street scenes, existing methods suffer from rapidly escalating per-viewpoint reconstruction costs as scene size increases, leading to significant computational overhead. After revisiting the conventional pipeline, we identify three key factors accounting for this issue: unnecessary local-to-global transformations, excessive 3D-to-2D projections, and inefficient rendering of distant content. To address these challenges, we propose S3R-GS, a 3DGS framework that Streamlines the pipeline for large-scale Street Scene Reconstruction, effectively mitigating these limitations. Moreover, most existing street 3DGS methods rely on ground-truth 3D bounding boxes to separate dynamic and static components, but 3D bounding boxes are difficult to obtain, limiting real-world applicability. To address this, we propose an alternative solution with 2D boxes, which are easier to annotate or can be predicted by off-the-shelf vision foundation models. Such designs together make S3R-GS readily adapt to large, in-the-wild scenarios. Extensive experiments demonstrate that S3R-GS enhances rendering quality and significantly accelerates reconstruction. Remarkably, when applied to videos from the challenging Argoverse2 dataset, it achieves state-of-the-art PSNR and SSIM, reducing reconstruction time to below 50 methods.

The recent trend for multi-camera 3D object detection is through the unified bird's-eye view (BEV) representation. However, directly transforming features extracted from the image-plane view to BEV inevitably results in feature distortion, especially around the objects of interest, making the objects blur into the background. To this end, we propose OA-BEV, a network that can be plugged into the BEV-based 3D object detection framework to bring out the objects by incorporating object-aware pseudo-3D features and depth features. Such features contain information about the object's position and 3D structures. First, we explicitly guide the network to learn the depth distribution by object-level supervision from each 3D object's center. Then, we select the foreground pixels by a 2D object detector and project them into 3D space for pseudo-voxel feature encoding. Finally, the object-aware depth features and pseudo-voxel features are incorporated into the BEV representation with a deformable attention mechanism. We conduct extensive experiments on the nuScenes dataset to validate the merits of our proposed OA-BEV. Our method achieves consistent improvements over the BEV-based baselines in terms of both average precision and nuScenes detection score. Our codes will be published.

We propose PuriLight, a lightweight and efficient framework for self-supervised monocular depth estimation, to address the dual challenges of computational efficiency and detail preservation. While recent advances in self-supervised depth estimation have reduced reliance on ground truth supervision, existing approaches remain constrained by either bulky architectures compromising practicality or lightweight models sacrificing structural precision. These dual limitations underscore the critical need to develop lightweight yet structurally precise architectures. Our framework addresses these limitations through a three-stage architecture incorporating three novel modules: the Shuffle-Dilation Convolution (SDC) module for local feature extraction, the Rotation-Adaptive Kernel Attention (RAKA) module for hierarchical feature enhancement, and the Deep Frequency Signal Purification (DFSP) module for global feature purification. Through effective collaboration, these modules enable PuriLight to achieve both lightweight and accurate feature extraction and processing. Extensive experiments demonstrate that PuriLight achieves state-of-the-art performance with minimal training parameters while maintaining exceptional computational efficiency. Codes will be available at https://github.com/ishrouder/PuriLight.

Extrinsic calibration is a fundamental step in sensor fusion systems. However, existing methods often lack generalization capabilities when facing diverse hardware configurations, sensor poses, and environmental conditions, hindering their large-scale deployment. To address this limitation, we propose a general extrinsic calibration method, CalibWorkflow. Our core innovation lies in positioning multimodal large language models (MLLMs) as "visual guides" for the calibration process, leveraging their powerful vision-language understanding capabilities to guide parameter search and refinement. This reliance on visual scene understanding, rather than specific geometric features or sensor characteristics, enables the method to generalize effectively across diverse hardware and environmental conditions. Specifically, CalibWorkflow employs a three-stage calibration pipeline: initial parameter search, coarse optimization, and fine optimization. First, it utilizes the MLLM to assess the visual consistency between the projected point cloud and the image, rapidly determining an initial range for the extrinsic parameters. Next, the MLLM serves as a differential evaluator, giving simple "better" or "worse" feedback on parameter changes to guide the search through the parameter space. Finally, the method refines the calibration by matching edge features and performing non-linear optimization. Extensive experiments are conducted across six diverse scenarios and four heterogeneous sensor combinations. CalibWorkflow achieves state-of-the-art sub-degree and centimeter-level accuracy on four datasets and demonstrates highly competitive performance on others. These results thoroughly validate the generalization and robustness when facing various scenarios. Codes will be available.

Autonomous systems often employ multiple LiDARs to leverage the integrated advantages, enhancing perception and robustness. The most critical prerequisite under this setting is the estimating the extrinsic between each LiDAR, i.e., calibration. Despite the exciting progress in multi-LiDAR calibration efforts, a universal, sensor-agnostic calibration method remains elusive. According to the coarse-to-fine framework, we first design a spherical descriptor TERRA for 3-DoF rotation initialization with no prior knowledge. To further optimize, we present JEEP for the joint estimation of extrinsic and pose, integrating geometric and motion information to overcome factors affecting the point cloud registration. Finally, the LiDAR poses optimized by the hierarchical optimization module are input to time syn- chronization module to produce the ultimate calibration results, including the time offset. To verify the effectiveness, we conduct extensive experiments on eight datasets, where 16 diverse types of LiDARs in total and dozens of calibration tasks are tested. In the challenging tasks, the calibration errors can still be controlled within 5cm and 1 with a high success rate.

Fusing the data of millimeter-wave Radar sensors and high-definition cameras has emerged as a viable approach to achieving precise 3D object detection for roadside traffic surveillance. For roadside perception systems, earlier studies have pointed out that it is better to perform the fusion on the 2D image plane than on the BEV plane (which is popular for on-car perception systems), especially when the perception range is large (e.g., > 150 m). Image-plane fusion requires critical transformations, like perspective projection from the Radar's BEV to the camera's 2D plane and reverse IPM. However, real-world issues like uneven terrain and sensor movement degrade these transformations' precision, impacting fusion effectiveness. To alleviate these issues, we propose a geometry-based Radar-camera fusion method on the ground, namely FARFusion V2. Specifically, we extend the ground-plane assumption in FARFusion [20] to support arbitrary shapes by formulating the ground height as an implicit representation based on geometric transformations. By incorporating the ground information, we can enhance Radar data with target height measurements. Consequently, we can thus project the enhanced Radar data onto the 2D plane to obtain more accurate depth information, thereby assisting the IPM process. A real-time parameterized transformation parameters estimation module is further introduced to refine the view transformation processes. Moreover, considering various measurement noises across these two sensors, we introduce an uncertainty-based depth fusion strategy into the 2D fusion process to maximize the probability of obtaining the optimal depth value. Extensive experiments are conducted on our collected roadside OWL benchmark, demonstrating the excellent localization capacity of FARFusion V2 in far-range scenarios. Our method achieves an average location accuracy of 0.771m when we extend the detection range up to 500m.

Multi-modal systems enhance performance in autonomous driving but face inefficiencies due to indiscriminate processing within each modality. Additionally, the independent feature learning of each modality lacks interaction, which results in extracted features that do not possess the complementary characteristics. These issue increases the cost of fusing redundant information across modalities. To address these challenges, we propose targeting driving-relevant elements, which reduces the volume of LiDAR features while preserving critical information. This approach enhances lane level interaction between the image and LiDAR branches, allowing for the extraction and fusion of their respective advantageous features. Building upon the camera-only framework PHP, we introduce the Lane-level camera-LiDAR Fusion Planning (LFP) method, which balances efficiency with performance by using lanes as the unit for sensor fusion. Specifically, we design three modules to enhance efficiency and performance. For efficiency, we propose an image-guided coarse lane prior generation module that forecasts the region of interest (ROI) for lanes and assigns a confidence score, guiding LiDAR processing. The LiDAR feature extraction modules leverages lane-aware priors from the image branch to guide sampling for pillar, retaining essential pillars. For performance, the lane-level cross-modal query integration and feature enhancement module uses confidence score from ROI to combine low-confidence image queries with LiDAR queries, extracting complementary depth features. These features enhance the low-confidence image features, compensating for the lack of depth. Experiments on the Carla benchmarks show that our method achieves state-of-the-art performance in both driving score and infraction score, with maximum improvement of 15 high frame rate of 19.27 FPS.

The recent advances in query-based multi-camera 3D object detection are featured by initializing object queries in the 3D space, and then sampling features from perspective-view images to perform multi-round query refinement. In such a framework, query points near the same camera ray are likely to sample similar features from very close pixels, resulting in ambiguous query features and degraded detection accuracy. To this end, we introduce RayFormer, a camera-ray-inspired query-based 3D object detector that aligns the initialization and feature extraction of object queries with the optical characteristics of cameras. Specifically, RayFormer transforms perspective-view image features into bird's eye view (BEV) via the lift-splat-shoot method and segments the BEV map to sectors based on the camera rays. Object queries are uniformly and sparsely initialized along each camera ray, facilitating the projection of different queries onto different areas in the image to extract distinct features. Besides, we leverage the instance information of images to supplement the uniformly initialized object queries by further involving additional queries along the ray from 2D object detection boxes. To extract unique object-level features that cater to distinct queries, we design a ray sampling method that suitably organizes the distribution of feature sampling points on both images and bird's eye view. Extensive experiments are conducted on the nuScenes dataset to validate our proposed ray-inspired model design. The proposed RayFormer achieves 55.5% mAP and 63.3% NDS, respectively.

Tensor program tuning is a non-convex objective optimization problem, to which search-based approaches have proven to be effective. At the core of the search-based approaches lies the design of the cost model. Though deep learning-based cost models perform significantly better than other methods, they still fall short and suffer from the following problems. First, their feature extraction heavily relies on expert-level domain knowledge in hardware architectures. Even so, the extracted features are often unsatisfactory and require separate considerations for CPUs and GPUs. Second, a cost model trained on one hardware platform usually performs poorly on another, a problem we call cross-hardware unavailability. In order to address these problems, we propose TLP and MTLTLP. TLP is a deep learning-based cost model that facilitates tensor program tuning. Instead of extracting features from the tensor program itself, TLP extracts features from the schedule primitives. We treat schedule primitives as tensor languages. TLP is thus a Tensor Language Processing task. In this way, the task of predicting the tensor program latency through the cost model is transformed into a natural language processing (NLP) regression task. MTL-TLP combines Multi-Task Learning and TLP to cope with the cross-hardware unavailability problem. We incorporate these techniques into the Ansor framework and conduct detailed experiments. Results show that TLP can speed up the average search time by 9.1X and 3.0X on CPU and GPU workloads, respectively, compared to the state-of-the-art implementation. MTL-TLP can achieve a speed-up of 4.7X and 2.9X on CPU and GPU workloads, respectively, using only 7% of the target hardware data.

It is important for deep reinforcement learning (DRL) algorithms to transfer their learned policies to new environments that have different visual inputs. In this paper, we introduce Prompt based Proximal Policy Optimization ($P^{3}O$), a three-stage DRL algorithm that transfers visual representations from a target to a source environment by applying prompting. The process of $P^{3}O$ consists of three stages: pre-training, prompting, and predicting. In particular, we specify a prompt-transformer for representation conversion and propose a two-step training process to train the prompt-transformer for the target environment, while the rest of the DRL pipeline remains unchanged. We implement $P^{3}O$ and evaluate it on the OpenAI CarRacing video game. The experimental results show that $P^{3}O$ outperforms the state-of-the-art visual transferring schemes. In particular, $P^{3}O$ allows the learned policies to perform well in environments with different visual inputs, which is much more effective than retraining the policies in these environments.

As cameras are increasingly deployed in new application domains such as autonomous driving, performing 3D object detection on monocular images becomes an important task for visual scene understanding. Recent advances on monocular 3D object detection mainly rely on the "pseudo-LiDAR'' generation, which performs monocular depth estimation and lifts the 2D pixels to pseudo 3D points. However, depth estimation from monocular images, due to its poor accuracy, leads to inevitable position shift of pseudo-LiDAR points within the object. Therefore, the predicted bounding boxes may suffer from inaccurate location and deformed shape. In this paper, we present a novel neighbor-voting method that incorporates neighbor predictions to ameliorate object detection from severely deformed pseudo-LiDAR point clouds. Specifically, each feature point around the object forms their own predictions, and then the "consensus'' is achieved through voting. In this way, we can effectively combine the neighbors' predictions with local prediction and achieve more accurate 3D detection. To further enlarge the difference between the foreground region of interest (ROI) pseudo-LiDAR points and the background points, we also encode the ROI prediction scores of 2D foreground pixels into the corresponding pseudo-LiDAR points. We conduct extensive experiments on the KITTI benchmark to validate the merits of our proposed method. Our results on the bird's eye view detection outperform the state-of-the-art performance, especially for the "hard" level detection. The code is available at https://github.com/cxmomo/Neighbor-Vote.

As cameras are increasingly deployed in new application domains such as autonomous driving, performing 3D object detection on monocular images becomes an important task for visual scene understanding. Recent advances on monocular 3D object detection mainly rely on the "pseudo-LiDAR'' generation, which performs monocular depth estimation and lifts the 2D pixels to pseudo 3D points. However, depth estimation from monocular images, due to its poor accuracy, leads to inevitable position shift of pseudo-LiDAR points within the object. Therefore, the predicted bounding boxes may suffer from inaccurate location and deformed shape. In this paper, we present a novel neighbor-voting method that incorporates neighbor predictions to ameliorate object detection from severely deformed pseudo-LiDAR point clouds. Specifically, each feature point around the object forms their own predictions, and then the "consensus'' is achieved through voting. In this way, we can effectively combine the neighbors' predictions with local prediction and achieve more accurate 3D detection. To further enlarge the difference between the foreground region of interest (ROI) pseudo-LiDAR points and the background points, we also encode the ROI prediction scores of 2D foreground pixels into the corresponding pseudo-LiDAR points. We conduct extensive experiments on the KITTI benchmark to validate the merits of our proposed method. Our results on the bird's eye view detection outperform the state-of-the-art performance, especially for the "hard" level detection. The code is available at https://github.com/cxmomo/Neighbor-Vote.