Bencheng Liao

Bo Jiang, Shaoyu Chen, Qing Xu et al.

Autonomous driving requires a comprehensive understanding of the surrounding environment for reliable trajectory planning. Previous works rely on dense rasterized scene representation (e.g., agent occupancy and semantic map) to perform planning, which is computationally intensive and misses the instance-level structure information. In this paper, we propose VAD, an end-to-end vectorized paradigm for autonomous driving, which models the driving scene as a fully vectorized representation. The proposed vectorized paradigm has two significant advantages. On one hand, VAD exploits the vectorized agent motion and map elements as explicit instance-level planning constraints which effectively improves planning safety. On the other hand, VAD runs much faster than previous end-to-end planning methods by getting rid of computation-intensive rasterized representation and hand-designed post-processing steps. VAD achieves state-of-the-art end-to-end planning performance on the nuScenes dataset, outperforming the previous best method by a large margin. Our base model, VAD-Base, greatly reduces the average collision rate by 29.0% and runs 2.5x faster. Besides, a lightweight variant, VAD-Tiny, greatly improves the inference speed (up to 9.3x) while achieving comparable planning performance. We believe the excellent performance and the high efficiency of VAD are critical for the real-world deployment of an autonomous driving system. Code and models are available at https://github.com/hustvl/VAD for facilitating future research.

Bencheng Liao, Shaoyu Chen, Xinggang Wang et al.

High-definition (HD) map provides abundant and precise environmental information of the driving scene, serving as a fundamental and indispensable component for planning in autonomous driving system. We present MapTR, a structured end-to-end Transformer for efficient online vectorized HD map construction. We propose a unified permutation-equivalent modeling approach, i.e., modeling map element as a point set with a group of equivalent permutations, which accurately describes the shape of map element and stabilizes the learning process. We design a hierarchical query embedding scheme to flexibly encode structured map information and perform hierarchical bipartite matching for map element learning. MapTR achieves the best performance and efficiency with only camera input among existing vectorized map construction approaches on nuScenes dataset. In particular, MapTR-nano runs at real-time inference speed ($25.1$ FPS) on RTX 3090, $8\times$ faster than the existing state-of-the-art camera-based method while achieving $5.0$ higher mAP. Even compared with the existing state-of-the-art multi-modality method, MapTR-nano achieves $0.7$ higher mAP, and MapTR-tiny achieves $13.5$ higher mAP and $3\times$ faster inference speed. Abundant qualitative results show that MapTR maintains stable and robust map construction quality in complex and various driving scenes. MapTR is of great application value in autonomous driving. Code and more demos are available at \url{https://github.com/hustvl/MapTR}.

Bencheng Liao, Shaoyu Chen, Yunchi Zhang et al.

High-definition (HD) map provides abundant and precise static environmental information of the driving scene, serving as a fundamental and indispensable component for planning in autonomous driving system. In this paper, we present \textbf{Map} \textbf{TR}ansformer, an end-to-end framework for online vectorized HD map construction. We propose a unified permutation-equivalent modeling approach, \ie, modeling map element as a point set with a group of equivalent permutations, which accurately describes the shape of map element and stabilizes the learning process. We design a hierarchical query embedding scheme to flexibly encode structured map information and perform hierarchical bipartite matching for map element learning. To speed up convergence, we further introduce auxiliary one-to-many matching and dense supervision. The proposed method well copes with various map elements with arbitrary shapes. It runs at real-time inference speed and achieves state-of-the-art performance on both nuScenes and Argoverse2 datasets. Abundant qualitative results show stable and robust map construction quality in complex and various driving scenes. Code and more demos are available at \url{https://github.com/hustvl/MapTR} for facilitating further studies and applications.

Bencheng Liao, Shaoyu Chen, Bo Jiang et al.

Online lane graph construction is a promising but challenging task in autonomous driving. Previous methods usually model the lane graph at the pixel or piece level, and recover the lane graph by pixel-wise or piece-wise connection, which breaks down the continuity of the lane and results in suboptimal performance. Human drivers focus on and drive along the continuous and complete paths instead of considering lane pieces. Autonomous vehicles also require path-specific guidance from lane graph for trajectory planning. We argue that the path, which indicates the traffic flow, is the primitive of the lane graph. Motivated by this, we propose to model the lane graph in a novel path-wise manner, which well preserves the continuity of the lane and encodes traffic information for planning. We present a path-based online lane graph construction method, termed LaneGAP, which end-to-end learns the path and recovers the lane graph via a Path2Graph algorithm. We qualitatively and quantitatively demonstrate the superior accuracy and efficiency of LaneGAP over conventional pixel-based and piece-based methods on the challenging nuScenes and Argoverse2 datasets under controllable and fair conditions. Compared to the recent state-of-the-art piece-wise method TopoNet on the OpenLane-V2 dataset, LaneGAP still outperforms by 1.6 mIoU, further validating the effectiveness of path-wise modeling. Abundant visualizations in the supplementary material show LaneGAP can cope with diverse traffic conditions. Code is released at \url{https://github.com/hustvl/LaneGAP}.

Shaoyu Chen, Yunchi Zhang, Bencheng Liao et al.

High-definition (HD) map serves as the essential infrastructure of autonomous driving. In this work, we build up a systematic vectorized map annotation framework (termed VMA) for efficiently generating HD map of large-scale driving scene. We design a divide-and-conquer annotation scheme to solve the spatial extensibility problem of HD map generation, and abstract map elements with a variety of geometric patterns as unified point sequence representation, which can be extended to most map elements in the driving scene. VMA is highly efficient and extensible, requiring negligible human effort, and flexible in terms of spatial scale and element type. We quantitatively and qualitatively validate the annotation performance on real-world urban and highway scenes, as well as NYC Planimetric Database. VMA can significantly improve map generation efficiency and require little human effort. On average VMA takes 160min for annotating a scene with a range of hundreds of meters, and reduces 52.3% of the human cost, showing great application value. Code: https://github.com/hustvl/VMA.

Bo Jiang, Shaoyu Chen, Xinggang Wang et al.

Motion prediction is highly relevant to the perception of dynamic objects and static map elements in the scenarios of autonomous driving. In this work, we propose PIP, the first end-to-end Transformer-based framework which jointly and interactively performs online mapping, object detection and motion prediction. PIP leverages map queries, agent queries and mode queries to encode the instance-wise information of map elements, agents and motion intentions, respectively. Based on the unified query representation, a differentiable multi-task interaction scheme is proposed to exploit the correlation between perception and prediction. Even without human-annotated HD map or agent's historical tracking trajectory as guidance information, PIP realizes end-to-end multi-agent motion prediction and achieves better performance than tracking-based and HD-map-based methods. PIP provides comprehensive high-level information of the driving scene (vectorized static map and dynamic objects with motion information), and contributes to the downstream planning and control. Code and models will be released for facilitating further research.

Lianghui Zhu, Yuxin Fang, Bencheng Liao et al.

Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .

Hongyuan Tao, Bencheng Liao, Shaoyu Chen et al.

Window attention and linear attention represent two principal strategies for mitigating the quadratic complexity and ever-growing KV cache in Vision-Language Models (VLMs). However, we observe that window-based VLMs suffer performance degradation when sequence length exceeds the window size, while linear attention underperforms on information-intensive tasks such as OCR and document understanding. To overcome these limitations, we propose InfiniteVL, a linear-complexity VLM architecture that synergizes sliding window attention (SWA) with Gated DeltaNet. For achieving competitive multimodal performance under constrained resources, we design a three-stage training strategy comprising distillation pretraining, instruction tuning, and long-sequence SFT. Remarkably, using less than 2\% of the training data required by leading VLMs, InfiniteVL not only substantially outperforms previous linear-complexity VLMs but also matches the performance of leading Transformer-based VLMs, while demonstrating effective long-term memory retention. Compared to similar-sized Transformer-based VLMs accelerated by FlashAttention-2, InfiniteVL achieves over 3.6\times inference speedup while maintaining constant latency and memory footprint. In streaming video understanding scenarios, it sustains a stable 24 FPS real-time prefill speed while preserving long-term memory cache. Code and models are available at https://github.com/hustvl/InfiniteVL.

Jialv Zou, Shaoyu Chen, Bencheng Liao et al.

Generative diffusion models for end-to-end autonomous driving often suffer from mode collapse, tending to generate conservative and homogeneous behaviors. While DiffusionDrive employs predefined anchors representing different driving intentions to partition the action space and generate diverse trajectories, its reliance on imitation learning lacks sufficient constraints, resulting in a dilemma between diversity and consistent high quality. In this work, we propose DiffusionDriveV2, which leverages reinforcement learning to both constrain low-quality modes and explore for superior trajectories. This significantly enhances the overall output quality while preserving the inherent multimodality of its core Gaussian Mixture Model. First, we use scale-adaptive multiplicative noise, ideal for trajectory planning, to promote broad exploration. Second, we employ intra-anchor GRPO to manage advantage estimation among samples generated from a single anchor, and inter-anchor truncated GRPO to incorporate a global perspective across different anchors, preventing improper advantage comparisons between distinct intentions (e.g., turning vs. going straight), which can lead to further mode collapse. DiffusionDriveV2 achieves 91.2 PDMS on the NAVSIM v1 dataset and 85.5 EPDMS on the NAVSIM v2 dataset in closed-loop evaluation with an aligned ResNet-34 backbone, setting a new record. Further experiments validate that our approach resolves the dilemma between diversity and consistent high quality for truncated diffusion models, achieving the best trade-off. Code and model will be available at https://github.com/hustvl/DiffusionDriveV2

Lianghui Zhu, Bencheng Liao, Qian Zhang et al.

Recently the state space models (SSMs) with efficient hardware-aware designs, i.e., the Mamba deep learning model, have shown great potential for long sequence modeling. Meanwhile building efficient and generic vision backbones purely upon SSMs is an appealing direction. However, representing visual data is challenging for SSMs due to the position-sensitivity of visual data and the requirement of global context for visual understanding. In this paper, we show that the reliance on self-attention for visual representation learning is not necessary and propose a new generic vision backbone with bidirectional Mamba blocks (Vim), which marks the image sequences with position embeddings and compresses the visual representation with bidirectional state space models. On ImageNet classification, COCO object detection, and ADE20k semantic segmentation tasks, Vim achieves higher performance compared to well-established vision transformers like DeiT, while also demonstrating significantly improved computation & memory efficiency. For example, Vim is 2.8$\times$ faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on images with a resolution of 1248$\times$1248. The results demonstrate that Vim is capable of overcoming the computation & memory constraints on performing Transformer-style understanding for high-resolution images and it has great potential to be the next-generation backbone for vision foundation models. Code is available at https://github.com/hustvl/Vim.

Bencheng Liao, Shaoyu Chen, Haoran Yin et al.

Recently, the diffusion model has emerged as a powerful generative technique for robotic policy learning, capable of modeling multi-mode action distributions. Leveraging its capability for end-to-end autonomous driving is a promising direction. However, the numerous denoising steps in the robotic diffusion policy and the more dynamic, open-world nature of traffic scenes pose substantial challenges for generating diverse driving actions at a real-time speed. To address these challenges, we propose a novel truncated diffusion policy that incorporates prior multi-mode anchors and truncates the diffusion schedule, enabling the model to learn denoising from anchored Gaussian distribution to the multi-mode driving action distribution. Additionally, we design an efficient cascade diffusion decoder for enhanced interaction with conditional scene context. The proposed model, DiffusionDrive, demonstrates 10$\times$ reduction in denoising steps compared to vanilla diffusion policy, delivering superior diversity and quality in just 2 steps. On the planning-oriented NAVSIM dataset, with the aligned ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS without bells and whistles, setting a new record, while running at a real-time speed of 45 FPS on an NVIDIA 4090. Qualitative results on challenging scenarios further confirm that DiffusionDrive can robustly generate diverse plausible driving actions. Code and model will be available at https://github.com/hustvl/DiffusionDrive.

Bo Jiang, Shaoyu Chen, Bencheng Liao et al.

End-to-end autonomous driving demonstrates strong planning capabilities with large-scale data but still struggles in complex, rare scenarios due to limited commonsense. In contrast, Large Vision-Language Models (LVLMs) excel in scene understanding and reasoning. The path forward lies in merging the strengths of both approaches. Previous methods using LVLMs to predict trajectories or control signals yield suboptimal results, as LVLMs are not well-suited for precise numerical predictions. This paper presents Senna, an autonomous driving system combining an LVLM (Senna-VLM) with an end-to-end model (Senna-E2E). Senna decouples high-level planning from low-level trajectory prediction. Senna-VLM generates planning decisions in natural language, while Senna-E2E predicts precise trajectories. Senna-VLM utilizes a multi-image encoding approach and multi-view prompts for efficient scene understanding. Besides, we introduce planning-oriented QAs alongside a three-stage training strategy, which enhances Senna-VLM's planning performance while preserving commonsense. Extensive experiments on two datasets show that Senna achieves state-of-the-art planning performance. Notably, with pre-training on a large-scale dataset DriveX and fine-tuning on nuScenes, Senna significantly reduces average planning error by 27.12% and collision rate by 33.33% over model without pre-training. We believe Senna's cross-scenario generalization and transferability are essential for achieving fully autonomous driving. Code and models will be released at https://github.com/hustvl/Senna.

Hao Gao, Shaoyu Chen, Bo Jiang et al.

Existing end-to-end autonomous driving (AD) algorithms typically follow the Imitation Learning (IL) paradigm, which faces challenges such as causal confusion and an open-loop gap. In this work, we propose RAD, a 3DGS-based closed-loop Reinforcement Learning (RL) framework for end-to-end Autonomous Driving. By leveraging 3DGS techniques, we construct a photorealistic digital replica of the real physical world, enabling the AD policy to extensively explore the state space and learn to handle out-of-distribution scenarios through large-scale trial and error. To enhance safety, we design specialized rewards to guide the policy in effectively responding to safety-critical events and understanding real-world causal relationships. To better align with human driving behavior, we incorporate IL into RL training as a regularization term. We introduce a closed-loop evaluation benchmark consisting of diverse, previously unseen 3DGS environments. Compared to IL-based methods, RAD achieves stronger performance in most closed-loop metrics, particularly exhibiting a 3x lower collision rate. Abundant closed-loop results are presented in the supplementary material. Code is available at https://github.com/hustvl/RAD for facilitating future research.

Bencheng Liao, Hongyuan Tao, Qian Zhang et al.

Recent Multimodal Large Language Models (MLLMs) have achieved remarkable performance but face deployment challenges due to their quadratic computational complexity, growing Key-Value cache requirements, and reliance on separate vision encoders. We propose mmMamba, a framework for developing linear-complexity native multimodal state space models through progressive distillation from existing MLLMs using moderate academic computational resources. Our approach enables the direct conversion of trained decoder-only MLLMs to linear-complexity architectures without requiring pre-trained RNN-based LLM or vision encoders. We propose an seeding strategy to carve Mamba from trained Transformer and a three-stage distillation recipe, which can effectively transfer the knowledge from Transformer to Mamba while preserving multimodal capabilities. Our method also supports flexible hybrid architectures that combine Transformer and Mamba layers for customizable efficiency-performance trade-offs. Distilled from the Transformer-based decoder-only HoVLE, mmMamba-linear achieves competitive performance against existing linear and quadratic-complexity VLMs, while mmMamba-hybrid further improves performance significantly, approaching HoVLE's capabilities. At 103K tokens, mmMamba-linear demonstrates 20.6$\times$ speedup and 75.8% GPU memory reduction compared to HoVLE, while mmMamba-hybrid achieves 13.5$\times$ speedup and 60.2% memory savings. Code and models are released at https://github.com/hustvl/mmMamba

Yongkang Li, Lijun Zhou, Sixu Yan et al.

Vision-Language-Action (VLA) models have recently emerged in autonomous driving, with the promise of leveraging rich world knowledge to improve the cognitive capabilities of driving systems. However, adapting such models for driving tasks currently faces a critical dilemma between spatial perception and semantic reasoning. Consequently, existing VLA systems are forced into suboptimal compromises: directly adopting 2D Vision-Language Models yields limited spatial perception, whereas enhancing them with 3D spatial representations often impairs the native reasoning capacity of VLMs. We argue that this dilemma largely stems from the coupled optimization of spatial perception and semantic reasoning within shared model parameters. To overcome this, we propose UniDriveVLA, a Unified Driving Vision-Language-Action model based on Mixture-of-Transformers that addresses the perception-reasoning conflict via expert decoupling. Specifically, it comprises three experts for driving understanding, scene perception, and action planning, which are coordinated through masked joint attention. In addition, we combine a sparse perception paradigm with a three-stage progressive training strategy to improve spatial perception while maintaining semantic reasoning capability. Extensive experiments show that UniDriveVLA achieves state-of-the-art performance in open-loop evaluation on nuScenes and closed-loop evaluation on Bench2Drive. Moreover, it demonstrates strong performance across a broad range of perception, prediction, and understanding tasks, including 3D detection, online mapping, motion forecasting, and driving-oriented VQA, highlighting its broad applicability as a unified model for autonomous driving. Code and model have been released at https://github.com/xiaomi-research/unidrivevla

Jialv Zou, Bencheng Liao, Qian Zhang et al.

Learning robust and scalable visual representations from massive multi-view video data remains a challenge in computer vision and autonomous driving. Existing pre-training methods either rely on expensive supervised learning with 3D annotations, limiting the scalability, or focus on single-frame or monocular inputs, neglecting the temporal information. We propose MIM4D, a novel pre-training paradigm based on dual masked image modeling (MIM). MIM4D leverages both spatial and temporal relations by training on masked multi-view video inputs. It constructs pseudo-3D features using continuous scene flow and projects them onto 2D plane for supervision. To address the lack of dense 3D supervision, MIM4D reconstruct pixels by employing 3D volumetric differentiable rendering to learn geometric representations. We demonstrate that MIM4D achieves state-of-the-art performance on the nuScenes dataset for visual representation learning in autonomous driving. It significantly improves existing methods on multiple downstream tasks, including BEV segmentation (8.7% IoU), 3D object detection (3.5% mAP), and HD map construction (1.4% mAP). Our work offers a new choice for learning representation at scale in autonomous driving. Code and models are released at https://github.com/hustvl/MIM4D

Jialv Zou, Bencheng Liao, Qian Zhang et al.

Recent advancements in unified multimodal understanding and visual generation (or multimodal generation) models have been hindered by their quadratic computational complexity and dependence on large-scale training data. We present OmniMamba, the first linear-architecture-based multimodal generation model that generates both text and images through a unified next-token prediction paradigm. The model fully leverages Mamba-2's high computational and memory efficiency, extending its capabilities from text generation to multimodal generation. To address the data inefficiency of existing unified models, we propose two key innovations: (1) decoupled vocabularies to guide modality-specific generation, and (2) task-specific LoRA for parameter-efficient adaptation. Furthermore, we introduce a decoupled two-stage training strategy to mitigate data imbalance between two tasks. Equipped with these techniques, OmniMamba achieves competitive performance with JanusFlow while surpassing Show-o across benchmarks, despite being trained on merely 2M image-text pairs, which is 1,000 times fewer than Show-o. Notably, OmniMamba stands out with outstanding inference efficiency, achieving up to a 119.2 times speedup and 63% GPU memory reduction for long-sequence generation compared to Transformer-based counterparts. Code and models are released at https://github.com/hustvl/OmniMamba

Yingyue Li, Bencheng Liao, Wenyu Liu et al.

With the advancement of RNN models with linear complexity, the quadratic complexity challenge of transformers has the potential to be overcome. Notably, the emerging Mamba-2 has demonstrated competitive performance, bridging the gap between RNN models and transformers. However, due to sequential processing and vanishing gradients, RNN models struggle to capture long-range dependencies, limiting contextual understanding. This results in slow convergence, high resource demands, and poor performance on downstream understanding and complex reasoning tasks. In this work, we present a hybrid model MaTVLM by substituting a portion of the transformer decoder layers in a pre-trained VLM with Mamba-2 layers. Leveraging the inherent relationship between attention and Mamba-2, we initialize Mamba-2 with corresponding attention weights to accelerate convergence. Subsequently, we employ a single-stage distillation process, using the pre-trained VLM as the teacher model to transfer knowledge to the MaTVLM, further enhancing convergence speed and performance. Furthermore, we investigate the impact of differential distillation loss within our training framework. We evaluate the MaTVLM on multiple benchmarks, demonstrating competitive performance against the teacher model and existing VLMs while surpassing both Mamba-based VLMs and models of comparable parameter scales. Remarkably, the MaTVLM achieves up to 3.6x faster inference than the teacher model while reducing GPU memory consumption by 27.5%, all without compromising performance. Code and models are released at http://github.com/hustvl/MaTVLM.

Shaoyu Chen, Bo Jiang, Hao Gao et al.

Learning a human-like driving policy from large-scale driving demonstrations is promising, but the uncertainty and non-deterministic nature of planning make it challenging. In this work, to cope with the uncertainty problem, we propose VADv2, an end-to-end driving model based on probabilistic planning. VADv2 takes multi-view image sequences as input in a streaming manner, transforms sensor data into environmental token embeddings, outputs the probabilistic distribution of action, and samples one action to control the vehicle. Only with camera sensors, VADv2 achieves state-of-the-art closed-loop performance on the CARLA Town05 benchmark, significantly outperforming all existing methods. It runs stably in a fully end-to-end manner, even without the rule-based wrapper. Closed-loop demos are presented at https://hgao-cv.github.io/VADv2.

Lunbin Zeng, Jingfeng Yao, Bencheng Liao et al.

In recent multimodal research, the diffusion paradigm has emerged as a promising alternative to the autoregressive paradigm (AR), owing to its unique decoding advantages. However, due to the capability limitations of the base diffusion language model, the performance of the diffusion vision language model (dVLM) still lags significantly behind that of mainstream models. This leads to a simple yet fundamental question: Is it possible to construct dVLMs based on existing powerful AR models? In response, we propose DiffusionVL, a dVLM family that could be translated from any powerful AR models. Through simple fine-tuning, we successfully adapt AR pre-trained models into the diffusion paradigm. This approach yields two key observations: (1) The paradigm shift from AR-based multimodal models to diffusion is remarkably effective. (2) Direct conversion of an AR language model to a dVLM is also feasible, achieving performance competitive with LLaVA-style visual-instruction-tuning. Further, we introduce a block-decoding design into dVLMs that supports arbitrary-length generation and KV cache reuse, achieving a significant inference speedup. We conduct a large number of experiments. Despite training with less than 5% of the data required by prior methods, DiffusionVL achieves a comprehensive performance improvement-a 34.4% gain on the MMMU-Pro (vision) bench and 37.5% gain on the MME (Cog.) bench-alongside a 2x inference speedup. The model and code are released at https://github.com/hustvl/DiffusionVL.

Dong Guo, Faming Wu, Feida Zhu et al. · pku

We present Seed1.5-VL, a vision-language foundation model designed to advance general-purpose multimodal understanding and reasoning. Seed1.5-VL is composed with a 532M-parameter vision encoder and a Mixture-of-Experts (MoE) LLM of 20B active parameters. Despite its relatively compact architecture, it delivers strong performance across a wide spectrum of public VLM benchmarks and internal evaluation suites, achieving the state-of-the-art performance on 38 out of 60 public benchmarks. Moreover, in agent-centric tasks such as GUI control and gameplay, Seed1.5-VL outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7. Beyond visual and video understanding, it also demonstrates strong reasoning abilities, making it particularly effective for multimodal reasoning challenges such as visual puzzles. We believe these capabilities will empower broader applications across diverse tasks. In this report, we mainly provide a comprehensive review of our experiences in building Seed1.5-VL across model design, data construction, and training at various stages, hoping that this report can inspire further research. Seed1.5-VL is now accessible at https://www.volcengine.com/ (Volcano Engine Model ID: doubao-1-5-thinking-vision-pro-250428)

Yuxin Fang, Bencheng Liao, Xinggang Wang et al.

Can Transformer perform 2D object- and region-level recognition from a pure sequence-to-sequence perspective with minimal knowledge about the 2D spatial structure? To answer this question, we present You Only Look at One Sequence (YOLOS), a series of object detection models based on the vanilla Vision Transformer with the fewest possible modifications, region priors, as well as inductive biases of the target task. We find that YOLOS pre-trained on the mid-sized ImageNet-1k dataset only can already achieve quite competitive performance on the challenging COCO object detection benchmark, e.g., YOLOS-Base directly adopted from BERT-Base architecture can obtain 42.0 box AP on COCO val. We also discuss the impacts as well as limitations of current pre-train schemes and model scaling strategies for Transformer in vision through YOLOS. Code and pre-trained models are available at https://github.com/hustvl/YOLOS.

Xinggang Wang, Zhaojin Huang, Bencheng Liao et al.

Video object detection is a fundamental problem in computer vision and has a wide spectrum of applications. Based on deep networks, video object detection is actively studied for pushing the limits of detection speed and accuracy. To reduce the computation cost, we sparsely sample key frames in video and treat the rest frames are non-key frames; a large and deep network is used to extract features for key frames and a tiny network is used for non-key frames. To enhance the features of non-key frames, we propose a novel short-term feature aggregation method to propagate the rich information in key frame features to non-key frame features in a fast way. The fast feature aggregation is enabled by the freely available motion cues in compressed videos. Further, key frame features are also aggregated based on optical flow. The propagated deep features are then integrated with the directly extracted features for object detection. The feature extraction and feature integration parameters are optimized in an end-to-end manner. The proposed video object detection network is evaluated on the large-scale ImageNet VID benchmark and achieves 77.2\% mAP, which is on-par with state-of-the-art accuracy, at the speed of 30 FPS using a Titan X GPU. The source codes are available at \url{https://github.com/hustvl/LSFA}.

Ziying Song, Lin Liu, Hongyu Pan et al.

Most end-to-end autonomous driving methods rely on imitation learning from single expert demonstrations, often leading to conservative and homogeneous behaviors that limit generalization in complex real-world scenarios. In this work, we propose DIVER, an end-to-end driving framework that integrates reinforcement learning with diffusion-based generation to produce diverse and feasible trajectories. At the core of DIVER lies a reinforced diffusion-based generation mechanism. First, the model conditions on map elements and surrounding agents to generate multiple reference trajectories from a single ground-truth trajectory, alleviating the limitations of imitation learning that arise from relying solely on single expert demonstrations. Second, reinforcement learning is employed to guide the diffusion process, where reward-based supervision enforces safety and diversity constraints on the generated trajectories, thereby enhancing their practicality and generalization capability. Furthermore, to address the limitations of L2-based open-loop metrics in capturing trajectory diversity, we propose a novel Diversity metric to evaluate the diversity of multi-mode predictions.Extensive experiments on the closed-loop NAVSIM and Bench2Drive benchmarks, as well as the open-loop nuScenes dataset, demonstrate that DIVER significantly improves trajectory diversity, effectively addressing the mode collapse problem inherent in imitation learning.

Lin Liu, Caiyan Jia, Ziying Song et al.

Current end-to-end autonomous driving methods typically learn only from expert planning data collected from a single ego vehicle, severely limiting the diversity of learnable driving policies and scenarios. However, a critical yet overlooked fact is that in any driving scenario, multiple high-quality trajectories from other vehicles coexist with a specific ego vehicle's trajectory. Existing methods fail to fully exploit this valuable resource, missing important opportunities to improve the models' performance (including long-tail scenarios) through learning from other experts. Intuitively, Jointly learning from both ego and other vehicles' expert data is beneficial for planning tasks. However, this joint learning faces two critical challenges. (1) Different scene observation perspectives across vehicles hinder inter-vehicle alignment of scene feature representations; (2) The absence of partial modality in other vehicles' data (e.g., vehicle states) compared to ego-vehicle data introduces learning bias. To address these challenges, we propose FUMP (Fully Unified Motion Planning), a novel two-stage trajectory generation framework. Building upon probabilistic decomposition, we model the planning task as a specialized subtask of motion prediction. Specifically, our approach decouples trajectory planning into two stages. In Stage 1, a shared decoder jointly generates initial trajectories for both tasks. In Stage 2, the model performs planning-specific refinement conditioned on an ego-vehicle's state. The transition between the two stages is bridged by a state predictor trained exclusively on ego-vehicle data. To address the cross-vehicle discrepancy in observational perspectives, we propose an Equivariant Context-Sharing Adapter (ECSA) before Stage 1 for improving cross-vehicle generalization of scene representations.