Shiyi Lan

Niket Agarwal, Arslan Ali, Maciej Bala et al. · nvidia

Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.

Hassan Abu Alhaija, Jose Alvarez, Maciej Bala et al. · nvidia

We introduce Cosmos-Transfer, a conditional world generation model that can generate world simulations based on multiple spatial control inputs of various modalities such as segmentation, depth, and edge. In the design, the spatial conditional scheme is adaptive and customizable. It allows weighting different conditional inputs differently at different spatial locations. This enables highly controllable world generation and finds use in various world-to-world transfer use cases, including Sim2Real. We conduct extensive evaluations to analyze the proposed model and demonstrate its applications for Physical AI, including robotics Sim2Real and autonomous vehicle data enrichment. We further demonstrate an inference scaling strategy to achieve real-time world generation with an NVIDIA GB200 NVL72 rack. To help accelerate research development in the field, we open-source our models and code at https://github.com/nvidia-cosmos/cosmos-transfer1.

Yilun Chen, Zhiding Yu, Yukang Chen et al.

False negatives (FN) in 3D object detection, {\em e.g.}, missing predictions of pedestrians, vehicles, or other obstacles, can lead to potentially dangerous situations in autonomous driving. While being fatal, this issue is understudied in many current 3D detection methods. In this work, we propose Hard Instance Probing (HIP), a general pipeline that identifies \textit{FN} in a multi-stage manner and guides the models to focus on excavating difficult instances. For 3D object detection, we instantiate this method as FocalFormer3D, a simple yet effective detector that excels at excavating difficult objects and improving prediction recall. FocalFormer3D features a multi-stage query generation to discover hard objects and a box-level transformer decoder to efficiently distinguish objects from massive object candidates. Experimental results on the nuScenes and Waymo datasets validate the superior performance of FocalFormer3D. The advantage leads to strong performance on both detection and tracking, in both LiDAR and multi-modal settings. Notably, FocalFormer3D achieves a 70.5 mAP and 73.9 NDS on nuScenes detection benchmark, while the nuScenes tracking benchmark shows 72.1 AMOTA, both ranking 1st place on the nuScenes LiDAR leaderboard. Our code is available at \url{https://github.com/NVlabs/FocalFormer3D}.

Zhiqi Li, Zhiding Yu, David Austin et al.

This technical report summarizes the winning solution for the 3D Occupancy Prediction Challenge, which is held in conjunction with the CVPR 2023 Workshop on End-to-End Autonomous Driving and CVPR 23 Workshop on Vision-Centric Autonomous Driving Workshop. Our proposed solution FB-OCC builds upon FB-BEV, a cutting-edge camera-based bird's-eye view perception design using forward-backward projection. On top of FB-BEV, we further study novel designs and optimization tailored to the 3D occupancy prediction task, including joint depth-semantic pre-training, joint voxel-BEV representation, model scaling up, and effective post-processing strategies. These designs and optimization result in a state-of-the-art mIoU score of 54.19% on the nuScenes dataset, ranking the 1st place in the challenge track. Code and models will be released at: https://github.com/NVlabs/FB-BEV.

Zhuolin Yang, Wei Ping, Zihan Liu et al. · stanford

Augmenting pretrained language models (LMs) with a vision encoder (e.g., Flamingo) has obtained the state-of-the-art results in image-to-text generation. However, these models store all the knowledge within their parameters, thus often requiring enormous model parameters to model the abundant visual concepts and very rich textual descriptions. Additionally, they are inefficient in incorporating new data, requiring a computational-expensive fine-tuning process. In this work, we introduce a Retrieval-augmented Visual Language Model, Re-ViLM, built upon the Flamingo, that supports retrieving the relevant knowledge from the external database for zero and in-context few-shot image-to-text generations. By storing certain knowledge explicitly in the external database, our approach reduces the number of model parameters and can easily accommodate new data during evaluation by simply updating the database. We also construct an interleaved image and text data that facilitates in-context few-shot learning capabilities. We demonstrate that Re-ViLM significantly boosts performance for image-to-text generation tasks, especially for zero-shot and few-shot generation in out-of-domain settings with 4 times less parameters compared with baseline methods.

Junfei Xiao, Zhichao Xu, Shiyi Lan et al.

This report describes the winning solution to the Robust Vision Challenge (RVC) semantic segmentation track at ECCV 2022. Our method adopts the FAN-B-Hybrid model as the encoder and uses SegFormer as the segmentation framework. The model is trained on a composite dataset consisting of images from 9 datasets (ADE20K, Cityscapes, Mapillary Vistas, ScanNet, VIPER, WildDash 2, IDD, BDD, and COCO) with a simple dataset balancing strategy. All the original labels are projected to a 256-class unified label space, and the model is trained using a cross-entropy loss. Without significant hyperparameter tuning or any specific loss weighting, our solution ranks the first place on all the testing semantic segmentation benchmarks from multiple domains (ADE20K, Cityscapes, Mapillary Vistas, ScanNet, VIPER, and WildDash 2). The proposed method can serve as a strong baseline for the multi-domain segmentation task and benefit future works. Code will be available at https://github.com/lambert-x/RVC_Segmentation.

Lingchen Meng, Shiyi Lan, Hengduo Li et al.

In-context segmentation aims at segmenting novel images using a few labeled example images, termed as "in-context examples", exploring content similarities between examples and the target. The resulting models can be generalized seamlessly to novel segmentation tasks, significantly reducing the labeling and training costs compared with conventional pipelines. However, in-context segmentation is more challenging than classic ones requiring the model to learn segmentation rules conditioned on a few samples. Unlike previous work with ad-hoc or non-end-to-end designs, we propose SEGIC, an end-to-end segment-in-context framework built upon a single vision foundation model (VFM). In particular, SEGIC leverages the emergent correspondence within VFM to capture dense relationships between target images and in-context samples. As such, information from in-context samples is then extracted into three types of instructions, i.e. geometric, visual, and meta instructions, serving as explicit conditions for the final mask prediction. SEGIC is a straightforward yet effective approach that yields state-of-the-art performance on one-shot segmentation benchmarks. Notably, SEGIC can be easily generalized to diverse tasks, including video object segmentation and open-vocabulary segmentation. Code will be available at https://github.com/MengLcool/SEGIC.

Wujian Peng, Sicheng Xie, Zuyao You et al.

Vision language models (VLM) have demonstrated remarkable performance across various downstream tasks. However, understanding fine-grained visual-linguistic concepts, such as attributes and inter-object relationships, remains a significant challenge. While several benchmarks aim to evaluate VLMs in finer granularity, their primary focus remains on the linguistic aspect, neglecting the visual dimension. Here, we highlight the importance of evaluating VLMs from both a textual and visual perspective. We introduce a progressive pipeline to synthesize images that vary in a specific attribute while ensuring consistency in all other aspects. Utilizing this data engine, we carefully design a benchmark, SPEC, to diagnose the comprehension of object size, position, existence, and count. Subsequently, we conduct a thorough evaluation of four leading VLMs on SPEC. Surprisingly, their performance is close to random guess, revealing significant limitations. With this in mind, we propose a simple yet effective approach to optimize VLMs in fine-grained understanding, achieving significant improvements on SPEC without compromising the zero-shot performance. Results on two additional fine-grained benchmarks also show consistent improvements, further validating the transferability of our approach. Code and data are available at https://github.com/wjpoom/SPEC.

Ali Hatamizadeh, Michael Ranzinger, Shiyi Lan et al.

Vision Transformers (ViTs) have attracted a lot of popularity in recent years, due to their exceptional capabilities in modeling long-range spatial dependencies and scalability for large scale training. Although the training parallelism of self-attention mechanism plays an important role in retaining great performance, its quadratic complexity baffles the application of ViTs in many scenarios which demand fast inference. This effect is even more pronounced in applications in which autoregressive modeling of input features is required. In Natural Language Processing (NLP), a new stream of efforts has proposed parallelizable models with recurrent formulation that allows for efficient inference in generative applications. Inspired by this trend, we propose a new class of computer vision models, dubbed Vision Retention Networks (ViR), with dual parallel and recurrent formulations, which strike an optimal balance between fast inference and parallel training with competitive performance. In particular, ViR scales favorably for image throughput and memory consumption in tasks that require higher-resolution images due to its flexible formulation in processing large sequence lengths. The ViR is the first attempt to realize dual parallel and recurrent equivalency in a general vision backbone for recognition tasks. We have validated the effectiveness of ViR through extensive experiments with different dataset sizes and various image resolutions and achieved competitive performance. Code: https://github.com/NVlabs/ViR

Shiyi Lan, Xitong Yang, Zhiding Yu et al.

We propose Mask Auto-Labeler (MAL), a high-quality Transformer-based mask auto-labeling framework for instance segmentation using only box annotations. MAL takes box-cropped images as inputs and conditionally generates their mask pseudo-labels.We show that Vision Transformers are good mask auto-labelers. Our method significantly reduces the gap between auto-labeling and human annotation regarding mask quality. Instance segmentation models trained using the MAL-generated masks can nearly match the performance of their fully-supervised counterparts, retaining up to 97.4\% performance of fully supervised models. The best model achieves 44.1\% mAP on COCO instance segmentation (test-dev 2017), outperforming state-of-the-art box-supervised methods by significant margins. Qualitative results indicate that masks produced by MAL are, in some cases, even better than human annotations.

Zhiqi Li, Zhiding Yu, Shiyi Lan et al.

End-to-end autonomous driving recently emerged as a promising research direction to target autonomy from a full-stack perspective. Along this line, many of the latest works follow an open-loop evaluation setting on nuScenes to study the planning behavior. In this paper, we delve deeper into the problem by conducting thorough analyses and demystifying more devils in the details. We initially observed that the nuScenes dataset, characterized by relatively simple driving scenarios, leads to an under-utilization of perception information in end-to-end models incorporating ego status, such as the ego vehicle's velocity. These models tend to rely predominantly on the ego vehicle's status for future path planning. Beyond the limitations of the dataset, we also note that current metrics do not comprehensively assess the planning quality, leading to potentially biased conclusions drawn from existing benchmarks. To address this issue, we introduce a new metric to evaluate whether the predicted trajectories adhere to the road. We further propose a simple baseline able to achieve competitive results without relying on perception annotations. Given the current limitations on the benchmark and metrics, we suggest the community reassess relevant prevailing research and be cautious whether the continued pursuit of state-of-the-art would yield convincing and universal conclusions. Code and models are available at \url{https://github.com/NVlabs/BEV-Planner}

Chengyue Wu, Shiyi Lan, Yonggan Fu et al.

Vision-language models (VLMs) predominantly rely on autoregressive decoding, which generates tokens one at a time and fundamentally limits inference throughput. This limitation is especially acute in physical AI scenarios such as robotics and autonomous driving, where VLMs are deployed on edge devices at batch size one, making AR decoding memory-bandwidth-bound and leaving hardware parallelism underutilized. While block-wise discrete diffusion has shown promise for parallel text generation, extending it to VLMs remains challenging due to the need to jointly handle continuous visual representations and discrete text tokens while preserving pretrained multimodal capabilities. We present Fast-dVLM, a block-diffusion-based VLM that enables KV-cache-compatible parallel decoding and speculative block decoding for inference acceleration. We systematically compare two AR-to-diffusion conversion strategies: a two-stage approach that first adapts the LLM backbone with text-only diffusion fine-tuning before multimodal training, and a direct approach that converts the full AR VLM in one stage. Under comparable training budgets, direct conversion proves substantially more efficient by leveraging the already multimodally aligned VLM; we therefore adopt it as our recommended recipe. We introduce a suite of multimodal diffusion adaptations, block size annealing, causal context attention, auto-truncation masking, and vision efficient concatenation, that collectively enable effective block diffusion in the VLM setting. Extensive experiments across 11 multimodal benchmarks show Fast-dVLM matches its autoregressive counterpart in generation quality. With SGLang integration and FP8 quantization, Fast-dVLM achieves over 6x end-to-end inference speedup over the AR baseline.

Zhenxin Li, Shiyi Lan, Jose M. Alvarez et al.

Recently, the rise of query-based Transformer decoders is reshaping camera-based 3D object detection. These query-based decoders are surpassing the traditional dense BEV (Bird's Eye View)-based methods. However, we argue that dense BEV frameworks remain important due to their outstanding abilities in depth estimation and object localization, depicting 3D scenes accurately and comprehensively. This paper aims to address the drawbacks of the existing dense BEV-based 3D object detectors by introducing our proposed enhanced components, including a CRF-modulated depth estimation module enforcing object-level consistencies, a long-term temporal aggregation module with extended receptive fields, and a two-stage object decoder combining perspective techniques with CRF-modulated depth embedding. These enhancements lead to a "modernized" dense BEV framework dubbed BEVNeXt. On the nuScenes benchmark, BEVNeXt outperforms both BEV-based and query-based frameworks under various settings, achieving a state-of-the-art result of 64.2 NDS on the nuScenes test set. Code will be available at \url{https://github.com/woxihuanjiangguo/BEVNeXt}.

Zhiqi Li, Guo Chen, Shilong Liu et al.

Recently, promising progress has been made by open-source vision-language models (VLMs) in bringing their capabilities closer to those of proprietary frontier models. However, most open-source models only publish their final model weights, leaving the critical details of data strategies and implementation largely opaque. In this work, we address VLM post-training from a data-centric perspective, showing the key role of data strategy in developing frontier VLMs. By studying and building our post-training data strategy from scratch, we share detailed insights into the development processes, aiming to benefit the development of competitive models for the open-source community. Our introduced data strategy, together with training recipes and model design, leads to a family of performant VLMs named Eagle2. Specifically, Eagle2-9B achieves state-of-the-art results across various multimodal benchmarks, matching certain competitive models with up to 70B parameters.

Barath Lakshmanan, Joshua Chen, Shiyi Lan et al.

The cornerstone of autonomous vehicles (AV) is a solid perception system, where camera encoders play a crucial role. Existing works usually leverage pre-trained Convolutional Neural Networks (CNN) or Vision Transformers (ViTs) designed for general vision tasks, such as image classification, segmentation, and 2D detection. Although those well-known architectures have achieved state-of-the-art accuracy in AV-related tasks, e.g., 3D Object Detection, there remains significant potential for improvement in network design due to the nuanced complexities of industrial-level AV dataset. Moreover, existing public AV benchmarks usually contain insufficient data, which might lead to inaccurate evaluation of those architectures.To reveal the AV-specific model insights, we start from a standard general-purpose encoder, ConvNeXt and progressively transform the design. We adjust different design parameters including width and depth of the model, stage compute ratio, attention mechanisms, and input resolution, supported by systematic analysis to each modifications. This customization yields an architecture optimized for AV camera encoder achieving 8.79% mAP improvement over the baseline. We believe our effort could become a sweet cookbook of image encoders for AV and pave the way to the next-level drive system.

Wenhao Yao, Xinglong Sun, Zhenxin Li et al.

End-to-end planning has emerged as a dominant paradigm for autonomous driving, where recent models often adopt a scoring-selection framework to choose trajectories from a large set of candidates, with diffusion-based decoding showing strong promise. However, directly selecting from the entire candidate space remains difficult to optimize, and Gaussian perturbations used in diffusion often introduce unrealistic trajectories that complicate the denoising process. In addition, for training these models, reinforcement learning (RL) has shown promise, but existing end-to-end RL approaches typically rely on a single coupled reward without structured signals, limiting optimization effectiveness. To address these challenges, we propose HAD, an end-to-end planning framework with a Hierarchical Diffusion Policy that decomposes planning into a coarse-to-fine process. To improve trajectory generation, we introduce Structure-Preserved Trajectory Expansion, which produces realistic candidates while maintaining kinematic structure. For policy learning, we develop Metric-Decoupled Policy Optimization (MDPO) to enable structured RL optimization across multiple driving objectives. Extensive experiments show that HAD achieves new state-of-the-art performance on both NAVSIM and HUGSIM, outperforming prior arts by a huge margin: +2.3 EPDMS on NAVSIM and +4.9 Route Completion on HUGSIM.

Zhenxin Li, Shihao Wang, Shiyi Lan et al.

End-to-end autonomous driving research currently faces a critical challenge in bridging the gap between open-loop training and closed-loop deployment. Current approaches are trained to predict trajectories in an open-loop environment, which struggle with quick reactions to other agents in closed-loop environments and risk generating kinematically infeasible plans due to the gap between open-loop training and closed-loop driving. In this paper, we introduce Hydra-NeXt, a novel multi-branch planning framework that unifies trajectory prediction, control prediction, and a trajectory refinement network in one model. Unlike current open-loop trajectory prediction models that only handle general-case planning, Hydra-NeXt further utilizes a control decoder to focus on short-term actions, which enables faster responses to dynamic situations and reactive agents. Moreover, we propose the Trajectory Refinement module to augment and refine the planning decisions by effectively adhering to kinematic constraints in closed-loop environments. This unified approach bridges the gap between open-loop training and closed-loop driving, demonstrating superior performance of 65.89 Driving Score (DS) and 48.20% Success Rate (SR) on the Bench2Drive dataset without relying on external experts for data collection. Hydra-NeXt surpasses the previous state-of-the-art by 22.98 DS and 17.49 SR, marking a significant advancement in autonomous driving. Code will be available at https://github.com/woxihuanjiangguo/Hydra-NeXt.

Bingyin Zhao, Zhiding Yu, Shiyi Lan et al.

Recent studies indicate that Vision Transformers (ViTs) are robust against out-of-distribution scenarios. In particular, the Fully Attentional Network (FAN) - a family of ViT backbones, has achieved state-of-the-art robustness. In this paper, we revisit the FAN models and improve their pre-training with a self-emerging token labeling (STL) framework. Our method contains a two-stage training framework. Specifically, we first train a FAN token labeler (FAN-TL) to generate semantically meaningful patch token labels, followed by a FAN student model training stage that uses both the token labels and the original class label. With the proposed STL framework, our best model based on FAN-L-Hybrid (77.3M parameters) achieves 84.8% Top-1 accuracy and 42.1% mCE on ImageNet-1K and ImageNet-C, and sets a new state-of-the-art for ImageNet-A (46.1%) and ImageNet-R (56.6%) without using extra data, outperforming the original FAN counterpart by significant margins. The proposed framework also demonstrates significantly enhanced performance on downstream tasks such as semantic segmentation, with up to 1.7% improvement in robustness over the counterpart model. Code is available at https://github.com/NVlabs/STL.

Junfei Xiao, Ziqi Zhou, Wenxuan Li et al.

This paper introduces ProLab, a novel approach using property-level label space for creating strong interpretable segmentation models. Instead of relying solely on category-specific annotations, ProLab uses descriptive properties grounded in common sense knowledge for supervising segmentation models. It is based on two core designs. First, we employ Large Language Models (LLMs) and carefully crafted prompts to generate descriptions of all involved categories that carry meaningful common sense knowledge and follow a structured format. Second, we introduce a description embedding model preserving semantic correlation across descriptions and then cluster them into a set of descriptive properties (e.g., 256) using K-Means. These properties are based on interpretable common sense knowledge consistent with theories of human recognition. We empirically show that our approach makes segmentation models perform stronger on five classic benchmarks (e.g., ADE20K, COCO-Stuff, Pascal Context, Cityscapes, and BDD). Our method also shows better scalability with extended training steps than category-level supervision. Our interpretable segmentation framework also emerges with the generalization ability to segment out-of-domain or unknown categories using only in-domain descriptive properties. Code is available at https://github.com/lambert-x/ProLab.

Zhenxin Li, Wenhao Yao, Zi Wang et al.

End-to-end autonomous driving maps raw sensor inputs directly into ego-vehicle trajectories to avoid cascading errors from perception modules and to leverage rich semantic cues. Existing frameworks largely rely on Imitation Learning (IL), which can be limited by sub-optimal expert demonstrations and covariate shift during deployment. On the other hand, Reinforcement Learning (RL) has recently shown potential in scaling up with simulations, but is typically confined to low-dimensional symbolic inputs (e.g. 3D objects and maps), falling short of full end-to-end learning from raw sensor data. We introduce ZTRS (Zero-Imitation End-to-End Autonomous Driving with Trajectory Scoring), a framework that combines the strengths of both worlds: sensor inputs without losing information and RL training for robust planning. To the best of our knowledge, ZTRS is the first framework that eliminates IL entirely by only learning from rewards while operating directly on high-dimensional sensor data. ZTRS utilizes offline reinforcement learning with our proposed Exhaustive Policy Optimization (EPO), a variant of policy gradient tailored for enumerable actions and rewards. ZTRS demonstrates strong performance across three benchmarks: Navtest (generic real-world open-loop planning), Navhard (open-loop planning in challenging real-world and synthetic scenarios), and HUGSIM (simulated closed-loop driving). Specifically, ZTRS achieves the state-of-the-art result on Navhard and outperforms IL-based baselines on HUGSIM. Code will be available at https://github.com/woxihuanjiangguo/ZTRS.

Zhenxin Li, Kailin Li, Shihao Wang et al.

We propose Hydra-MDP, a novel paradigm employing multiple teachers in a teacher-student model. This approach uses knowledge distillation from both human and rule-based teachers to train the student model, which features a multi-head decoder to learn diverse trajectory candidates tailored to various evaluation metrics. With the knowledge of rule-based teachers, Hydra-MDP learns how the environment influences the planning in an end-to-end manner instead of resorting to non-differentiable post-processing. This method achieves the $1^{st}$ place in the Navsim challenge, demonstrating significant improvements in generalization across diverse driving environments and conditions. More details by visiting \url{https://github.com/NVlabs/Hydra-MDP}.

Tianrui Guan, Jun Wang, Shiyi Lan et al.

We present a novel architecture for 3D object detection, M3DeTR, which combines different point cloud representations (raw, voxels, bird-eye view) with different feature scales based on multi-scale feature pyramids. M3DeTR is the first approach that unifies multiple point cloud representations, feature scales, as well as models mutual relationships between point clouds simultaneously using transformers. We perform extensive ablation experiments that highlight the benefits of fusing representation and scale, and modeling the relationships. Our method achieves state-of-the-art performance on the KITTI 3D object detection dataset and Waymo Open Dataset. Results show that M3DeTR improves the baseline significantly by 1.48% mAP for all classes on Waymo Open Dataset. In particular, our approach ranks 1st on the well-known KITTI 3D Detection Benchmark for both car and cyclist classes, and ranks 1st on Waymo Open Dataset with single frame point cloud input. Our code is available at: https://github.com/rayguan97/M3DETR.

Hexiang Hu, Shiyi Lan, Yuning Jiang et al.

Objects appear to scale differently in natural images. This fact requires methods dealing with object-centric tasks (e.g. object proposal) to have robust performance over variances in object scales. In the paper, we present a novel segment proposal framework, namely FastMask, which takes advantage of hierarchical features in deep convolutional neural networks to segment multi-scale objects in one shot. Innovatively, we adapt segment proposal network into three different functional components (body, neck and head). We further propose a weight-shared residual neck module as well as a scale-tolerant attentional head module for efficient one-shot inference. On MS COCO benchmark, the proposed FastMask outperforms all state-of-the-art segment proposal methods in average recall being 2~5 times faster. Moreover, with a slight trade-off in accuracy, FastMask can segment objects in near real time (~13 fps) with 800*600 resolution images, demonstrating its potential in practical applications. Our implementation is available on https://github.com/voidrank/FastMask.

Shihao Wang, Zhiding Yu, Xiaohui Jiang et al.

The advances in vision-language models (VLMs) have led to a growing interest in autonomous driving to leverage their strong reasoning capabilities. However, extending these capabilities from 2D to full 3D understanding is crucial for real-world applications. To address this challenge, we propose OmniDrive, a holistic vision-language dataset that aligns agent models with 3D driving tasks through counterfactual reasoning. This approach enhances decision-making by evaluating potential scenarios and their outcomes, similar to human drivers considering alternative actions. Our counterfactual-based synthetic data annotation process generates large-scale, high-quality datasets, providing denser supervision signals that bridge planning trajectories and language-based reasoning. Futher, we explore two advanced OmniDrive-Agent frameworks, namely Omni-L and Omni-Q, to assess the importance of vision-language alignment versus 3D perception, revealing critical insights into designing effective LLM-agents. Significant improvements on the DriveLM Q\&A benchmark and nuScenes open-loop planning demonstrate the effectiveness of our dataset and methods.

Shihao Wang, Zhiding Yu, Xiaohui Jiang et al.

The advances in vision-language models (VLMs) have led to a growing interest in autonomous driving to leverage their strong reasoning capabilities. However, extending these capabilities from 2D to full 3D understanding is crucial for real-world applications. To address this challenge, we propose OmniDrive, a holistic vision-language dataset that aligns agent models with 3D driving tasks through counterfactual reasoning. This approach enhances decision-making by evaluating potential scenarios and their outcomes, similar to human drivers considering alternative actions. Our counterfactual-based synthetic data annotation process generates large-scale, high-quality datasets, providing denser supervision signals that bridge planning trajectories and language-based reasoning. Futher, we explore two advanced OmniDrive-Agent frameworks, namely Omni-L and Omni-Q, to assess the importance of vision-language alignment versus 3D perception, revealing critical insights into designing effective LLM-agents. Significant improvements on the DriveLM Q\&A benchmark and nuScenes open-loop planning demonstrate the effectiveness of our dataset and methods.

Kailin Li, Zhenxin Li, Shiyi Lan et al.

Hydra-MDP++ introduces a novel teacher-student knowledge distillation framework with a multi-head decoder that learns from human demonstrations and rule-based experts. Using a lightweight ResNet-34 network without complex components, the framework incorporates expanded evaluation metrics, including traffic light compliance (TL), lane-keeping ability (LK), and extended comfort (EC) to address unsafe behaviors not captured by traditional NAVSIM-derived teachers. Like other end-to-end autonomous driving approaches, \hydra processes raw images directly without relying on privileged perception signals. Hydra-MDP++ achieves state-of-the-art performance by integrating these components with a 91.0% drive score on NAVSIM through scaling to a V2-99 image encoder, demonstrating its effectiveness in handling diverse driving scenarios while maintaining computational efficiency.

Chonghao Sima, Kashyap Chitta, Zhiding Yu et al.

How can we rely on an end-to-end autonomous vehicle's complex decision-making system during deployment? One common solution is to have a ``fallback layer'' that checks the planned trajectory for rule violations and replaces it with a pre-defined safe action if necessary. Another approach involves adjusting the planner's decisions to minimize a pre-defined ``cost function'' using additional system predictions such as road layouts and detected obstacles. However, these pre-programmed rules or cost functions cannot learn and improve with new training data, often resulting in overly conservative behaviors. In this work, we propose Centaur (Cluster Entropy for Test-time trAining using Uncertainty) which updates a planner's behavior via test-time training, without relying on hand-engineered rules or cost functions. Instead, we measure and minimize the uncertainty in the planner's decisions. For this, we develop a novel uncertainty measure, called Cluster Entropy, which is simple, interpretable, and compatible with state-of-the-art planning algorithms. Using data collected at prior test-time time-steps, we perform an update to the model's parameters using a gradient that minimizes the Cluster Entropy. With only this sole gradient update prior to inference, Centaur exhibits significant improvements, ranking first on the navtest leaderboard with notable gains in safety-critical metrics such as time to collision. To provide detailed insights on a per-scenario basis, we also introduce navsafe, a challenging new benchmark, which highlights previously undiscovered failure modes of driving models.

Zi Wang, Shiyi Lan, Xinglong Sun et al.

Autonomous vehicle safety is crucial for the successful deployment of self-driving cars. However, most existing planning methods rely heavily on imitation learning, which limits their ability to leverage collision data effectively. Moreover, collecting collision or near-collision data is inherently challenging, as it involves risks and raises ethical and practical concerns. In this paper, we propose SafeFusion, a training framework to learn from collision data. Instead of over-relying on imitation learning, SafeFusion integrates safety-oriented metrics during training to enable collision avoidance learning. In addition, to address the scarcity of collision data, we propose CollisionGen, a scalable data generation pipeline to generate diverse, high-quality scenarios using natural language prompts, generative models, and rule-based filtering. Experimental results show that our approach improves planning performance in collision-prone scenarios by 56\% over previous state-of-the-art planners while maintaining effectiveness in regular driving situations. Our work provides a scalable and effective solution for advancing the safety of autonomous driving systems.

Jihao Liu, Zhiding Yu, Shiyi Lan et al.

This paper presents StreamChat, a novel approach that enhances the interaction capabilities of Large Multimodal Models (LMMs) with streaming video content. In streaming interaction scenarios, existing methods rely solely on visual information available at the moment a question is posed, resulting in significant delays as the model remains unaware of subsequent changes in the streaming video. StreamChat addresses this limitation by innovatively updating the visual context at each decoding step, ensuring that the model utilizes up-to-date video content throughout the decoding process. Additionally, we introduce a flexible and efficient crossattention-based architecture to process dynamic streaming inputs while maintaining inference efficiency for streaming interactions. Furthermore, we construct a new dense instruction dataset to facilitate the training of streaming interaction models, complemented by a parallel 3D-RoPE mechanism that encodes the relative temporal information of visual and text tokens. Experimental results demonstrate that StreamChat achieves competitive performance on established image and video benchmarks and exhibits superior capabilities in streaming interaction scenarios compared to state-of-the-art video LMM.

Shuaiyi Huang, De-An Huang, Zhiding Yu et al.

Video instance segmentation (VIS) is a challenging vision task that aims to detect, segment, and track objects in videos. Conventional VIS methods rely on densely-annotated object masks which are expensive. We reduce the human annotations to only one point for each object in a video frame during training, and obtain high-quality mask predictions close to fully supervised models. Our proposed training method consists of a class-agnostic proposal generation module to provide rich negative samples and a spatio-temporal point-based matcher to match the object queries with the provided point annotations. Comprehensive experiments on three VIS benchmarks demonstrate competitive performance of the proposed framework, nearly matching fully supervised methods.

Yunfei Xie, Yinsong Ma, Shiyi Lan et al.

Developing reasoning capabilities in multimodal large language models (MLLMs) remains challenging. Motivated by literature suggesting that gameplay promotes transferable reasoning skills, we propose a novel post-training method, Visual Game Learning (ViGaL), where MLLMs develop generalizable reasoning skills through playing arcade-like games. Specifically, we show that training a 7B-parameter MLLM via reinforcement learning (RL) on simple games like Snake significantly enhances the downstream performance on multimodal math benchmarks like MathVista, on multi-discipline questions like MMMU and on 3D spatial reasoning benchmarks like VSI-Bench, without seeing any worked solutions, equations, or diagrams during RL. Remarkably, our model outperforms specialist models post-trained on benchmark-oriented multimodal reasoning data, while preserving the model's performance on general visual benchmarks, a challenge where specialist models often fall short. Our findings suggest that multimodal reasoning can emerge from gameplay, pointing to a promising strategy of designing surrogate tasks for RL post-training.

Shijia Liao, Shiyi Lan, Arun George Zachariah

The advent of Large Models marks a new era in machine learning, significantly outperforming smaller models by leveraging vast datasets to capture and synthesize complex patterns. Despite these advancements, the exploration into scaling, especially in the audio generation domain, remains limited, with previous efforts didn't extend into the high-fidelity (HiFi) 44.1kHz domain and suffering from both spectral discontinuities and blurriness in the high-frequency domain, alongside a lack of robustness against out-of-domain data. These limitations restrict the applicability of models to diverse use cases, including music and singing generation. Our work introduces Enhanced Various Audio Generation via Scalable Generative Adversarial Networks (EVA-GAN), yields significant improvements over previous state-of-the-art in spectral and high-frequency reconstruction and robustness in out-of-domain data performance, enabling the generation of HiFi audios by employing an extensive dataset of 36,000 hours of 44.1kHz audio, a context-aware module, a Human-In-The-Loop artifact measurement toolkit, and expands the model to approximately 200 million parameters. Demonstrations of our work are available at https://double-blind-eva-gan.cc.

Xinglong Sun, Barath Lakshmanan, Maying Shen et al.

Current structural pruning methods face two significant limitations: (i) they often limit pruning to finer-grained levels like channels, making aggressive parameter reduction challenging, and (ii) they focus heavily on parameter and FLOP reduction, with existing latency-aware methods frequently relying on simplistic, suboptimal linear models that fail to generalize well to transformers, where multiple interacting dimensions impact latency. In this paper, we address both limitations by introducing Multi-Dimensional Pruning (MDP), a novel paradigm that jointly optimizes across a variety of pruning granularities-including channels, query, key, heads, embeddings, and blocks. MDP employs an advanced latency modeling technique to accurately capture latency variations across all prunable dimensions, achieving an optimal balance between latency and accuracy. By reformulating pruning as a Mixed-Integer Nonlinear Program (MINLP), MDP efficiently identifies the optimal pruned structure across all prunable dimensions while respecting latency constraints. This versatile framework supports both CNNs and transformers. Extensive experiments demonstrate that MDP significantly outperforms previous methods, especially at high pruning ratios. On ImageNet, MDP achieves a 28% speed increase with a +1.4 Top-1 accuracy improvement over prior work like HALP for ResNet50 pruning. Against the latest transformer pruning method, Isomorphic, MDP delivers an additional 37% acceleration with a +0.7 Top-1 accuracy improvement.

Jingyu Song, Zhenxin Li, Shiyi Lan et al.

Benchmarking autonomous driving planners to align with human judgment remains a critical challenge, as state-of-the-art metrics like the Extended Predictive Driver Model Score (EPDMS) lack context awareness in nuanced scenarios. To address this, we introduce DriveCritic, a novel framework featuring two key contributions: the DriveCritic dataset, a curated collection of challenging scenarios where context is critical for correct judgment and annotated with pairwise human preferences, and the DriveCritic model, a Vision-Language Model (VLM) based evaluator. Fine-tuned using a two-stage supervised and reinforcement learning pipeline, the DriveCritic model learns to adjudicate between trajectory pairs by integrating visual and symbolic context. Experiments show DriveCritic significantly outperforms existing metrics and baselines in matching human preferences and demonstrates strong context awareness. Overall, our work provides a more reliable, human-aligned foundation to evaluating autonomous driving systems.

Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen et al. · nvidia, utoronto

Recent advances in large generative models have greatly enhanced both image editing and in-context image generation, yet a critical gap remains in ensuring physical consistency, where edited objects must remain coherent. This capability is especially vital for world simulation related tasks. In this paper, we present ChronoEdit, a framework that reframes image editing as a video generation problem. First, ChronoEdit treats the input and edited images as the first and last frames of a video, allowing it to leverage large pretrained video generative models that capture not only object appearance but also the implicit physics of motion and interaction through learned temporal consistency. Second, ChronoEdit introduces a temporal reasoning stage that explicitly performs editing at inference time. Under this setting, target frame is jointly denoised with reasoning tokens to imagine a plausible editing trajectory that constrains the solution space to physically viable transformations. The reasoning tokens are then dropped after a few steps to avoid the high computational cost of rendering a full video. To validate ChronoEdit, we introduce PBench-Edit, a new benchmark of image-prompt pairs for contexts that require physical consistency, and demonstrate that ChronoEdit surpasses state-of-the-art baselines in both visual fidelity and physical plausibility. Project page for code and models: https://research.nvidia.com/labs/toronto-ai/chronoedit

Xinglong Sun, Barath Lakshmanan, Maying Shen et al.

As we push the boundaries of performance in various vision tasks, the models grow in size correspondingly. To keep up with this growth, we need very aggressive pruning techniques for efficient inference and deployment on edge devices. Existing pruning approaches are limited to channel pruning and struggle with aggressive parameter reductions. In this paper, we propose a novel multi-dimensional pruning framework that jointly optimizes pruning across channels, layers, and blocks while adhering to latency constraints. We develop a latency modeling technique that accurately captures model-wide latency variations during pruning, which is crucial for achieving an optimal latency-accuracy trade-offs at high pruning ratio. We reformulate pruning as a Mixed-Integer Nonlinear Program (MINLP) to efficiently determine the optimal pruned structure with only a single pass. Our extensive results demonstrate substantial improvements over previous methods, particularly at large pruning ratios. In classification, our method significantly outperforms prior art HALP with a Top-1 accuracy of 70.0(v.s. 68.6) and an FPS of 5262 im/s(v.s. 4101 im/s). In 3D object detection, we establish a new state-of-the-art by pruning StreamPETR at a 45% pruning ratio, achieving higher FPS (37.3 vs. 31.7) and mAP (0.451 vs. 0.449) than the dense baseline.

Lingchen Meng, Hengduo Li, Bor-Chun Chen et al.

Built on top of self-attention mechanisms, vision transformers have demonstrated remarkable performance on a variety of vision tasks recently. While achieving excellent performance, they still require relatively intensive computational cost that scales up drastically as the numbers of patches, self-attention heads and transformer blocks increase. In this paper, we argue that due to the large variations among images, their need for modeling long-range dependencies between patches differ. To this end, we introduce AdaViT, an adaptive computation framework that learns to derive usage policies on which patches, self-attention heads and transformer blocks to use throughout the backbone on a per-input basis, aiming to improve inference efficiency of vision transformers with a minimal drop of accuracy for image recognition. Optimized jointly with a transformer backbone in an end-to-end manner, a light-weight decision network is attached to the backbone to produce decisions on-the-fly. Extensive experiments on ImageNet demonstrate that our method obtains more than 2x improvement on efficiency compared to state-of-the-art vision transformers with only 0.8% drop of accuracy, achieving good efficiency/accuracy trade-offs conditioned on different computational budgets. We further conduct quantitative and qualitative analysis on learned usage polices and provide more insights on the redundancy in vision transformers.

Shiyi Lan, Zhiding Yu, Christopher Choy et al.

We introduce DiscoBox, a novel framework that jointly learns instance segmentation and semantic correspondence using bounding box supervision. Specifically, we propose a self-ensembling framework where instance segmentation and semantic correspondence are jointly guided by a structured teacher in addition to the bounding box supervision. The teacher is a structured energy model incorporating a pairwise potential and a cross-image potential to model the pairwise pixel relationships both within and across the boxes. Minimizing the teacher energy simultaneously yields refined object masks and dense correspondences between intra-class objects, which are taken as pseudo-labels to supervise the task network and provide positive/negative correspondence pairs for dense constrastive learning. We show a symbiotic relationship where the two tasks mutually benefit from each other. Our best model achieves 37.9% AP on COCO instance segmentation, surpassing prior weakly supervised methods and is competitive to supervised methods. We also obtain state of the art weakly supervised results on PASCAL VOC12 and PF-PASCAL with real-time inference.

Jun Wang, Shiyi Lan, Mingfei Gao et al.

Real-time 3D object detection is crucial for autonomous cars. Achieving promising performance with high efficiency, voxel-based approaches have received considerable attention. However, previous methods model the input space with features extracted from equally divided sub-regions without considering that point cloud is generally non-uniformly distributed over the space. To address this issue, we propose a novel 3D object detection framework with dynamic information modeling. The proposed framework is designed in a coarse-to-fine manner. Coarse predictions are generated in the first stage via a voxel-based region proposal network. We introduce InfoFocus, which improves the coarse detections by adaptively refining features guided by the information of point cloud density. Experiments are conducted on the large-scale nuScenes 3D detection benchmark. Results show that our framework achieves the state-of-the-art performance with 31 FPS and improves our baseline significantly by 9.0% mAP on the nuScenes test set.

Shiyi Lan, Zhou Ren, Yi Wu et al.

Object detection is an essential step towards holistic scene understanding. Most existing object detection algorithms attend to certain object areas once and then predict the object locations. However, neuroscientists have revealed that humans do not look at the scene in fixed steadiness. Instead, human eyes move around, locating informative parts to understand the object location. This active perceiving movement process is called \textit{saccade}. %In this paper, Inspired by such mechanism, we propose a fast and accurate object detector called \textit{SaccadeNet}. It contains four main modules, the \cenam, the \coram, the \atm, and the \aggatt, which allows it to attend to different informative object keypoints, and predict object locations from coarse to fine. The \coram~is used only during training to extract more informative corner features which brings free-lunch performance boost. On the MS COCO dataset, we achieve the performance of 40.4\% mAP at 28 FPS and 30.5\% mAP at 118 FPS. Among all the real-time object detectors, %that can run faster than 25 FPS, our SaccadeNet achieves the best detection performance, which demonstrates the effectiveness of the proposed detection mechanism.

Shiyi Lan, Ruichi Yu, Gang Yu et al.

Recent advances in deep convolutional neural networks (CNNs) have motivated researchers to adapt CNNs to directly model points in 3D point clouds. Modeling local structure has been proven to be important for the success of convolutional architectures, and researchers exploited the modeling of local point sets in the feature extraction hierarchy. However, limited attention has been paid to explicitly model the geometric structure amongst points in a local region. To address this problem, we propose Geo-CNN, which applies a generic convolution-like operation dubbed as GeoConv to each point and its local neighborhood. Local geometric relationships among points are captured when extracting edge features between the center and its neighboring points. We first decompose the edge feature extraction process onto three orthogonal bases, and then aggregate the extracted features based on the angles between the edge vector and the bases. This encourages the network to preserve the geometric structure in Euclidean space throughout the feature extraction hierarchy. GeoConv is a generic and efficient operation that can be easily integrated into 3D point cloud analysis pipelines for multiple applications. We evaluate Geo-CNN on ModelNet40 and KITTI and achieve state-of-the-art performance.