Trung Pham

Arslan Ali, Junjie Bai, Maciej Bala et al. · nvidia

We introduce [Cosmos-Predict2.5], the latest generation of the Cosmos World Foundation Models for Physical AI. Built on a flow-based architecture, [Cosmos-Predict2.5] unifies Text2World, Image2World, and Video2World generation in a single model and leverages [Cosmos-Reason1], a Physical AI vision-language model, to provide richer text grounding and finer control of world simulation. Trained on 200M curated video clips and refined with reinforcement learning-based post-training, [Cosmos-Predict2.5] achieves substantial improvements over [Cosmos-Predict1] in video quality and instruction alignment, with models released at 2B and 14B scales. These capabilities enable more reliable synthetic data generation, policy evaluation, and closed-loop simulation for robotics and autonomous systems. We further extend the family with [Cosmos-Transfer2.5], a control-net style framework for Sim2Real and Real2Real world translation. Despite being 3.5$\times$ smaller than [Cosmos-Transfer1], it delivers higher fidelity and robust long-horizon video generation. Together, these advances establish [Cosmos-Predict2.5] and [Cosmos-Transfer2.5] as versatile tools for scaling embodied intelligence. To accelerate research and deployment in Physical AI, we release source code, pretrained checkpoints, and curated benchmarks under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-predict2.5 and https://github.com/nvidia-cosmos/cosmos-transfer2.5. We hope these open resources lower the barrier to adoption and foster innovation in building the next generation of embodied intelligence.

Aditi, Niket Agarwal, Arslan Ali et al.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Trung Pham, Chaoning Zhang, Axi Niu et al.

Exponential Moving Average (EMA or momentum) is widely used in modern self-supervised learning (SSL) approaches, such as MoCo, for enhancing performance. We demonstrate that such momentum can also be plugged into momentum-free SSL frameworks, such as SimCLR, for a performance boost. Despite its wide use as a fundamental component in modern SSL frameworks, the benefit caused by momentum is not well understood. We find that its success can be at least partly attributed to the stability effect. In the first attempt, we analyze how EMA affects each part of the encoder and reveal that the portion near the encoder's input plays an insignificant role while the latter parts have much more influence. By monitoring the gradient of the overall loss with respect to the output of each block in the encoder, we observe that the final layers tend to fluctuate much more than other layers during backpropagation, i.e. less stability. Interestingly, we show that using EMA to the final part of the SSL encoder, i.e. projector, instead of the whole deep network encoder can give comparable or preferable performance. Our proposed projector-only momentum helps maintain the benefit of EMA but avoids the double forward computation.

Thanh Nguyen, Trung Pham, Chaoning Zhang et al.

Self-supervised learning (SSL) has gained remarkable success, for which contrastive learning (CL) plays a key role. However, the recent development of new non-CL frameworks has achieved comparable or better performance with high improvement potential, prompting researchers to enhance these frameworks further. Assimilating CL into non-CL frameworks has been thought to be beneficial, but empirical evidence indicates no visible improvements. In view of that, this paper proposes a strategy of performing CL along the dimensional direction instead of along the batch direction as done in conventional contrastive learning, named Dimensional Contrastive Learning (DimCL). DimCL aims to enhance the feature diversity, and it can serve as a regularizer to prior SSL frameworks. DimCL has been found to be effective, and the hardness-aware property is identified as a critical reason for its success. Extensive experimental results reveal that assimilating DimCL into SSL frameworks leads to performance improvement by a non-trivial margin on various datasets and backbone architectures.

Yang Zheng, Oles Andrienko, Yonglei Zhao et al.

Regular object detection methods output rectangle bounding boxes, which are unable to accurately describe the actual object shapes. Instance segmentation methods output pixel-level labels, which are computationally expensive for real-time applications. Therefore, a polygon representation is needed to achieve precise shape alignment, while retaining low computation cost. We develop a novel Deformable Polar Polygon Object Detection method (DPPD) to detect objects in polygon shapes. In particular, our network predicts, for each object, a sparse set of flexible vertices to construct the polygon, where each vertex is represented by a pair of angle and distance in the Polar coordinate system. To enable training, both ground truth and predicted polygons are densely resampled to have the same number of vertices with equal-spaced raypoints. The resampling operation is fully differentable, allowing gradient back-propagation. Sparse polygon predicton ensures high-speed runtime inference while dense resampling allows the network to learn object shapes with high precision. The polygon detection head is established on top of an anchor-free and NMS-free network architecture. DPPD has been demonstrated successfully in various object detection tasks for autonomous driving such as traffic-sign, crosswalk, vehicle and pedestrian objects.

Trung Pham, Mehran Maghoumi, Wanli Jiang et al.

Achieving robust and real-time 3D perception is fundamental for autonomous vehicles. While most existing 3D perception methods prioritize detection accuracy, they often overlook critical aspects such as computational efficiency, onboard chip deployment friendliness, resilience to sensor mounting deviations, and adaptability to various vehicle types. To address these challenges, we present NVAutoNet: a specialized Bird's-Eye-View (BEV) perception network tailored explicitly for automated vehicles. NVAutoNet takes synchronized camera images as input and predicts 3D signals like obstacles, freespaces, and parking spaces. The core of NVAutoNet's architecture (image and BEV backbones) relies on efficient convolutional networks, optimized for high performance using TensorRT. More importantly, our image-to-BEV transformation employs simple linear layers and BEV look-up tables, ensuring rapid inference speed. Trained on an extensive proprietary dataset, NVAutoNet consistently achieves elevated perception accuracy, operating remarkably at 53 frames per second on the NVIDIA DRIVE Orin SoC. Notably, NVAutoNet demonstrates resilience to sensor mounting deviations arising from diverse car models. Moreover, NVAutoNet excels in adapting to varied vehicle types, facilitated by inexpensive model fine-tuning procedures that expedite compatibility adjustments.

Duy-Tho Le, Trung Pham, Jianfei Cai et al.

Optimizing the similarity between parametric shapes is crucial for numerous computer vision tasks, where Intersection over Union (IoU) stands as the canonical measure. However, existing optimization methods exhibit significant shortcomings: regression-based losses like L1/L2 lack correlation with IoU, IoU-based losses are unstable and limited to simple shapes, and task-specific methods are computationally intensive and not generalizable accross domains. As a result, the current landscape of parametric shape objective functions has become scattered, with each domain proposing distinct IoU approximations. To address this, we unify the parametric shape optimization objective functions by introducing Marginalized Generalized IoU (MGIoU), a novel loss function that overcomes these challenges by projecting structured convex shapes onto their unique shape Normals to compute one-dimensional normalized GIoU. MGIoU offers a simple, efficient, fully differentiable approximation strongly correlated with IoU. We then extend MGIoU to MGIoU+ that supports optimizing unstructured convex shapes. Together, MGIoU and MGIoU+ unify parametric shape optimization across diverse applications. Experiments on standard benchmarks demonstrate that MGIoU and MGIoU+ consistently outperform existing losses while reducing loss computation latency by 10-40x. Additionally, MGIoU and MGIoU+ satisfy metric properties and scale-invariance, ensuring robustness as an objective function. We further propose MGIoU- for minimizing overlaps in tasks like collision-free trajectory prediction. Code is available at https://ldtho.github.io/MGIoU

Thanh Nguyen, Tung Luu, Trung Pham et al.

Model agnostic meta-learning (MAML) is a popular state-of-the-art meta-learning algorithm that provides good weight initialization of a model given a variety of learning tasks. The model initialized by provided weight can be fine-tuned to an unseen task despite only using a small amount of samples and within a few adaptation steps. MAML is simple and versatile but requires costly learning rate tuning and careful design of the task distribution which affects its scalability and generalization. This paper proposes a more robust MAML based on an adaptive learning scheme and a prioritization task buffer(PTB) referred to as Robust MAML (RMAML) for improving scalability of training process and alleviating the problem of distribution mismatch. RMAML uses gradient-based hyper-parameter optimization to automatically find the optimal learning rate and uses the PTB to gradually adjust train-ing task distribution toward testing task distribution over the course of training. Experimental results on meta reinforcement learning environments demonstrate a substantial performance gain as well as being less sensitive to hyper-parameter choice and robust to distribution mismatch.

Junyeong Kim, Minuk Ma, Trung Pham et al.

This paper considers a network referred to as Modality Shifting Attention Network (MSAN) for Multimodal Video Question Answering (MVQA) task. MSAN decomposes the task into two sub-tasks: (1) localization of temporal moment relevant to the question, and (2) accurate prediction of the answer based on the localized moment. The modality required for temporal localization may be different from that for answer prediction, and this ability to shift modality is essential for performing the task. To this end, MSAN is based on (1) the moment proposal network (MPN) that attempts to locate the most appropriate temporal moment from each of the modalities, and also on (2) the heterogeneous reasoning network (HRN) that predicts the answer using an attention mechanism on both modalities. MSAN is able to place importance weight on the two modalities for each sub-task using a component referred to as Modality Importance Modulation (MIM). Experimental results show that MSAN outperforms previous state-of-the-art by achieving 71.13\% test accuracy on TVQA benchmark dataset. Extensive ablation studies and qualitative analysis are conducted to validate various components of the network.

Trung Pham, Vijay Kumar B G, Thanh-Toan Do et al.

This paper addresses the semantic instance segmentation task in the open-set conditions, where input images can contain known and unknown object classes. The training process of existing semantic instance segmentation methods requires annotation masks for all object instances, which is expensive to acquire or even infeasible in some realistic scenarios, where the number of categories may increase boundlessly. In this paper, we present a novel open-set semantic instance segmentation approach capable of segmenting all known and unknown object classes in images, based on the output of an object detector trained on known object classes. We formulate the problem using a Bayesian framework, where the posterior distribution is approximated with a simulated annealing optimization equipped with an efficient image partition sampler. We show empirically that our method is competitive with state-of-the-art supervised methods on known classes, but also performs well on unknown classes when compared with unsupervised methods.

Mehdi Hosseinzadeh, Yasir Latif, Trung Pham et al.

Simultaneous Localization And Mapping (SLAM) is a fundamental problem in mobile robotics. While point-based SLAM methods provide accurate camera localization, the generated maps lack semantic information. On the other hand, state of the art object detection methods provide rich information about entities present in the scene from a single image. This work marries the two and proposes a method for representing generic objects as quadrics which allows object detections to be seamlessly integrated in a SLAM framework. For scene coverage, additional dominant planar structures are modeled as infinite planes. Experiments show that the proposed points-planes-quadrics representation can easily incorporate Manhattan and object affordance constraints, greatly improving camera localization and leading to semantically meaningful maps. The performance of our SLAM system is demonstrated in https://youtu.be/dR-rB9keF8M .

Thanh-Toan Do, Ming Cai, Trung Pham et al.

Detecting objects and their 6D poses from only RGB images is an important task for many robotic applications. While deep learning methods have made significant progress in visual object detection and segmentation, the object pose estimation task is still challenging. In this paper, we introduce an end-toend deep learning framework, named Deep-6DPose, that jointly detects, segments, and most importantly recovers 6D poses of object instances from a single RGB image. In particular, we extend the recent state-of-the-art instance segmentation network Mask R-CNN with a novel pose estimation branch to directly regress 6D object poses without any post-refinements. Our key technical contribution is the decoupling of pose parameters into translation and rotation so that the rotation can be regressed via a Lie algebra representation. The resulting pose regression loss is differential and unconstrained, making the training tractable. The experiments on two standard pose benchmarking datasets show that our proposed approach compares favorably with the state-of-the-art RGB-based multi-stage pose estimation methods. Importantly, due to the end-to-end architecture, Deep-6DPose is considerably faster than competing multi-stage methods, offers an inference speed of 10 fps that is well suited for robotic applications.

Thanh-Toan Do, Tuan Hoang, Dang-Khoa Le Tan et al.

Learning compact binary codes for image retrieval task using deep neural networks has attracted increasing attention recently. However, training deep hashing networks for the task is challenging due to the binary constraints on the hash codes, the similarity preserving property, and the requirement for a vast amount of labelled images. To the best of our knowledge, none of the existing methods has tackled all of these challenges completely in a unified framework. In this work, we propose a novel end-to-end deep learning approach for the task, in which the network is trained to produce binary codes directly from image pixels without the need of manual annotation. In particular, to deal with the non-smoothness of binary constraints, we propose a novel pairwise constrained loss function, which simultaneously encodes the distances between pairs of hash codes, and the binary quantization error. In order to train the network with the proposed loss function, we propose an efficient parameter learning algorithm. In addition, to provide similar / dissimilar training images to train the network, we exploit 3D models reconstructed from unlabelled images for automatic generation of enormous training image pairs. The extensive experiments on image retrieval benchmark datasets demonstrate the improvements of the proposed method over the state-of-the-art compact representation methods on the image retrieval problem.

Toan Tran, Trung Pham, Gustavo Carneiro et al.

Data augmentation is an essential part of the training process applied to deep learning models. The motivation is that a robust training process for deep learning models depends on large annotated datasets, which are expensive to be acquired, stored and processed. Therefore a reasonable alternative is to be able to automatically generate new annotated training samples using a process known as data augmentation. The dominant data augmentation approach in the field assumes that new training samples can be obtained via random geometric or appearance transformations applied to annotated training samples, but this is a strong assumption because it is unclear if this is a reliable generative model for producing new training samples. In this paper, we provide a novel Bayesian formulation to data augmentation, where new annotated training points are treated as missing variables and generated based on the distribution learned from the training set. For learning, we introduce a theoretically sound algorithm --- generalised Monte Carlo expectation maximisation, and demonstrate one possible implementation via an extension of the Generative Adversarial Network (GAN). Classification results on MNIST, CIFAR-10 and CIFAR-100 show the better performance of our proposed method compared to the current dominant data augmentation approach mentioned above --- the results also show that our approach produces better classification results than similar GAN models.

Trung Pham, Thanh-Toan Do, Niko Sünderhauf et al.

This paper presents SceneCut, a novel approach to jointly discover previously unseen objects and non-object surfaces using a single RGB-D image. SceneCut's joint reasoning over scene semantics and geometry allows a robot to detect and segment object instances in complex scenes where modern deep learning-based methods either fail to separate object instances, or fail to detect objects that were not seen during training. SceneCut automatically decomposes a scene into meaningful regions which either represent objects or scene surfaces. The decomposition is qualified by an unified energy function over objectness and geometric fitting. We show how this energy function can be optimized efficiently by utilizing hierarchical segmentation trees. Moreover, we leverage a pre-trained convolutional oriented boundary network to predict accurate boundaries from images, which are used to construct high-quality region hierarchies. We evaluate SceneCut on several different indoor environments, and the results show that SceneCut significantly outperforms all the existing methods.

Jürgen Leitner, Adam W. Tow, Jake E. Dean et al.

Robotic challenges like the Amazon Picking Challenge (APC) or the DARPA Challenges are an established and important way to drive scientific progress. They make research comparable on a well-defined benchmark with equal test conditions for all participants. However, such challenge events occur only occasionally, are limited to a small number of contestants, and the test conditions are very difficult to replicate after the main event. We present a new physical benchmark challenge for robotic picking: the ACRV Picking Benchmark (APB). Designed to be reproducible, it consists of a set of 42 common objects, a widely available shelf, and exact guidelines for object arrangement using stencils. A well-defined evaluation protocol enables the comparison of \emph{complete} robotic systems -- including perception and manipulation -- instead of sub-systems only. Our paper also describes and reports results achieved by an open baseline system based on a Baxter robot.