Francesco Ferroni

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.

Alisson Azzolini, Junjie Bai, Hannah Brandon et al. · nvidia

Physical AI systems need to perceive, understand, and perform complex actions in the physical world. In this paper, we present the Cosmos-Reason1 models that can understand the physical world and generate appropriate embodied decisions (e.g., next step action) in natural language through long chain-of-thought reasoning processes. We begin by defining key capabilities for Physical AI reasoning, with a focus on physical common sense and embodied reasoning. To represent physical common sense, we use a hierarchical ontology that captures fundamental knowledge about space, time, and physics. For embodied reasoning, we rely on a two-dimensional ontology that generalizes across different physical embodiments. Building on these capabilities, we develop two multimodal large language models, Cosmos-Reason1-7B and Cosmos-Reason1-56B. We curate data and train our models in two stages: Physical AI supervised fine-tuning (SFT) and Physical AI reinforcement learning (RL). To evaluate our models, we build comprehensive benchmarks for physical common sense and embodied reasoning according to our ontologies. Evaluation results show that Physical AI SFT and RL bring significant improvements. To facilitate the development of Physical AI, we make our code and pre-trained models available under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-reason1.

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.

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 .

Aarti Basant, Amlan Kar, Despoina Paschalidou et al. · nvidia

As autonomous vehicle capabilities advance, the safe evaluation of driving policies in long-tail scenarios remains a critical bottleneck. In closed-loop simulation, the driving policy model actively interacts with the environment, where its actions dynamically update the simulator state and directly influence the next set of generated sensor observations. While recent reconstruction-based neural simulators offer photorealism, they are fundamentally constrained by their initial captured data and struggle to generalize to highly dynamic or novel scenes. To overcome these limitations, we introduce OmniDreams, a foundation generative world model mid- and post-trained from the Cosmos diffusion model to autoregressively generate action-conditioned videos in real time. By leveraging the rich visual priors of Cosmos and mid- and post-training on 21k hours of driving scenarios, OmniDreams synthesizes complex, unobserved phenomena that are hard for traditional simulators to capture, such as extreme weather and unpredictable dynamic agent behaviors. Crucially, it autoregressively conditions its photorealistic sensor generation on past frames, the current simulator state, and immediate driving actions. Deployed in a closed-loop system with the Alpamayo 1 policy model and AlpaSim orchestrator, OmniDreams acts as a highly responsive, reactive environment, providing a scalable and comprehensive solution for training and evaluating next-generation autonomous driving policies. We additionally show preliminary results indicating that a world-action model (WAM) post-trained from OmniDreams achieves strong performance on the Physical AI Autonomous Vehicles NuRec dataset, surpassing the VLA-based Alpamayo 1.5 research policy model while using only 1/5 the total parameters. These results highlight the potential for a real-time world model like OmniDreams to also serve as a backbone for policy architectures.

Shubham Gupta, Jeet Kanjani, Mengtian Li et al. · gatech

We focus on the task of far-field 3D detection (Far3Det) of objects beyond a certain distance from an observer, e.g., $>$50m. Far3Det is particularly important for autonomous vehicles (AVs) operating at highway speeds, which require detections of far-field obstacles to ensure sufficient braking distances. However, contemporary AV benchmarks such as nuScenes underemphasize this problem because they evaluate performance only up to a certain distance (50m). One reason is that obtaining far-field 3D annotations is difficult, particularly for lidar sensors that produce very few point returns for far-away objects. Indeed, we find that almost 50% of far-field objects (beyond 50m) contain zero lidar points. Secondly, current metrics for 3D detection employ a "one-size-fits-all" philosophy, using the same tolerance thresholds for near and far objects, inconsistent with tolerances for both human vision and stereo disparities. Both factors lead to an incomplete analysis of the Far3Det task. For example, while conventional wisdom tells us that high-resolution RGB sensors should be vital for 3D detection of far-away objects, lidar-based methods still rank higher compared to RGB counterparts on the current benchmark leaderboards. As a first step towards a Far3Det benchmark, we develop a method to find well-annotated scenes from the nuScenes dataset and derive a well-annotated far-field validation set. We also propose a Far3Det evaluation protocol and explore various 3D detection methods for Far3Det. Our result convincingly justifies the long-held conventional wisdom that high-resolution RGB improves 3D detection in the far-field. We further propose a simple yet effective method that fuses detections from RGB and lidar detectors based on non-maximum suppression, which remarkably outperforms state-of-the-art 3D detectors in the far-field.

Haithem Turki, Jason Y. Zhang, Francesco Ferroni et al.

We extend neural radiance fields (NeRFs) to dynamic large-scale urban scenes. Prior work tends to reconstruct single video clips of short durations (up to 10 seconds). Two reasons are that such methods (a) tend to scale linearly with the number of moving objects and input videos because a separate model is built for each and (b) tend to require supervision via 3D bounding boxes and panoptic labels, obtained manually or via category-specific models. As a step towards truly open-world reconstructions of dynamic cities, we introduce two key innovations: (a) we factorize the scene into three separate hash table data structures to efficiently encode static, dynamic, and far-field radiance fields, and (b) we make use of unlabeled target signals consisting of RGB images, sparse LiDAR, off-the-shelf self-supervised 2D descriptors, and most importantly, 2D optical flow. Operationalizing such inputs via photometric, geometric, and feature-metric reconstruction losses enables SUDS to decompose dynamic scenes into the static background, individual objects, and their motions. When combined with our multi-branch table representation, such reconstructions can be scaled to tens of thousands of objects across 1.2 million frames from 1700 videos spanning geospatial footprints of hundreds of kilometers, (to our knowledge) the largest dynamic NeRF built to date. We present qualitative initial results on a variety of tasks enabled by our representations, including novel-view synthesis of dynamic urban scenes, unsupervised 3D instance segmentation, and unsupervised 3D cuboid detection. To compare to prior work, we also evaluate on KITTI and Virtual KITTI 2, surpassing state-of-the-art methods that rely on ground truth 3D bounding box annotations while being 10x quicker to train.

Jia Shi, Gautam Gare, Jinjin Tian et al.

We tackle the challenge of predicting models' Out-of-Distribution (OOD) performance using in-distribution (ID) measurements without requiring OOD data. Existing evaluations with "Effective Robustness", which use ID accuracy as an indicator of OOD accuracy, encounter limitations when models are trained with diverse supervision and distributions, such as class labels (Vision Models, VMs, on ImageNet) and textual descriptions (Visual-Language Models, VLMs, on LAION). VLMs often generalize better to OOD data than VMs despite having similar or lower ID performance. To improve the prediction of models' OOD performance from ID measurements, we introduce the Lowest Common Ancestor (LCA)-on-the-Line framework. This approach revisits the established concept of LCA distance, which measures the hierarchical distance between labels and predictions within a predefined class hierarchy, such as WordNet. We assess 75 models using ImageNet as the ID dataset and five significantly shifted OOD variants, uncovering a strong linear correlation between ID LCA distance and OOD top-1 accuracy. Our method provides a compelling alternative for understanding why VLMs tend to generalize better. Additionally, we propose a technique to construct a taxonomic hierarchy on any dataset using K-means clustering, demonstrating that LCA distance is robust to the constructed taxonomic hierarchy. Moreover, we demonstrate that aligning model predictions with class taxonomies, through soft labels or prompt engineering, can enhance model generalization. Open source code in our Project Page: https://elvishelvis.github.io/papers/lca/.

Xueqian Li, Jianqiao Zheng, Francesco Ferroni et al.

Neural Scene Flow Prior (NSFP) is of significant interest to the vision community due to its inherent robustness to out-of-distribution (OOD) effects and its ability to deal with dense lidar points. The approach utilizes a coordinate neural network to estimate scene flow at runtime, without any training. However, it is up to 100 times slower than current state-of-the-art learning methods. In other applications such as image, video, and radiance function reconstruction innovations in speeding up the runtime performance of coordinate networks have centered upon architectural changes. In this paper, we demonstrate that scene flow is different -- with the dominant computational bottleneck stemming from the loss function itself (i.e., Chamfer distance). Further, we rediscover the distance transform (DT) as an efficient, correspondence-free loss function that dramatically speeds up the runtime optimization. Our fast neural scene flow (FNSF) approach reports for the first time real-time performance comparable to learning methods, without any training or OOD bias on two of the largest open autonomous driving (AV) lidar datasets Waymo Open and Argoverse.

Chittesh Thavamani, Mengtian Li, Francesco Ferroni et al.

Many perception systems in mobile computing, autonomous navigation, and AR/VR face strict compute constraints that are particularly challenging for high-resolution input images. Previous works propose nonuniform downsamplers that "learn to zoom" on salient image regions, reducing compute while retaining task-relevant image information. However, for tasks with spatial labels (such as 2D/3D object detection and semantic segmentation), such distortions may harm performance. In this work (LZU), we "learn to zoom" in on the input image, compute spatial features, and then "unzoom" to revert any deformations. To enable efficient and differentiable unzooming, we approximate the zooming warp with a piecewise bilinear mapping that is invertible. LZU can be applied to any task with 2D spatial input and any model with 2D spatial features, and we demonstrate this versatility by evaluating on a variety of tasks and datasets: object detection on Argoverse-HD, semantic segmentation on Cityscapes, and monocular 3D object detection on nuScenes. Interestingly, we observe boosts in performance even when high-resolution sensor data is unavailable, implying that LZU can be used to "learn to upsample" as well.

Abhinav Agarwalla, Xuhua Huang, Jason Ziglar et al.

State-of-the-art lidar panoptic segmentation (LPS) methods follow bottom-up segmentation-centric fashion wherein they build upon semantic segmentation networks by utilizing clustering to obtain object instances. In this paper, we re-think this approach and propose a surprisingly simple yet effective detection-centric network for both LPS and tracking. Our network is modular by design and optimized for all aspects of both the panoptic segmentation and tracking task. One of the core components of our network is the object instance detection branch, which we train using point-level (modal) annotations, as available in segmentation-centric datasets. In the absence of amodal (cuboid) annotations, we regress modal centroids and object extent using trajectory-level supervision that provides information about object size, which cannot be inferred from single scans due to occlusions and the sparse nature of the lidar data. We obtain fine-grained instance segments by learning to associate lidar points with detected centroids. We evaluate our method on several 3D/4D LPS benchmarks and observe that our model establishes a new state-of-the-art among open-sourced models, outperforming recent query-based models.

Xindi Wu, KwunFung Lau, Francesco Ferroni et al.

Self-driving vehicles rely on urban street maps for autonomous navigation. In this paper, we introduce Pix2Map, a method for inferring urban street map topology directly from ego-view images, as needed to continually update and expand existing maps. This is a challenging task, as we need to infer a complex urban road topology directly from raw image data. The main insight of this paper is that this problem can be posed as cross-modal retrieval by learning a joint, cross-modal embedding space for images and existing maps, represented as discrete graphs that encode the topological layout of the visual surroundings. We conduct our experimental evaluation using the Argoverse dataset and show that it is indeed possible to accurately retrieve street maps corresponding to both seen and unseen roads solely from image data. Moreover, we show that our retrieved maps can be used to update or expand existing maps and even show proof-of-concept results for visual localization and image retrieval from spatial graphs.

Nadine Chang, Francesco Ferroni, Michael J. Tarr et al.

Large-scale datasets are essential to modern day deep learning. Advocates argue that understanding these methods requires dataset transparency (e.g. "dataset curation, motivation, composition, collection process, etc..."). However, almost no one has suggested the release of the detailed definitions and visual category examples provided to annotators - information critical to understanding the structure of the annotations present in each dataset. These labels are at the heart of public datasets, yet few datasets include the instructions that were used to generate them. We introduce a new task, Labeling Instruction Generation, to address missing publicly available labeling instructions. In Labeling Instruction Generation, we take a reasonably annotated dataset and: 1) generate a set of examples that are visually representative of each category in the dataset; 2) provide a text label that corresponds to each of the examples. We introduce a framework that requires no model training to solve this task and includes a newly created rapid retrieval system that leverages a large, pre-trained vision and language model. This framework acts as a proxy to human annotators that can help to both generate a final labeling instruction set and evaluate its quality. Our framework generates multiple diverse visual and text representations of dataset categories. The optimized instruction set outperforms our strongest baseline across 5 folds by 7.06 mAP for NuImages and 12.9 mAP for COCO.

Di Feng, Francesco Ferroni

Transfomer-based approaches advance the recent development of multi-camera 3D detection both in academia and industry. In a vanilla transformer architecture, queries are randomly initialised and optimised for the whole dataset, without considering the differences among input frames. In this work, we propose to leverage the predictions from an image backbone, which is often highly optimised for 2D tasks, as priors to the transformer part of a 3D detection network. The method works by (1). augmenting image feature maps with 2D priors, (2). sampling query locations via ray-casting along 2D box centroids, as well as (3). initialising query features with object-level image features. Experimental results shows that 2D priors not only help the model converge faster, but also largely improve the baseline approach by up to 12% in terms of average precision.

Aljoša Ošep, Tim Meinhardt, Francesco Ferroni et al.

We propose the SAL (Segment Anything in Lidar) method consisting of a text-promptable zero-shot model for segmenting and classifying any object in Lidar, and a pseudo-labeling engine that facilitates model training without manual supervision. While the established paradigm for Lidar Panoptic Segmentation (LPS) relies on manual supervision for a handful of object classes defined a priori, we utilize 2D vision foundation models to generate 3D supervision ``for free''. Our pseudo-labels consist of instance masks and corresponding CLIP tokens, which we lift to Lidar using calibrated multi-modal data. By training our model on these labels, we distill the 2D foundation models into our Lidar SAL model. Even without manual labels, our model achieves $91\%$ in terms of class-agnostic segmentation and $54\%$ in terms of zero-shot Lidar Panoptic Segmentation of the fully supervised state-of-the-art. Furthermore, we outperform several baselines that do not distill but only lift image features to 3D. More importantly, we demonstrate that SAL supports arbitrary class prompts, can be easily extended to new datasets, and shows significant potential to improve with increasing amounts of self-labeled data. Code and models are available at this $\href{https://github.com/nv-dvl/segment-anything-lidar}{URL}$.

Jenny Seidenschwarz, Aljoša Ošep, Francesco Ferroni et al.

We tackle semi-supervised object detection based on motion cues. Recent results suggest that heuristic-based clustering methods in conjunction with object trackers can be used to pseudo-label instances of moving objects and use these as supervisory signals to train 3D object detectors in Lidar data without manual supervision. We re-think this approach and suggest that both, object detection, as well as motion-inspired pseudo-labeling, can be tackled in a data-driven manner. We leverage recent advances in scene flow estimation to obtain point trajectories from which we extract long-term, class-agnostic motion patterns. Revisiting correlation clustering in the context of message passing networks, we learn to group those motion patterns to cluster points to object instances. By estimating the full extent of the objects, we obtain per-scan 3D bounding boxes that we use to supervise a Lidar object detection network. Our method not only outperforms prior heuristic-based approaches (57.5 AP, +14 improvement over prior work), more importantly, we show we can pseudo-label and train object detectors across datasets.

Nutan Chen, Djalel Benbouzid, Francesco Ferroni et al.

The use of machine learning in artistic music generation leads to controversial discussions of the quality of art, for which objective quantification is nonsensical. We therefore consider a music-generating algorithm as a counterpart to a human musician, in a setting where reciprocal interplay is to lead to new experiences, both for the musician and the audience. To obtain this behaviour, we resort to the framework of recurrent Variational Auto-Encoders (VAE) and learn to generate music, seeded by a human musician. In the learned model, we generate novel musical sequences by interpolation in latent space. Standard VAEs however do not guarantee any form of smoothness in their latent representation. This translates into abrupt changes in the generated music sequences. To overcome these limitations, we regularise the decoder and endow the latent space with a flat Riemannian manifold, i.e., a manifold that is isometric to the Euclidean space. As a result, linearly interpolating in the latent space yields realistic and smooth musical changes that fit the type of machine--musician interactions we aim for. We provide empirical evidence for our method via a set of experiments on music datasets and we deploy our model for an interactive jam session with a professional drummer. The live performance provides qualitative evidence that the latent representation can be intuitively interpreted and exploited by the drummer to drive the interplay. Beyond the musical application, our approach showcases an instance of human-centred design of machine-learning models, driven by interpretability and the interaction with the end user.

Nutan Chen, Alexej Klushyn, Francesco Ferroni et al.

Measuring the similarity between data points often requires domain knowledge, which can in parts be compensated by relying on unsupervised methods such as latent-variable models, where similarity/distance is estimated in a more compact latent space. Prevalent is the use of the Euclidean metric, which has the drawback of ignoring information about similarity of data stored in the decoder, as captured by the framework of Riemannian geometry. We propose an extension to the framework of variational auto-encoders allows learning flat latent manifolds, where the Euclidean metric is a proxy for the similarity between data points. This is achieved by defining the latent space as a Riemannian manifold and by regularising the metric tensor to be a scaled identity matrix. Additionally, we replace the compact prior typically used in variational auto-encoders with a recently presented, more expressive hierarchical one---and formulate the learning problem as a constrained optimisation problem. We evaluate our method on a range of data-sets, including a video-tracking benchmark, where the performance of our unsupervised approach nears that of state-of-the-art supervised approaches, while retaining the computational efficiency of straight-line-based approaches.

Janis Postels, Francesco Ferroni, Huseyin Coskun et al.

We present a sampling-free approach for computing the epistemic uncertainty of a neural network. Epistemic uncertainty is an important quantity for the deployment of deep neural networks in safety-critical applications, since it represents how much one can trust predictions on new data. Recently promising works were proposed using noise injection combined with Monte-Carlo sampling at inference time to estimate this quantity (e.g. Monte-Carlo dropout). Our main contribution is an approximation of the epistemic uncertainty estimated by these methods that does not require sampling, thus notably reducing the computational overhead. We apply our approach to large-scale visual tasks (i.e., semantic segmentation and depth regression) to demonstrate the advantages of our method compared to sampling-based approaches in terms of quality of the uncertainty estimates as well as of computational overhead.

Nutan Chen, Francesco Ferroni, Alexej Klushyn et al.

The length of the geodesic between two data points along a Riemannian manifold, induced by a deep generative model, yields a principled measure of similarity. Current approaches are limited to low-dimensional latent spaces, due to the computational complexity of solving a non-convex optimisation problem. We propose finding shortest paths in a finite graph of samples from the aggregate approximate posterior, that can be solved exactly, at greatly reduced runtime, and without a notable loss in quality. Our approach, therefore, is hence applicable to high-dimensional problems, e.g., in the visual domain. We validate our approach empirically on a series of experiments using variational autoencoders applied to image data, including the Chair, FashionMNIST, and human movement data sets.