An-Chieh Cheng

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 .

An-Chieh Cheng, Xueting Li, Sifei Liu et al.

With the capacity of modeling long-range dependencies in sequential data, transformers have shown remarkable performances in a variety of generative tasks such as image, audio, and text generation. Yet, taming them in generating less structured and voluminous data formats such as high-resolution point clouds have seldom been explored due to ambiguous sequentialization processes and infeasible computation burden. In this paper, we aim to further exploit the power of transformers and employ them for the task of 3D point cloud generation. The key idea is to decompose point clouds of one category into semantically aligned sequences of shape compositions, via a learned canonical space. These shape compositions can then be quantized and used to learn a context-rich composition codebook for point cloud generation. Experimental results on point cloud reconstruction and unconditional generation show that our model performs favorably against state-of-the-art approaches. Furthermore, our model can be easily extended to multi-modal shape completion as an application for conditional shape generation.

An-Chieh Cheng, Yang Fu, Yatai Ji et al.

We present GR3D, a spatial vision language model equipped with three complementary grounding capabilities--explicit 2D grounding, implicit 2D grounding, and monocular 3D grounding--within a single framework. GR3D introduces an implicit grounding mechanism that identifies entity mentions during generation and inserts the corresponding region tokens into the text stream, allowing the model to reference visual evidence on the fly when producing spatial chain-of-thought responses. In parallel, a region-prompted monocular 3D grounding design predicts 3D bounding boxes in the camera view from grounded region queries, supported by intrinsic-aware normalization and dense geometric supervision. Together, these grounding capabilities enable GR3D to decompose complex spatial understanding problems into grounded 2D perception followed by 3D inference. GR3D achieves consistent improvements across grounded and non-grounded spatial benchmarks, demonstrating grounding as an effective inductive bias for strengthening spatial understanding in VLMs. These grounding capabilities collectively enhance general spatial understanding beyond the grounding task itself.

Zhijian Liu, Ligeng Zhu, Baifeng Shi et al.

Visual language models (VLMs) have made significant advances in accuracy in recent years. However, their efficiency has received much less attention. This paper introduces NVILA, a family of open VLMs designed to optimize both efficiency and accuracy. Building on top of VILA, we improve its model architecture by first scaling up the spatial and temporal resolutions, and then compressing visual tokens. This "scale-then-compress" approach enables NVILA to efficiently process high-resolution images and long videos. We also conduct a systematic investigation to enhance the efficiency of NVILA throughout its entire lifecycle, from training and fine-tuning to deployment. NVILA matches or surpasses the accuracy of many leading open and proprietary VLMs across a wide range of image and video benchmarks. At the same time, it reduces training costs by 4.5X, fine-tuning memory usage by 3.4X, pre-filling latency by 1.6-2.2X, and decoding latency by 1.2-2.8X. We will soon make our code and models available to facilitate reproducibility.

Hanrong Ye, Chao-Han Huck Yang, Arushi Goel et al.

Advancing machine intelligence requires developing the ability to perceive across multiple modalities, much as humans sense the world. We introduce OmniVinci, an initiative to build a strong, open-source, omni-modal LLM. We carefully study the design choices across model architecture and data curation. For model architecture, we present three key innovations: (i) OmniAlignNet for strengthening alignment between vision and audio embeddings in a shared omni-modal latent space; (ii) Temporal Embedding Grouping for capturing relative temporal alignment between vision and audio signals; and (iii) Constrained Rotary Time Embedding for encoding absolute temporal information in omni-modal embeddings. We introduce a curation and synthesis pipeline that generates 24M single-modal and omni-modal conversations. We find that modalities reinforce one another in both perception and reasoning. Our model, OmniVinci, outperforms Qwen2.5-Omni with +19.05 on DailyOmni (cross-modal understanding), +1.7 on MMAR (audio), and +3.9 on Video-MME (vision), while using just 0.2T training tokens - a 6 times reduction compared to Qwen2.5-Omni's 1.2T. We finally demonstrate omni-modal advantages in downstream applications spanning robotics, medical AI, and smart factory.

An-Chieh Cheng, Yandong Ji, Zhaojing Yang et al.

This paper proposes to solve the problem of Vision-and-Language Navigation with legged robots, which not only provides a flexible way for humans to command but also allows the robot to navigate through more challenging and cluttered scenes. However, it is non-trivial to translate human language instructions all the way to low-level leg joint actions. We propose NaVILA, a 2-level framework that unifies a Vision-Language-Action model (VLA) with locomotion skills. Instead of directly predicting low-level actions from VLA, NaVILA first generates mid-level actions with spatial information in the form of language, (e.g., "moving forward 75cm"), which serves as an input for a visual locomotion RL policy for execution. NaVILA substantially improves previous approaches on existing benchmarks. The same advantages are demonstrated in our newly developed benchmarks with IsaacLab, featuring more realistic scenes, low-level controls, and real-world robot experiments. We show more results at https://navila-bot.github.io/

Ruihan Yang, Qinxi Yu, Yecheng Wu et al.

Real robot data collection for imitation learning has led to significant advancements in robotic manipulation. However, the requirement for robot hardware in the process fundamentally constrains the scale of the data. In this paper, we explore training Vision-Language-Action (VLA) models using egocentric human videos. The benefit of using human videos is not only for their scale but more importantly for the richness of scenes and tasks. With a VLA trained on human video that predicts human wrist and hand actions, we can perform Inverse Kinematics and retargeting to convert the human actions to robot actions. We fine-tune the model using a few robot manipulation demonstrations to obtain the robot policy, namely EgoVLA. We propose a simulation benchmark called Ego Humanoid Manipulation Benchmark, where we design diverse bimanual manipulation tasks with demonstrations. We fine-tune and evaluate EgoVLA with Ego Humanoid Manipulation Benchmark and show significant improvements over baselines and ablate the importance of human data. Videos can be found on our website: https://rchalyang.github.io/EgoVLA

Isabella Liu, An-Chieh Cheng, Rui Yan et al.

Long-horizon manipulation remains challenging for vision-language-action (VLA) policies: real tasks are multi-step, progress-dependent, and brittle to compounding execution errors. We present LoHo-Manip, a modular framework that scales short-horizon VLA execution to long-horizon instruction following via a dedicated task-management VLM. The manager is decoupled from the executor and is invoked in a receding-horizon manner: given the current observation, it predicts a progress-aware remaining plan that combines (i) a subtask sequence with an explicit done + remaining split as lightweight language memory, and (ii) a visual trace -- a compact 2D keypoint trajectory prompt specifying where to go and what to approach next. The executor VLA is adapted to condition on the rendered trace, thereby turning long-horizon decision-making into repeated local control by following the trace. Crucially, predicting the remaining plan at each step yields an implicit closed loop: failed steps persist in subsequent outputs, and traces update accordingly, enabling automatic continuation and replanning without hand-crafted recovery logic or brittle visual-history buffers. Extensive experiments spanning embodied planning, long-horizon reasoning, trajectory prediction, and end-to-end manipulation in simulation and on a real Franka robot demonstrate strong gains in long-horizon success, robustness, and out-of-distribution generalization. Project page: https://www.liuisabella.com/LoHoManip

An-Chieh Cheng, Yang Fu, Yukang Chen et al.

We present Spatial Region 3D (SR-3D) aware vision-language model that connects single-view 2D images and multi-view 3D data through a shared visual token space. SR-3D supports flexible region prompting, allowing users to annotate regions with bounding boxes, segmentation masks on any frame, or directly in 3D, without the need for exhaustive multi-frame labeling. We achieve this by enriching 2D visual features with 3D positional embeddings, which allows the 3D model to draw upon strong 2D priors for more accurate spatial reasoning across frames, even when objects of interest do not co-occur within the same view. Extensive experiments on both general 2D vision language and specialized 3D spatial benchmarks demonstrate that SR-3D achieves state-of-the-art performance, underscoring its effectiveness for unifying 2D and 3D representation space on scene understanding. Moreover, we observe applicability to in-the-wild videos without sensory 3D inputs or ground-truth 3D annotations, where SR-3D accurately infers spatial relationships and metric measurements.

An-Chieh Cheng, Hongxu Yin, Yang Fu et al.

Vision Language Models (VLMs) have demonstrated remarkable performance in 2D vision and language tasks. However, their ability to reason about spatial arrangements remains limited. In this work, we introduce Spatial Region GPT (SpatialRGPT) to enhance VLMs' spatial perception and reasoning capabilities. SpatialRGPT advances VLMs' spatial understanding through two key innovations: (1) a data curation pipeline that enables effective learning of regional representation from 3D scene graphs, and (2) a flexible plugin module for integrating depth information into the visual encoder of existing VLMs. During inference, when provided with user-specified region proposals, SpatialRGPT can accurately perceive their relative directions and distances. Additionally, we propose SpatialRGBT-Bench, a benchmark with ground-truth 3D annotations encompassing indoor, outdoor, and simulated environments, for evaluating 3D spatial cognition in VLMs. Our results demonstrate that SpatialRGPT significantly enhances performance in spatial reasoning tasks, both with and without local region prompts. The model also exhibits strong generalization capabilities, effectively reasoning about complex spatial relations and functioning as a region-aware dense reward annotator for robotic tasks. Code, dataset, and benchmark are released at https://www.anjiecheng.me/SpatialRGPT

An-Chieh Cheng, Xueting Li, Sifei Liu et al.

Textures are a vital aspect of creating visually appealing and realistic 3D models. In this paper, we study the problem of generating high-fidelity texture given shapes of 3D assets, which has been relatively less explored compared with generic 3D shape modeling. Our goal is to facilitate a controllable texture generation process, such that one texture code can correspond to a particular appearance style independent of any input shapes from a category. We introduce Texture UV Radiance Fields (TUVF) that generate textures in a learnable UV sphere space rather than directly on the 3D shape. This allows the texture to be disentangled from the underlying shape and transferable to other shapes that share the same UV space, i.e., from the same category. We integrate the UV sphere space with the radiance field, which provides a more efficient and accurate representation of textures than traditional texture maps. We perform our experiments on synthetic and real-world object datasets where we achieve not only realistic synthesis but also substantial improvements over state-of-the-arts on texture controlling and editing. Project Page: https://www.anjiecheng.me/TUVF

An-Chieh Cheng, Xueting Li, Min Sun et al.

We propose a canonical point autoencoder (CPAE) that predicts dense correspondences between 3D shapes of the same category. The autoencoder performs two key functions: (a) encoding an arbitrarily ordered point cloud to a canonical primitive, e.g., a sphere, and (b) decoding the primitive back to the original input instance shape. As being placed in the bottleneck, this primitive plays a key role to map all the unordered point clouds on the canonical surface and to be reconstructed in an ordered fashion. Once trained, points from different shape instances that are mapped to the same locations on the primitive surface are determined to be a pair of correspondence. Our method does not require any form of annotation or self-supervised part segmentation network and can handle unaligned input point clouds. Experimental results on 3D semantic keypoint transfer and part segmentation transfer show that our model performs favorably against state-of-the-art correspondence learning methods.

An-Chieh Cheng, Chieh Hubert Lin, Da-Cheng Juan et al.

Conventional Neural Architecture Search (NAS) aims at finding a single architecture that achieves the best performance, which usually optimizes task related learning objectives such as accuracy. However, a single architecture may not be representative enough for the whole dataset with high diversity and variety. Intuitively, electing domain-expert architectures that are proficient in domain-specific features can further benefit architecture related objectives such as latency. In this paper, we propose InstaNAS---an instance-aware NAS framework---that employs a controller trained to search for a "distribution of architectures" instead of a single architecture; This allows the model to use sophisticated architectures for the difficult samples, which usually comes with large architecture related cost, and shallow architectures for those easy samples. During the inference phase, the controller assigns each of the unseen input samples with a domain expert architecture that can achieve high accuracy with customized inference costs. Experiments within a search space inspired by MobileNetV2 show InstaNAS can achieve up to 48.8% latency reduction without compromising accuracy on a series of datasets against MobileNetV2.

Hsuan-Kung Yang, An-Chieh Cheng, Kuan-Wei Ho et al.

In this paper, we investigate the use of an unsupervised label clustering technique and demonstrate that it enables substantial improvements in visual relationship prediction accuracy on the Person in Context (PIC) dataset. We propose to group object labels with similar patterns of relationship distribution in the dataset into fewer categories. Label clustering not only mitigates both the large classification space and class imbalance issues, but also potentially increases data samples for each clustered category. We further propose to incorporate depth information as an additional feature into the instance segmentation model. The additional depth prediction path supplements the relationship prediction model in a way that bounding boxes or segmentation masks are unable to deliver. We have rigorously evaluated the proposed techniques and performed various ablation analysis to validate the benefits of them.

An-Chieh Cheng, Jin-Dong Dong, Chi-Hung Hsu et al.

Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performance in many tasks such as image classification and language understanding. However, most existing works only optimize for model accuracy and largely ignore other important factors imposed by the underlying hardware and devices, such as latency and energy, when making inference. In this paper, we first introduce the problem of NAS and provide a survey on recent works. Then we deep dive into two recent advancements on extending NAS into multiple-objective frameworks: MONAS and DPP-Net. Both MONAS and DPP-Net are capable of optimizing accuracy and other objectives imposed by devices, searching for neural architectures that can be best deployed on a wide spectrum of devices: from embedded systems and mobile devices to workstations. Experimental results are poised to show that architectures found by MONAS and DPP-Net achieves Pareto optimality w.r.t the given objectives for various devices.

Jin-Dong Dong, An-Chieh Cheng, Da-Cheng Juan et al.

Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performances in applications such as image classification and language modeling. However, these techniques typically ignore device-related objectives such as inference time, memory usage, and power consumption. Optimizing neural architecture for device-related objectives is immensely crucial for deploying deep networks on portable devices with limited computing resources. We propose DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures, optimizing for both device-related (e.g., inference time and memory usage) and device-agnostic (e.g., accuracy and model size) objectives. DPP-Net employs a compact search space inspired by current state-of-the-art mobile CNNs, and further improves search efficiency by adopting progressive search (Liu et al. 2017). Experimental results on CIFAR-10 are poised to demonstrate the effectiveness of Pareto-optimal networks found by DPP-Net, for three different devices: (1) a workstation with Titan X GPU, (2) NVIDIA Jetson TX1 embedded system, and (3) mobile phone with ARM Cortex-A53. Compared to CondenseNet and NASNet (Mobile), DPP-Net achieves better performances: higher accuracy and shorter inference time on various devices. Additional experimental results show that models found by DPP-Net also achieve considerably-good performance on ImageNet as well.