Michael S. Ryoo

Anthony Brohan, Noah Brown, Justice Carbajal et al. · stanford

We study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning. Our goal is to enable a single end-to-end trained model to both learn to map robot observations to actions and enjoy the benefits of large-scale pretraining on language and vision-language data from the web. To this end, we propose to co-fine-tune state-of-the-art vision-language models on both robotic trajectory data and Internet-scale vision-language tasks, such as visual question answering. In contrast to other approaches, we propose a simple, general recipe to achieve this goal: in order to fit both natural language responses and robotic actions into the same format, we express the actions as text tokens and incorporate them directly into the training set of the model in the same way as natural language tokens. We refer to such category of models as vision-language-action models (VLA) and instantiate an example of such a model, which we call RT-2. Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data (such as placing an object onto a particular number or icon), and the ability to perform rudimentary reasoning in response to user commands (such as picking up the smallest or largest object, or the one closest to another object). We further show that incorporating chain of thought reasoning allows RT-2 to perform multi-stage semantic reasoning, for example figuring out which object to pick up for use as an improvised hammer (a rock), or which type of drink is best suited for someone who is tired (an energy drink).

Boyuan Chen, Fei Xia, Brian Ichter et al. · mit

Large language models (LLMs) have unlocked new capabilities of task planning from human instructions. However, prior attempts to apply LLMs to real-world robotic tasks are limited by the lack of grounding in the surrounding scene. In this paper, we develop NLMap, an open-vocabulary and queryable scene representation to address this problem. NLMap serves as a framework to gather and integrate contextual information into LLM planners, allowing them to see and query available objects in the scene before generating a context-conditioned plan. NLMap first establishes a natural language queryable scene representation with Visual Language models (VLMs). An LLM based object proposal module parses instructions and proposes involved objects to query the scene representation for object availability and location. An LLM planner then plans with such information about the scene. NLMap allows robots to operate without a fixed list of objects nor executable options, enabling real robot operation unachievable by previous methods. Project website: https://nlmap-saycan.github.io

Anthony Brohan, Noah Brown, Justice Carbajal et al.

By transferring knowledge from large, diverse, task-agnostic datasets, modern machine learning models can solve specific downstream tasks either zero-shot or with small task-specific datasets to a high level of performance. While this capability has been demonstrated in other fields such as computer vision, natural language processing or speech recognition, it remains to be shown in robotics, where the generalization capabilities of the models are particularly critical due to the difficulty of collecting real-world robotic data. We argue that one of the keys to the success of such general robotic models lies with open-ended task-agnostic training, combined with high-capacity architectures that can absorb all of the diverse, robotic data. In this paper, we present a model class, dubbed Robotics Transformer, that exhibits promising scalable model properties. We verify our conclusions in a study of different model classes and their ability to generalize as a function of the data size, model size, and data diversity based on a large-scale data collection on real robots performing real-world tasks. The project's website and videos can be found at robotics-transformer1.github.io

Michael S. Ryoo, Keerthana Gopalakrishnan, Kumara Kahatapitiya et al.

We propose Token Turing Machines (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which summarise the previous history (i.e., frames). This memory is efficiently addressed, read and written using a Transformer as the processing unit/controller at each step. The model's memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efficiently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential visual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token_turing

Srijan Das, Michael S. Ryoo

In this report, we introduce our adaptation of image-text models for long-term action anticipation. Our Video + CLIP framework makes use of a large-scale pre-trained paired image-text model: CLIP and a video encoder Slowfast network. The CLIP embedding provides fine-grained understanding of objects relevant for an action whereas the slowfast network is responsible for modeling temporal information within a video clip of few frames. We show that the features obtained from both encoders are complementary to each other, thus outperforming the baseline on Ego4D for the task of long-term action anticipation. Our code is available at github.com/srijandas07/clip_baseline_LTA_Ego4d.

Jinghuan Shang, Srijan Das, Michael S. Ryoo

Humans are remarkably flexible in understanding viewpoint changes due to visual cortex supporting the perception of 3D structure. In contrast, most of the computer vision models that learn visual representation from a pool of 2D images often fail to generalize over novel camera viewpoints. Recently, the vision architectures have shifted towards convolution-free architectures, visual Transformers, which operate on tokens derived from image patches. However, these Transformers do not perform explicit operations to learn viewpoint-agnostic representation for visual understanding. To this end, we propose a 3D Token Representation Layer (3DTRL) that estimates the 3D positional information of the visual tokens and leverages it for learning viewpoint-agnostic representations. The key elements of 3DTRL include a pseudo-depth estimator and a learned camera matrix to impose geometric transformations on the tokens, trained in an unsupervised fashion. These enable 3DTRL to recover the 3D positional information of the tokens from 2D patches. In practice, 3DTRL is easily plugged-in into a Transformer. Our experiments demonstrate the effectiveness of 3DTRL in many vision tasks including image classification, multi-view video alignment, and action recognition. The models with 3DTRL outperform their backbone Transformers in all the tasks with minimal added computation. Our code is available at https://github.com/elicassion/3DTRL.

Can Qin, Congying Xia, Krithika Ramakrishnan et al. · salesforce, stanford

We present xGen-VideoSyn-1, a text-to-video (T2V) generation model capable of producing realistic scenes from textual descriptions. Building on recent advancements, such as OpenAI's Sora, we explore the latent diffusion model (LDM) architecture and introduce a video variational autoencoder (VidVAE). VidVAE compresses video data both spatially and temporally, significantly reducing the length of visual tokens and the computational demands associated with generating long-sequence videos. To further address the computational costs, we propose a divide-and-merge strategy that maintains temporal consistency across video segments. Our Diffusion Transformer (DiT) model incorporates spatial and temporal self-attention layers, enabling robust generalization across different timeframes and aspect ratios. We have devised a data processing pipeline from the very beginning and collected over 13M high-quality video-text pairs. The pipeline includes multiple steps such as clipping, text detection, motion estimation, aesthetics scoring, and dense captioning based on our in-house video-LLM model. Training the VidVAE and DiT models required approximately 40 and 642 H100 days, respectively. Our model supports over 14-second 720p video generation in an end-to-end way and demonstrates competitive performance against state-of-the-art T2V models.

Andy Zeng, Maria Attarian, Brian Ichter et al.

Large pretrained (e.g., "foundation") models exhibit distinct capabilities depending on the domain of data they are trained on. While these domains are generic, they may only barely overlap. For example, visual-language models (VLMs) are trained on Internet-scale image captions, but large language models (LMs) are further trained on Internet-scale text with no images (e.g., spreadsheets, SAT questions, code). As a result, these models store different forms of commonsense knowledge across different domains. In this work, we show that this diversity is symbiotic, and can be leveraged through Socratic Models (SMs): a modular framework in which multiple pretrained models may be composed zero-shot i.e., via multimodal-informed prompting, to exchange information with each other and capture new multimodal capabilities, without requiring finetuning. With minimal engineering, SMs are not only competitive with state-of-the-art zero-shot image captioning and video-to-text retrieval, but also enable new applications such as (i) answering free-form questions about egocentric video, (ii) engaging in multimodal assistive dialogue with people (e.g., for cooking recipes) by interfacing with external APIs and databases (e.g., web search), and (iii) robot perception and planning.

Yu Fang, Kanchana Ranasinghe, Le Xue et al. · salesforce, stanford

Vision-Language-Action (VLA) models have achieved remarkable progress in robotic manipulation by mapping multimodal observations and instructions directly to actions. However, they typically mimic expert trajectories without predictive motion reasoning, which limits their ability to reason about what actions to take. To address this limitation, we propose joint learning with motion image diffusion, a novel strategy that enhances VLA models with motion reasoning capabilities. Our method extends the VLA architecture with a dual-head design: while the action head predicts action chunks as in vanilla VLAs, an additional motion head, implemented as a Diffusion Transformer (DiT), predicts optical-flow-based motion images that capture future dynamics. The two heads are trained jointly, enabling the shared VLM backbone to learn representations that couple robot control with motion knowledge. This joint learning builds temporally coherent and physically grounded representations without modifying the inference pathway of standard VLAs, thereby maintaining test-time latency. Experiments in both simulation and real-world environments demonstrate that joint learning with motion image diffusion improves the success rate of pi-series VLAs to 97.5% on the LIBERO benchmark and 58.0% on the RoboTwin benchmark, yielding a 23% improvement in real-world performance and validating its effectiveness in enhancing the motion reasoning capability of large-scale VLAs.

Xiang Li, Varun Belagali, Jinghuan Shang et al.

Sequence modeling approaches have shown promising results in robot imitation learning. Recently, diffusion models have been adopted for behavioral cloning in a sequence modeling fashion, benefiting from their exceptional capabilities in modeling complex data distributions. The standard diffusion-based policy iteratively generates action sequences from random noise conditioned on the input states. Nonetheless, the model for diffusion policy can be further improved in terms of visual representations. In this work, we propose Crossway Diffusion, a simple yet effective method to enhance diffusion-based visuomotor policy learning via a carefully designed state decoder and an auxiliary self-supervised learning (SSL) objective. The state decoder reconstructs raw image pixels and other state information from the intermediate representations of the reverse diffusion process. The whole model is jointly optimized by the SSL objective and the original diffusion loss. Our experiments demonstrate the effectiveness of Crossway Diffusion in various simulated and real-world robot tasks, confirming its consistent advantages over the standard diffusion-based policy and substantial improvements over the baselines.

Srijan Das, Tanmay Jain, Dominick Reilly et al.

Vision Transformers (ViTs) have become ubiquitous in computer vision. Despite their success, ViTs lack inductive biases, which can make it difficult to train them with limited data. To address this challenge, prior studies suggest training ViTs with self-supervised learning (SSL) and fine-tuning sequentially. However, we observe that jointly optimizing ViTs for the primary task and a Self-Supervised Auxiliary Task (SSAT) is surprisingly beneficial when the amount of training data is limited. We explore the appropriate SSL tasks that can be optimized alongside the primary task, the training schemes for these tasks, and the data scale at which they can be most effective. Our findings reveal that SSAT is a powerful technique that enables ViTs to leverage the unique characteristics of both the self-supervised and primary tasks, achieving better performance than typical ViTs pre-training with SSL and fine-tuning sequentially. Our experiments, conducted on 10 datasets, demonstrate that SSAT significantly improves ViT performance while reducing carbon footprint. We also confirm the effectiveness of SSAT in the video domain for deepfake detection, showcasing its generalizability. Our code is available at https://github.com/dominickrei/Limited-data-vits.

Ryan Burgert, Kanchana Ranasinghe, Xiang Li et al.

Recently, text-to-image diffusion models have shown remarkable capabilities in creating realistic images from natural language prompts. However, few works have explored using these models for semantic localization or grounding. In this work, we explore how an off-the-shelf text-to-image diffusion model, trained without exposure to localization information, can ground various semantic phrases without segmentation-specific re-training. We introduce an inference time optimization process capable of generating segmentation masks conditioned on natural language prompts. Our proposal, Peekaboo, is a first-of-its-kind zero-shot, open-vocabulary, unsupervised semantic grounding technique leveraging diffusion models without any training. We evaluate Peekaboo on the Pascal VOC dataset for unsupervised semantic segmentation and the RefCOCO dataset for referring segmentation, showing results competitive with promising results. We also demonstrate how Peekaboo can be used to generate images with transparency, even though the underlying diffusion model was only trained on RGB images - which to our knowledge we are the first to attempt. Please see our project page, including our code: https://ryanndagreat.github.io/peekaboo

Kumara Kahatapitiya, Anurag Arnab, Arsha Nagrani et al.

Vision-Language models (VLMs) have excelled in the image-domain -- especially in zero-shot settings -- thanks to the availability of vast pretraining data (i.e., paired image-text samples). However for videos, such paired data is not as abundant. Therefore, video-VLMs are usually designed by adapting pretrained image-VLMs to the video-domain, instead of training from scratch. All such recipes rely on augmenting visual embeddings with temporal information (i.e., image $\rightarrow$ video), often keeping text embeddings unchanged or even being discarded. In this paper, we argue the contrary, that better video-VLMs can be designed by focusing more on augmenting text, rather than visual information. More specifically, we introduce Video-conditioned Text Representations (VicTR): a form of text embeddings optimized w.r.t. visual embeddings, creating a more-flexible contrastive latent space. Our model can further make use of freely-available semantic information, in the form of visually-grounded auxiliary text (e.g. object or scene information). We evaluate our model on few-shot, zero-shot (HMDB-51, UCF-101), short-form (Kinetics-400) and long-form (Charades) activity recognition benchmarks, showing strong performance among video-VLMs.

Xiang Li, Jinghuan Shang, Srijan Das et al.

We investigate whether self-supervised learning (SSL) can improve online reinforcement learning (RL) from pixels. We extend the contrastive reinforcement learning framework (e.g., CURL) that jointly optimizes SSL and RL losses and conduct an extensive amount of experiments with various self-supervised losses. Our observations suggest that the existing SSL framework for RL fails to bring meaningful improvement over the baselines only taking advantage of image augmentation when the same amount of data and augmentation is used. We further perform evolutionary searches to find the optimal combination of multiple self-supervised losses for RL, but find that even such a loss combination fails to meaningfully outperform the methods that only utilize carefully designed image augmentations. After evaluating these approaches together in multiple different environments including a real-world robot environment, we confirm that no single self-supervised loss or image augmentation method can dominate all environments and that the current framework for joint optimization of SSL and RL is limited. Finally, we conduct the ablation study on multiple factors and demonstrate the properties of representations learned with different approaches.

AJ Piergiovanni, Isaac Noble, Dahun Kim et al.

One of the main challenges of multimodal learning is the need to combine heterogeneous modalities (e.g., video, audio, text). For example, video and audio are obtained at much higher rates than text and are roughly aligned in time. They are often not synchronized with text, which comes as a global context, e.g., a title, or a description. Furthermore, video and audio inputs are of much larger volumes, and grow as the video length increases, which naturally requires more compute dedicated to these modalities and makes modeling of long-range dependencies harder. We here decouple the multimodal modeling, dividing it into separate, focused autoregressive models, processing the inputs according to the characteristics of the modalities. We propose a multimodal model, called Mirasol3B, consisting of an autoregressive component for the time-synchronized modalities (audio and video), and an autoregressive component for the context modalities which are not necessarily aligned in time but are still sequential. To address the long-sequences of the video-audio inputs, we propose to further partition the video and audio sequences in consecutive snippets and autoregressively process their representations. To that end, we propose a Combiner mechanism, which models the audio-video information jointly within a timeframe. The Combiner learns to extract audio and video features from raw spatio-temporal signals, and then learns to fuse these features producing compact but expressive representations per snippet. Our approach achieves the state-of-the-art on well established multimodal benchmarks, outperforming much larger models. It effectively addresses the high computational demand of media inputs by both learning compact representations, controlling the sequence length of the audio-video feature representations, and modeling their dependencies in time.

AJ Piergiovanni, Kairo Morton, Weicheng Kuo et al.

Video question answering is a challenging task that requires understanding jointly the language input, the visual information in individual video frames, as well as the temporal information about the events occurring in the video. In this paper, we propose a novel multi-stream video encoder for video question answering that uses multiple video inputs and a new video-text iterative co-tokenization approach to answer a variety of questions related to videos. We experimentally evaluate the model on several datasets, such as MSRVTT-QA, MSVD-QA, IVQA, outperforming the previous state-of-the-art by large margins. Simultaneously, our model reduces the required GFLOPs from 150-360 to only 67, producing a highly efficient video question answering model.

Kanchana Ranasinghe, Michael Ryoo

Recent contrastive language image pre-training has led to learning highly transferable and robust image representations. However, adapting these models to video domains with minimal supervision remains an open problem. We explore a simple step in that direction, using language tied self-supervised learning to adapt an image CLIP model to the video domain. A backbone modified for temporal modeling is trained under self-distillation settings with train objectives operating in an action concept space. Feature vectors of various action concepts extracted from a language encoder using relevant textual prompts construct this space. We introduce two train objectives, concept distillation and concept alignment, that retain generality of original representations while enforcing relations between actions and their attributes. Our approach improves zero-shot and linear probing performance on three action recognition benchmarks.

Ryan Burgert, Jinghuan Shang, Xiang Li et al.

Unpaired image translation algorithms can be used for sim2real tasks, but many fail to generate temporally consistent results. We present a new approach that combines differentiable rendering with image translation to achieve temporal consistency over indefinite timescales, using surface consistency losses and \emph{neural neural textures}. We call this algorithm TRITON (Texture Recovering Image Translation Network): an unsupervised, end-to-end, stateless sim2real algorithm that leverages the underlying 3D geometry of input scenes by generating realistic-looking learnable neural textures. By settling on a particular texture for the objects in a scene, we ensure consistency between frames statelessly. Unlike previous algorithms, TRITON is not limited to camera movements -- it can handle the movement of objects as well, making it useful for downstream tasks such as robotic manipulation.

Jongwoo Park, Kumara Kahatapitiya, Donghyun Kim et al.

Vision Transformers (ViTs) have recently become the state-of-the-art across many computer vision tasks. In contrast to convolutional networks (CNNs), ViTs enable global information sharing even within shallow layers of a network, i.e., among high-resolution features. However, this perk was later overlooked with the success of pyramid architectures such as Swin Transformer, which show better performance-complexity trade-offs. In this paper, we present a simple and efficient add-on component (termed GrafT) that considers global dependencies and multi-scale information throughout the network, in both high- and low-resolution features alike. It has the flexibility of branching out at arbitrary depths and shares most of the parameters and computations of the backbone. GrafT shows consistent gains over various well-known models which includes both hybrid and pure Transformer types, both homogeneous and pyramid structures, and various self-attention methods. In particular, it largely benefits mobile-size models by providing high-level semantics. On the ImageNet-1k dataset, GrafT delivers +3.9%, +1.4%, and +1.9% top-1 accuracy improvement to DeiT-T, Swin-T, and MobileViT-XXS, respectively. Our code and models will be made available.

Alan Wu, Michael S. Ryoo

We present a novel framework using Energy-Based Models (EBMs) for localizing a ground vehicle mounted with a range sensor against satellite imagery in the absence of GPS. Lidar sensors have become ubiquitous on autonomous vehicles for describing its surrounding environment. Map priors are typically built using the same sensor modality for localization purposes. However, these map building endeavors using range sensors are often expensive and time-consuming. Alternatively, we leverage the use of satellite images as map priors, which are widely available, easily accessible, and provide comprehensive coverage. We propose a method using convolutional transformers that performs accurate metric-level localization in a cross-modal manner, which is challenging due to the drastic difference in appearance between the sparse range sensor readings and the rich satellite imagery. We train our model end-to-end and demonstrate our approach achieving higher accuracy than the state-of-the-art on KITTI, Pandaset, and a custom dataset.

Yoo Sung Jang, Kanchana Ranasinghe, Cristina Mata et al.

Cross-embodiment learning from human demonstrations is hindered by the visual gap between human and robot embodiments. While self-supervised learning (SSL) backbones encode rich inter-class semantics of general objects, we show they fail to establish correspondence between human and robot hands. We propose LACE, a framework that aligns human and robot visual representations in the latent space of these backbones by leveraging correspondences between shared body parts across embodiments as sparse supervision. These annotations can be automatically obtained via forward kinematics, and single robot demonstration is sufficient to train the model. Our semantic alignment loss matches distributions incurred by corresponding features, lifting patch-level supervision to semantic-level alignment, while a Gram loss preserves pretrained feature quality. This alignment enables robot policies to leverage abundant human data when robot demonstrations are scarce: in zero-shot transfer, policies using LACE-DINO outperform those using DINO by a large margin (65\%), with consistent gains in low-data regimes and out-of-distribution environments.

Jinghuan Shang, Michael S. Ryoo

In this work, we investigate Active Vision Reinforcement Learning (ActiveVision-RL), where an embodied agent simultaneously learns action policy for the task while also controlling its visual observations in partially observable environments. We denote the former as motor policy and the latter as sensory policy. For example, humans solve real world tasks by hand manipulation (motor policy) together with eye movements (sensory policy). ActiveVision-RL poses challenges on coordinating two policies given their mutual influence. We propose SUGARL, Sensorimotor Understanding Guided Active Reinforcement Learning, a framework that models motor and sensory policies separately, but jointly learns them using with an intrinsic sensorimotor reward. This learnable reward is assigned by sensorimotor reward module, incentivizes the sensory policy to select observations that are optimal to infer its own motor action, inspired by the sensorimotor stage of humans. Through a series of experiments, we show the effectiveness of our method across a range of observability conditions and its adaptability to existed RL algorithms. The sensory policies learned through our method are observed to exhibit effective active vision strategies.

Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park et al.

Language has become a prominent modality in computer vision with the rise of LLMs. Despite supporting long context-lengths, their effectiveness in handling long-term information gradually declines with input length. This becomes critical, especially in applications such as long-form video understanding. In this paper, we introduce a Language Repository (LangRepo) for LLMs, that maintains concise and structured information as an interpretable (i.e., all-textual) representation. Our repository is updated iteratively based on multi-scale video chunks. We introduce write and read operations that focus on pruning redundancies in text, and extracting information at various temporal scales. The proposed framework is evaluated on zero-shot visual question-answering benchmarks including EgoSchema, NExT-QA, IntentQA and NExT-GQA, showing state-of-the-art performance at its scale. Our code is available at https://github.com/kkahatapitiya/LangRepo.

Ryan Burgert, Yuancheng Xu, Wenqi Xian et al.

Generative modeling aims to transform random noise into structured outputs. In this work, we enhance video diffusion models by allowing motion control via structured latent noise sampling. This is achieved by just a change in data: we pre-process training videos to yield structured noise. Consequently, our method is agnostic to diffusion model design, requiring no changes to model architectures or training pipelines. Specifically, we propose a novel noise warping algorithm, fast enough to run in real time, that replaces random temporal Gaussianity with correlated warped noise derived from optical flow fields, while preserving the spatial Gaussianity. The efficiency of our algorithm enables us to fine-tune modern video diffusion base models using warped noise with minimal overhead, and provide a one-stop solution for a wide range of user-friendly motion control: local object motion control, global camera movement control, and motion transfer. The harmonization between temporal coherence and spatial Gaussianity in our warped noise leads to effective motion control while maintaining per-frame pixel quality. Extensive experiments and user studies demonstrate the advantages of our method, making it a robust and scalable approach for controlling motion in video diffusion models. Video results are available on our webpage: https://eyeline-labs.github.io/Go-with-the-Flow. Source code and model checkpoints are available on GitHub: https://github.com/Eyeline-Labs/Go-with-the-Flow.

Michael S. Ryoo, Honglu Zhou, Shrikant Kendre et al. · salesforce, stanford

We present xGen-MM-Vid (BLIP-3-Video): a multimodal language model for videos, particularly designed to efficiently capture temporal information over multiple frames. BLIP-3-Video takes advantage of the 'temporal encoder' in addition to the conventional visual tokenizer, which maps a sequence of tokens over multiple frames into a compact set of visual tokens. This enables BLIP3-Video to use much fewer visual tokens than its competing models (e.g., 32 vs. 4608 tokens). We explore different types of temporal encoders, including learnable spatio-temporal pooling as well as sequential models like Token Turing Machines. We experimentally confirm that BLIP-3-Video obtains video question-answering accuracies comparable to much larger state-of-the-art models (e.g., 34B), while being much smaller (i.e., 4B) and more efficient by using fewer visual tokens. The project website is at https://www.salesforceairesearch.com/opensource/xGen-MM-Vid/index.html

Ryan Burgert, Xiang Li, Abe Leite et al.

We explore the problem of computationally generating special `prime' images that produce optical illusions when physically arranged and viewed in a certain way. First, we propose a formal definition for this problem. Next, we introduce Diffusion Illusions, the first comprehensive pipeline designed to automatically generate a wide range of these illusions. Specifically, we both adapt the existing `score distillation loss' and propose a new `dream target loss' to optimize a group of differentially parametrized prime images, using a frozen text-to-image diffusion model. We study three types of illusions, each where the prime images are arranged in different ways and optimized using the aforementioned losses such that images derived from them align with user-chosen text prompts or images. We conduct comprehensive experiments on these illusions and verify the effectiveness of our proposed method qualitatively and quantitatively. Additionally, we showcase the successful physical fabrication of our illusions -- as they are all designed to work in the real world. Our code and examples are publicly available at our interactive project website: https://diffusionillusions.com

Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya et al.

Large Language Models (LLMs) have allowed recent LLM-based approaches to achieve excellent performance on long-video understanding benchmarks. We investigate how extensive world knowledge and strong reasoning skills of underlying LLMs influence this strong performance. Surprisingly, we discover that LLM-based approaches can yield surprisingly good accuracy on long-video tasks with limited video information, sometimes even with no video specific information. Building on this, we explore injecting video-specific information into an LLM-based framework. We utilize off-the-shelf vision tools to extract three object-centric information modalities from videos, and then leverage natural language as a medium for fusing this information. Our resulting Multimodal Video Understanding (MVU) framework demonstrates state-of-the-art performance across multiple video understanding benchmarks. Strong performance also on robotics domain tasks establish its strong generality. Code: https://github.com/kahnchana/mvu

Cristina Mata, Kanchana Ranasinghe, Michael S. Ryoo

Unsupervised domain adaptation (UDA) involves learning class semantics from labeled data within a source domain that generalize to an unseen target domain. UDA methods are particularly impactful for semantic segmentation, where annotations are more difficult to collect than in image classification. Despite recent advances in large-scale vision-language representation learning, UDA methods for segmentation have not taken advantage of the domain-agnostic properties of text. To address this, we present a novel Covariance-based Pixel-Text loss, CoPT, that uses domain-agnostic text embeddings to learn domain-invariant features in an image segmentation encoder. The text embeddings are generated through our LLM Domain Template process, where an LLM is used to generate source and target domain descriptions that are fed to a frozen CLIP model and combined. In experiments on four benchmarks we show that a model trained using CoPT achieves the new state of the art performance on UDA for segmentation. The code can be found at https://github.com/cfmata/CoPT.

Xiang Li, Cristina Mata, Jongwoo Park et al.

Vision Language Models (VLMs) have recently been leveraged to generate robotic actions, forming Vision-Language-Action (VLA) models. However, directly adapting a pretrained VLM for robotic control remains challenging, particularly when constrained by a limited number of robot demonstrations. In this work, we introduce LLaRA: Large Language and Robotics Assistant, a framework that formulates robot action policy as visuo-textual conversations and enables an efficient transfer of a pretrained VLM into a powerful VLA, motivated by the success of visual instruction tuning in Computer Vision. First, we present an automated pipeline to generate conversation-style instruction tuning data for robots from existing behavior cloning datasets, aligning robotic actions with image pixel coordinates. Further, we enhance this dataset in a self-supervised manner by defining six auxiliary tasks, without requiring any additional action annotations. We show that a VLM finetuned with a limited amount of such datasets can produce meaningful action decisions for robotic control. Through experiments across multiple simulated and real-world tasks, we demonstrate that LLaRA achieves state-of-the-art performance while preserving the generalization capabilities of large language models. The code, datasets, and pretrained models are available at https://github.com/LostXine/LLaRA.

Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya et al.

Long-form videos that span across wide temporal intervals are highly information redundant and contain multiple distinct events or entities that are often loosely related. Therefore, when performing long-form video question answering (LVQA), all information necessary to generate a correct response can often be contained within a small subset of frames. Recent literature leverage large language models (LLMs) in LVQA benchmarks, achieving exceptional performance, while relying on vision language models (VLMs) to convert all visual content within videos into natural language. Such VLMs often independently caption a large number of frames uniformly sampled from long videos, which is not efficient and can mostly be redundant. Motivated by this inefficiency, we propose LVNet, a modular and training-free framework featuring a novel Hierarchical Keyframe Selector (HKS) that efficiently selects a minimal set of informative frames tailored to each question. LVNet's modularity allows easy integration with existing approaches for more efficient LVQA. We achieve state-of-the-art performance among similarly configured models across four benchmark LVQA datasets: EgoSchema, NExT-QA, IntentQA, VideoMME. The code can be found at https://github.com/jongwoopark7978/LVNet

Kumara Kahatapitiya, Michael S. Ryoo

Modeling visual data as tokens (i.e., image patches) using attention mechanisms, feed-forward networks or convolutions has been highly effective in recent years. Such methods usually have a common pipeline: a tokenization method, followed by a set of layers/blocks for information mixing, both within and among tokens. When image patches are converted into tokens, they are often flattened, discarding the spatial structure within each patch. As a result, any processing that follows (eg: multi-head self-attention) may fail to recover and/or benefit from such information. In this paper, we argue that models can have significant gains when spatial structure is preserved during tokenization, and is explicitly used during the mixing stage. We propose two key contributions: (1) Structure-aware Tokenization and, (2) Structure-aware Mixing, both of which can be combined with existing models with minimal effort. We introduce a family of models (SWAT), showing improvements over the likes of DeiT, MLP-Mixer and Swin Transformer, across multiple benchmarks including ImageNet classification and ADE20K segmentation. Our code is available at https://github.com/kkahatapitiya/SWAT.

Kumara Kahatapitiya, Zhou Ren, Haoxiang Li et al.

Temporal Activity Detection aims to predict activity classes per frame, in contrast to video-level predictions in Activity Classification (i.e., Activity Recognition). Due to the expensive frame-level annotations required for detection, the scale of detection datasets is limited. Thus, commonly, previous work on temporal activity detection resorts to fine-tuning a classification model pretrained on large-scale classification datasets (e.g., Kinetics-400). However, such pretrained models are not ideal for downstream detection, due to the disparity between the pretraining and the downstream fine-tuning tasks. In this work, we propose a novel 'weakly-guided self-supervised' pretraining method for detection. We leverage weak labels (classification) to introduce a self-supervised pretext task (detection) by generating frame-level pseudo labels, multi-action frames, and action segments. Simply put, we design a detection task similar to downstream, on large-scale classification data, without extra annotations. We show that the models pretrained with the proposed weakly-guided self-supervised detection task outperform prior work on multiple challenging activity detection benchmarks, including Charades and MultiTHUMOS. Our extensive ablations further provide insights on when and how to use the proposed models for activity detection. Code is available at https://github.com/kkahatapitiya/SSDet.

Jinghuan Shang, Kumara Kahatapitiya, Xiang Li et al.

Reinforcement Learning (RL) can be considered as a sequence modeling task: given a sequence of past state-action-reward experiences, an agent predicts a sequence of next actions. In this work, we propose State-Action-Reward Transformer (StARformer) for visual RL, which explicitly models short-term state-action-reward representations (StAR-representations), essentially introducing a Markovian-like inductive bias to improve long-term modeling. Our approach first extracts StAR-representations by self-attending image state patches, action, and reward tokens within a short temporal window. These are then combined with pure image state representations -- extracted as convolutional features, to perform self-attention over the whole sequence. Our experiments show that StARformer outperforms the state-of-the-art Transformer-based method on image-based Atari and DeepMind Control Suite benchmarks, in both offline-RL and imitation learning settings. StARformer is also more compliant with longer sequences of inputs. Our code is available at https://github.com/elicassion/StARformer.

AJ Piergiovanni, Anelia Angelova, Michael S. Ryoo et al.

In this paper we address the problem of automatically discovering atomic actions in unsupervised manner from instructional videos. Instructional videos contain complex activities and are a rich source of information for intelligent agents, such as, autonomous robots or virtual assistants, which can, for example, automatically `read' the steps from an instructional video and execute them. However, videos are rarely annotated with atomic activities, their boundaries or duration. We present an unsupervised approach to learn atomic actions of structured human tasks from a variety of instructional videos. We propose a sequential stochastic autoregressive model for temporal segmentation of videos, which learns to represent and discover the sequential relationship between different atomic actions of the task, and which provides automatic and unsupervised self-labeling for videos. Our approach outperforms the state-of-the-art unsupervised methods with large margins. We will open source the code.

Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab et al.

In this paper, we introduce a novel visual representation learning which relies on a handful of adaptively learned tokens, and which is applicable to both image and video understanding tasks. Instead of relying on hand-designed splitting strategies to obtain visual tokens and processing a large number of densely sampled patches for attention, our approach learns to mine important tokens in visual data. This results in efficiently and effectively finding a few important visual tokens and enables modeling of pairwise attention between such tokens, over a longer temporal horizon for videos, or the spatial content in images. Our experiments demonstrate strong performance on several challenging benchmarks for both image and video recognition tasks. Importantly, due to our tokens being adaptive, we accomplish competitive results at significantly reduced compute amount. We obtain comparable results to the state-of-the-arts on ImageNet while being computationally more efficient. We also confirm the effectiveness of the approach on multiple video datasets, including Kinetics-400, Kinetics-600, Charades, and AViD. The code is available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token_learner

Kumara Kahatapitiya, Michael S. Ryoo

In this paper, we introduce Coarse-Fine Networks, a two-stream architecture which benefits from different abstractions of temporal resolution to learn better video representations for long-term motion. Traditional Video models process inputs at one (or few) fixed temporal resolution without any dynamic frame selection. However, we argue that, processing multiple temporal resolutions of the input and doing so dynamically by learning to estimate the importance of each frame can largely improve video representations, specially in the domain of temporal activity localization. To this end, we propose (1) Grid Pool, a learned temporal downsampling layer to extract coarse features, and, (2) Multi-stage Fusion, a spatio-temporal attention mechanism to fuse a fine-grained context with the coarse features. We show that our method outperforms the state-of-the-arts for action detection in public datasets including Charades with a significantly reduced compute and memory footprint. The code is available at https://github.com/kkahatapitiya/Coarse-Fine-Networks

AJ Piergiovanni, Anelia Angelova, Michael S. Ryoo

This paper proposes a novel algorithm which learns a formal regular grammar from real-world continuous data, such as videos. Learning latent terminals, non-terminals, and production rules directly from continuous data allows the construction of a generative model capturing sequential structures with multiple possibilities. Our model is fully differentiable, and provides easily interpretable results which are important in order to understand the learned structures. It outperforms the state-of-the-art on several challenging datasets and is more accurate for forecasting future activities in videos. We plan to open-source the code. https://sites.google.com/view/differentiable-grammars

AJ Piergiovanni, Anelia Angelova, Alexander Toshev et al.

We present a new method for finding video CNN architectures that capture rich spatio-temporal information in videos. Previous work, taking advantage of 3D convolutions, obtained promising results by manually designing video CNN architectures. We here develop a novel evolutionary search algorithm that automatically explores models with different types and combinations of layers to jointly learn interactions between spatial and temporal aspects of video representations. We demonstrate the generality of this algorithm by applying it to two meta-architectures, obtaining new architectures superior to manually designed architectures. Further, we propose a new component, the iTGM layer, which more efficiently utilizes its parameters to allow learning of space-time interactions over longer time horizons. The iTGM layer is often preferred by the evolutionary algorithm and allows building cost-efficient networks. The proposed approach discovers new and diverse video architectures that were previously unknown. More importantly they are both more accurate and faster than prior models, and outperform the state-of-the-art results on multiple datasets we test, including HMDB, Kinetics, and Moments in Time. We will open source the code and models, to encourage future model development.

Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed et al.

Integration of Large Language Models (LLMs) into visual domain tasks, resulting in visual-LLMs (V-LLMs), has enabled exceptional performance in vision-language tasks, particularly for visual question answering (VQA). However, existing V-LLMs (e.g. BLIP-2, LLaVA) demonstrate weak spatial reasoning and localization awareness. Despite generating highly descriptive and elaborate textual answers, these models fail at simple tasks like distinguishing a left vs right location. In this work, we explore how image-space coordinate based instruction fine-tuning objectives could inject spatial awareness into V-LLMs. We discover optimal coordinate representations, data-efficient instruction fine-tuning objectives, and pseudo-data generation strategies that lead to improved spatial awareness in V-LLMs. Additionally, our resulting model improves VQA across image and video domains, reduces undesired hallucination, and generates better contextual object descriptions. Experiments across 5 vision-language tasks involving 14 different datasets establish the clear performance improvements achieved by our proposed framework.

Kumara Kahatapitiya, Haozhe Liu, Sen He et al.

Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs) -- despite making significant headway in this context -- have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache), which is motivated by the fact that "not all videos are created equal": meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to 4.7x on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines.

Isabel Leal, Krzysztof Choromanski, Deepali Jain et al.

We present Self-Adaptive Robust Attention for Robotics Transformers (SARA-RT): a new paradigm for addressing the emerging challenge of scaling up Robotics Transformers (RT) for on-robot deployment. SARA-RT relies on the new method of fine-tuning proposed by us, called up-training. It converts pre-trained or already fine-tuned Transformer-based robotic policies of quadratic time complexity (including massive billion-parameter vision-language-action models or VLAs), into their efficient linear-attention counterparts maintaining high quality. We demonstrate the effectiveness of SARA-RT by speeding up: (a) the class of recently introduced RT-2 models, the first VLA robotic policies pre-trained on internet-scale data, as well as (b) Point Cloud Transformer (PCT) robotic policies operating on large point clouds. We complement our results with the rigorous mathematical analysis providing deeper insight into the phenomenon of SARA.

Kanchana Ranasinghe, Xiang Li, E-Ro Nguyen et al.

We present LangToMo, a vision-language-action framework structured as a dual-system architecture that uses pixel motion forecasts as intermediate representations. Our high-level System 2, an image diffusion model, generates text-conditioned pixel motion sequences from a single frame to guide robot control. Pixel motion-a universal, interpretable, and motion-centric representation-can be extracted from videos in a weakly-supervised manner, enabling diffusion model training on any video-caption data. Treating generated pixel motion as learned universal representations, our low level System 1 module translates these into robot actions via motion-to-action mapping functions, which can be either hand-crafted or learned with minimal supervision. System 2 operates as a high-level policy applied at sparse temporal intervals, while System 1 acts as a low-level policy at dense temporal intervals. This hierarchical decoupling enables flexible, scalable, and generalizable robot control under both unsupervised and supervised settings, bridging the gap between language, motion, and action. Checkout https://kahnchana.github.io/LangToMo

Jake Grigsby, Yuke Zhu, Michael Ryoo et al. · salesforce, stanford

Recent research looks to harness the general knowledge and reasoning of large language models (LLMs) into agents that accomplish user-specified goals in interactive environments. Vision-language models (VLMs) extend LLMs to multi-modal data and provide agents with the visual reasoning necessary for new applications in areas such as computer automation. However, agent tasks emphasize skills where accessible open-weight VLMs lag behind their LLM equivalents. For example, VLMs are less capable of following an environment's strict output syntax requirements and are more focused on open-ended question answering. Overcoming these limitations requires supervised fine-tuning (SFT) on task-specific expert demonstrations. Our work approaches these challenges from an offline-to-online reinforcement learning (RL) perspective. RL lets us fine-tune VLMs to agent tasks while learning from the unsuccessful decisions of our own model or more capable (larger) models. We explore an off-policy RL solution that retains the stability and simplicity of the widely used SFT workflow while allowing our agent to self-improve and learn from low-quality datasets. We demonstrate this technique with two open-weight VLMs across three multi-modal agent domains.

E-Ro Nguyen, Hieu Le, Dimitris Samaras et al.

Recent works on Generalized Referring Expression Segmentation (GRES) struggle with handling complex expressions referring to multiple distinct objects. This is because these methods typically employ an end-to-end foreground-background segmentation and lack a mechanism to explicitly differentiate and associate different object instances to the text query. To this end, we propose InstAlign, a method that incorporates object-level reasoning into the segmentation process. Our model leverages both text and image inputs to extract a set of object-level tokens that capture both the semantic information in the input prompt and the objects within the image. By modeling the text-object alignment via instance-level supervision, each token uniquely represents an object segment in the image, while also aligning with relevant semantic information from the text. Extensive experiments on the gRefCOCO and Ref-ZOM benchmarks demonstrate that our method significantly advances state-of-the-art performance, setting a new standard for precise and flexible GRES.

AJ Piergiovanni, Dahun Kim, Michael S. Ryoo et al.

Generating automatic dense captions for videos that accurately describe their contents remains a challenging area of research. Most current models require processing the entire video at once. Instead, we propose an efficient, online approach which outputs frequent, detailed and temporally aligned captions, without access to future frames. Our model uses a novel autoregressive factorized decoding architecture, which models the sequence of visual features for each time segment, outputting localized descriptions and efficiently leverages the context from the previous video segments. This allows the model to output frequent, detailed captions to more comprehensively describe the video, according to its actual local content, rather than mimic the training data. Second, we propose an optimization for efficient training and inference, which enables scaling to longer videos. Our approach shows excellent performance compared to both offline and online methods, and uses 20\% less compute. The annotations produced are much more comprehensive and frequent, and can further be utilized in automatic video tagging and in large-scale video data harvesting.

Ziyang Wang, Honglu Zhou, Shijie Wang et al.

Long video understanding (LVU) is challenging because answering real-world queries often depends on sparse, temporally dispersed cues buried in hours of mostly redundant and irrelevant content. While agentic pipelines improve video reasoning capabilities, prevailing frameworks rely on a query-agnostic captioner to perceive video information, which wastes computation on irrelevant content and blurs fine-grained temporal and spatial information. Motivated by active perception theory, we argue that LVU agents should actively decide what, when, and where to observe, and continuously assess whether the current observation is sufficient to answer the query. We present Active Video Perception (AVP), an evidence-seeking framework that treats the video as an interactive environment and acquires compact, queryrelevant evidence directly from pixels. Concretely, AVP runs an iterative plan-observe-reflect process with MLLM agents. In each round, a planner proposes targeted video interactions, an observer executes them to extract time-stamped evidence, and a reflector evaluates the sufficiency of the evidence for the query, either halting with an answer or triggering further observation. Across five LVU benchmarks, AVP achieves highest performance with significant improvements. Notably, AVP outperforms the best agentic method by 5.7% in average accuracy while only requires 18.4% inference time and 12.4% input tokens.

Shrikant Kendre, Austin Xu, Honglu Zhou et al.

Traditional evaluation metrics for textual and visual question answering, like ROUGE, METEOR, and Exact Match (EM), focus heavily on n-gram based lexical similarity, often missing the deeper semantic understanding needed for accurate assessment. While measures like BERTScore and MoverScore leverage contextual embeddings to address this limitation, they lack flexibility in balancing sentence-level and keyword-level semantics and ignore lexical similarity, which remains important. Large Language Model (LLM) based evaluators, though powerful, come with drawbacks like high costs, bias, inconsistency, and hallucinations. To address these issues, we introduce SMILE: Semantic Metric Integrating Lexical Exactness, a novel approach that combines sentence-level semantic understanding with keyword-level semantic understanding and easy keyword matching. This composite method balances lexical precision and semantic relevance, offering a comprehensive evaluation. Extensive benchmarks across text, image, and video QA tasks show SMILE is highly correlated with human judgments and computationally lightweight, bridging the gap between lexical and semantic evaluation.

Honglu Zhou, Xiangyu Peng, Shrikant Kendre et al. · salesforce, stanford

Next-generation AI companions must go beyond general video understanding to resolve spatial and temporal references in dynamic, real-world environments. Existing Video Large Language Models (Video LLMs), while capable of coarse-level comprehension, struggle with fine-grained, spatiotemporal reasoning, especially when user queries rely on time-based event references for temporal anchoring, or gestural cues for spatial anchoring to clarify object references and positions. To bridge this critical gap, we introduce Strefer, a synthetic instruction data generation framework designed to equip Video LLMs with spatiotemporal referring and reasoning capabilities. Strefer produces diverse instruction-tuning data using a data engine that pseudo-annotates temporally dense, fine-grained video metadata, capturing rich spatial and temporal information in a structured manner, including subjects, objects, their locations as masklets, and their action descriptions and timelines. Our approach enhances the ability of Video LLMs to interpret spatial and temporal references, fostering more versatile, space-time-aware reasoning essential for real-world AI companions. Without using proprietary models, costly human annotation, or the need to annotate large volumes of new videos, experimental evaluations show that models trained with data produced by Strefer outperform baselines on tasks requiring spatial and temporal disambiguation. Additionally, these models exhibit enhanced space-time-aware reasoning, establishing a new foundation for perceptually grounded, instruction-tuned Video LLMs.

Kanchana Ranasinghe, Sadeep Jayasumana, Andreas Veit et al.

Latent Diffusion Models (LDMs) produce high-quality, photo-realistic images, however, the latency incurred by multiple costly inference iterations can restrict their applicability. We introduce LatentCRF, a continuous Conditional Random Field (CRF) model, implemented as a neural network layer, that models the spatial and semantic relationships among the latent vectors in the LDM. By replacing some of the computationally-intensive LDM inference iterations with our lightweight LatentCRF, we achieve a superior balance between quality, speed and diversity. We increase inference efficiency by 33% with no loss in image quality or diversity compared to the full LDM. LatentCRF is an easy add-on, which does not require modifying the LDM.

Rui Dai, Srijan Das, Michael S. Ryoo et al.

The challenge of long-term video understanding remains constrained by the efficient extraction of object semantics and the modelling of their relationships for downstream tasks. Although the CLIP visual features exhibit discriminative properties for various vision tasks, particularly in object encoding, they are suboptimal for long-term video understanding. To address this issue, we present the Attributes-Aware Network (AAN), which consists of two key components: the Attributes Extractor and a Graph Reasoning block. These components facilitate the extraction of object-centric attributes and the modelling of their relationships within the video. By leveraging CLIP features, AAN outperforms state-of-the-art approaches on two popular action detection datasets: Charades and Toyota Smarthome Untrimmed datasets.