Shuran Song

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.

Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang et al. · allen-ai, stanford

Multimodal datasets are a critical component in recent breakthroughs such as Stable Diffusion and GPT-4, yet their design does not receive the same research attention as model architectures or training algorithms. To address this shortcoming in the ML ecosystem, we introduce DataComp, a testbed for dataset experiments centered around a new candidate pool of 12.8 billion image-text pairs from Common Crawl. Participants in our benchmark design new filtering techniques or curate new data sources and then evaluate their new dataset by running our standardized CLIP training code and testing the resulting model on 38 downstream test sets. Our benchmark consists of multiple compute scales spanning four orders of magnitude, which enables the study of scaling trends and makes the benchmark accessible to researchers with varying resources. Our baseline experiments show that the DataComp workflow leads to better training sets. In particular, our best baseline, DataComp-1B, enables training a CLIP ViT-L/14 from scratch to 79.2% zero-shot accuracy on ImageNet, outperforming OpenAI's CLIP ViT-L/14 by 3.7 percentage points while using the same training procedure and compute. We release DataComp and all accompanying code at www.datacomp.ai.

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 .

Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre et al.

Open-vocabulary models like CLIP achieve high accuracy across many image classification tasks. However, there are still settings where their zero-shot performance is far from optimal. We study model patching, where the goal is to improve accuracy on specific tasks without degrading accuracy on tasks where performance is already adequate. Towards this goal, we introduce PAINT, a patching method that uses interpolations between the weights of a model before fine-tuning and the weights after fine-tuning on a task to be patched. On nine tasks where zero-shot CLIP performs poorly, PAINT increases accuracy by 15 to 60 percentage points while preserving accuracy on ImageNet within one percentage point of the zero-shot model. PAINT also allows a single model to be patched on multiple tasks and improves with model scale. Furthermore, we identify cases of broad transfer, where patching on one task increases accuracy on other tasks even when the tasks have disjoint classes. Finally, we investigate applications beyond common benchmarks such as counting or reducing the impact of typographic attacks on CLIP. Our findings demonstrate that it is possible to expand the set of tasks on which open-vocabulary models achieve high accuracy without re-training them from scratch.

Isaac Kasahara, Shubham Agrawal, Selim Engin et al. · apple-ml, cmu

General scene reconstruction refers to the task of estimating the full 3D geometry and texture of a scene containing previously unseen objects. In many practical applications such as AR/VR, autonomous navigation, and robotics, only a single view of the scene may be available, making the scene reconstruction task challenging. In this paper, we present a method for scene reconstruction by structurally breaking the problem into two steps: rendering novel views via inpainting and 2D to 3D scene lifting. Specifically, we leverage the generalization capability of large visual language models (Dalle-2) to inpaint the missing areas of scene color images rendered from different views. Next, we lift these inpainted images to 3D by predicting normals of the inpainted image and solving for the missing depth values. By predicting for normals instead of depth directly, our method allows for robustness to changes in depth distributions and scale. With rigorous quantitative evaluation, we show that our method outperforms multiple baselines while providing generalization to novel objects and scenes.

Mengda Xu, Zhenjia Xu, Cheng Chi et al.

Human demonstration videos are a widely available data source for robot learning and an intuitive user interface for expressing desired behavior. However, directly extracting reusable robot manipulation skills from unstructured human videos is challenging due to the big embodiment difference and unobserved action parameters. To bridge this embodiment gap, this paper introduces XSkill, an imitation learning framework that 1) discovers a cross-embodiment representation called skill prototypes purely from unlabeled human and robot manipulation videos, 2) transfers the skill representation to robot actions using conditional diffusion policy, and finally, 3) composes the learned skill to accomplish unseen tasks specified by a human prompt video. Our experiments in simulation and real-world environments show that the discovered skill prototypes facilitate both skill transfer and composition for unseen tasks, resulting in a more general and scalable imitation learning framework. The benchmark, code, and qualitative results are on https://xskill.cs.columbia.edu/

Samir Yitzhak Gadre, Mitchell Wortsman, Gabriel Ilharco et al.

For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the Pasture benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 21 CoW baselines across Habitat, RoboTHOR, and Pasture. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions, but are proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration -- and no additional training -- matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on Habitat MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art RoboTHOR ZSON model.

Zhao Mandi, Shreeya Jain, Shuran Song

We propose a novel approach to multi-robot collaboration that harnesses the power of pre-trained large language models (LLMs) for both high-level communication and low-level path planning. Robots are equipped with LLMs to discuss and collectively reason task strategies. They then generate sub-task plans and task space waypoint paths, which are used by a multi-arm motion planner to accelerate trajectory planning. We also provide feedback from the environment, such as collision checking, and prompt the LLM agents to improve their plan and waypoints in-context. For evaluation, we introduce RoCoBench, a 6-task benchmark covering a wide range of multi-robot collaboration scenarios, accompanied by a text-only dataset for agent representation and reasoning. We experimentally demonstrate the effectiveness of our approach -- it achieves high success rates across all tasks in RoCoBench and adapts to variations in task semantics. Our dialog setup offers high interpretability and flexibility -- in real world experiments, we show RoCo easily incorporates human-in-the-loop, where a user can communicate and collaborate with a robot agent to complete tasks together. See project website https://project-roco.github.io for videos and code.

Zeyi Liu, Arpit Bahety, Shuran Song

The ability to detect and analyze failed executions automatically is crucial for an explainable and robust robotic system. Recently, Large Language Models (LLMs) have demonstrated strong reasoning abilities on textual inputs. To leverage the power of LLMs for robot failure explanation, we introduce REFLECT, a framework which queries LLM for failure reasoning based on a hierarchical summary of robot past experiences generated from multisensory observations. The failure explanation can further guide a language-based planner to correct the failure and complete the task. To systematically evaluate the framework, we create the RoboFail dataset with a variety of tasks and failure scenarios. We demonstrate that the LLM-based framework is able to generate informative failure explanations that assist successful correction planning.

Huy Ha, Shuran Song

We study open-world 3D scene understanding, a family of tasks that require agents to reason about their 3D environment with an open-set vocabulary and out-of-domain visual inputs - a critical skill for robots to operate in the unstructured 3D world. Towards this end, we propose Semantic Abstraction (SemAbs), a framework that equips 2D Vision-Language Models (VLMs) with new 3D spatial capabilities, while maintaining their zero-shot robustness. We achieve this abstraction using relevancy maps extracted from CLIP, and learn 3D spatial and geometric reasoning skills on top of those abstractions in a semantic-agnostic manner. We demonstrate the usefulness of SemAbs on two open-world 3D scene understanding tasks: 1) completing partially observed objects and 2) localizing hidden objects from language descriptions. Experiments show that SemAbs can generalize to novel vocabulary, materials/lighting, classes, and domains (i.e., real-world scans) from training on limited 3D synthetic data. Code and data is available at https://semantic-abstraction.cs.columbia.edu/

Jingyun Yang, Zi-ang Cao, Congyue Deng et al.

Building effective imitation learning methods that enable robots to learn from limited data and still generalize across diverse real-world environments is a long-standing problem in robot learning. We propose Equibot, a robust, data-efficient, and generalizable approach for robot manipulation task learning. Our approach combines SIM(3)-equivariant neural network architectures with diffusion models. This ensures that our learned policies are invariant to changes in scale, rotation, and translation, enhancing their applicability to unseen environments while retaining the benefits of diffusion-based policy learning such as multi-modality and robustness. We show on a suite of 6 simulation tasks that our proposed method reduces the data requirements and improves generalization to novel scenarios. In the real world, with 10 variations of 6 mobile manipulation tasks, we show that our method can easily generalize to novel objects and scenes after learning from just 5 minutes of human demonstrations in each task.

Zhao Mandi, Homanga Bharadhwaj, Vincent Moens et al.

Large-scale training have propelled significant progress in various sub-fields of AI such as computer vision and natural language processing. However, building robot learning systems at a comparable scale remains challenging. To develop robots that can perform a wide range of skills and adapt to new scenarios, efficient methods for collecting vast and diverse amounts of data on physical robot systems are required, as well as the capability to train high-capacity policies using such datasets. In this work, we propose a framework for scaling robot learning, with specific focus on multi-task and multi-scene manipulation in kitchen environments, both in simulation and in the real world. Our proposed framework, CACTI, comprises four stages that separately handle: data collection, data augmentation, visual representation learning, and imitation policy training, to enable scalability in robot learning . We make use of state-of-the-art generative models as part of the data augmentation stage, and use pre-trained out-of-domain visual representations to improve training efficiency. Experimental results demonstrate the effectiveness of our approach. On a real robot setup, CACTI enables efficient training of a single policy that can perform 10 manipulation tasks involving kitchen objects, and is robust to varying layouts of distractors. In a simulated kitchen environment, CACTI trains a single policy to perform 18 semantic tasks across 100 layout variations for each individual task. We will release the simulation task benchmark and augmented datasets in both real and simulated environments to facilitate future research.

Jimmy Wu, Xingyuan Sun, Andy Zeng et al.

We investigate pneumatic non-prehensile manipulation (i.e., blowing) as a means of efficiently moving scattered objects into a target receptacle. Due to the chaotic nature of aerodynamic forces, a blowing controller must (i) continually adapt to unexpected changes from its actions, (ii) maintain fine-grained control, since the slightest misstep can result in large unintended consequences (e.g., scatter objects already in a pile), and (iii) infer long-range plans (e.g., move the robot to strategic blowing locations). We tackle these challenges in the context of deep reinforcement learning, introducing a multi-frequency version of the spatial action maps framework. This allows for efficient learning of vision-based policies that effectively combine high-level planning and low-level closed-loop control for dynamic mobile manipulation. Experiments show that our system learns efficient behaviors for the task, demonstrating in particular that blowing achieves better downstream performance than pushing, and that our policies improve performance over baselines. Moreover, we show that our system naturally encourages emergent specialization between the different subpolicies spanning low-level fine-grained control and high-level planning. On a real mobile robot equipped with a miniature air blower, we show that our simulation-trained policies transfer well to a real environment and can generalize to novel objects.

Jiayi Chen, Mi Yan, Jiazhao Zhang et al.

In this work, we tackle the challenging task of jointly tracking hand object pose and reconstructing their shapes from depth point cloud sequences in the wild, given the initial poses at frame 0. We for the first time propose a point cloud based hand joint tracking network, HandTrackNet, to estimate the inter-frame hand joint motion. Our HandTrackNet proposes a novel hand pose canonicalization module to ease the tracking task, yielding accurate and robust hand joint tracking. Our pipeline then reconstructs the full hand via converting the predicted hand joints into a template-based parametric hand model MANO. For object tracking, we devise a simple yet effective module that estimates the object SDF from the first frame and performs optimization-based tracking. Finally, a joint optimization step is adopted to perform joint hand and object reasoning, which alleviates the occlusion-induced ambiguity and further refines the hand pose. During training, the whole pipeline only sees purely synthetic data, which are synthesized with sufficient variations and by depth simulation for the ease of generalization. The whole pipeline is pertinent to the generalization gaps and thus directly transferable to real in-the-wild data. We evaluate our method on two real hand object interaction datasets, e.g. HO3D and DexYCB, without any finetuning. Our experiments demonstrate that the proposed method significantly outperforms the previous state-of-the-art depth-based hand and object pose estimation and tracking methods, running at a frame rate of 9 FPS.

Jingxi Xu, Shuran Song, Matei Ciocarlie

Inspired by the human ability to perform complex manipulation in the complete absence of vision (like retrieving an object from a pocket), the robotic manipulation field is motivated to develop new methods for tactile-based object interaction. However, tactile sensing presents the challenge of being an active sensing modality: a touch sensor provides sparse, local data, and must be used in conjunction with effective exploration strategies in order to collect information. In this work, we focus on the process of guiding tactile exploration, and its interplay with task-related decision making. We propose TANDEM (TActile exploration aNd DEcision Making), an architecture to learn efficient exploration strategies in conjunction with decision making. Our approach is based on separate but co-trained modules for exploration and discrimination. We demonstrate this method on a tactile object recognition task, where a robot equipped with a touch sensor must explore and identify an object from a known set based on binary contact signals alone. TANDEM achieves higher accuracy with fewer actions than alternative methods and is also shown to be more robust to sensor noise.

Jingxi Xu, Han Lin, Shuran Song et al.

Tactile recognition of 3D objects remains a challenging task. Compared to 2D shapes, the complex geometry of 3D surfaces requires richer tactile signals, more dexterous actions, and more advanced encoding techniques. In this work, we propose TANDEM3D, a method that applies a co-training framework for exploration and decision making to 3D object recognition with tactile signals. Starting with our previous work, which introduced a co-training paradigm for 2D recognition problems, we introduce a number of advances that enable us to scale up to 3D. TANDEM3D is based on a novel encoder that builds 3D object representation from contact positions and normals using PointNet++. Furthermore, by enabling 6DOF movement, TANDEM3D explores and collects discriminative touch information with high efficiency. Our method is trained entirely in simulation and validated with real-world experiments. Compared to state-of-the-art baselines, TANDEM3D achieves higher accuracy and a lower number of actions in recognizing 3D objects and is also shown to be more robust to different types and amounts of sensor noise. Video is available at https://jxu.ai/tandem3d.

Siddharth Singi, Zhanpeng He, Alvin Pan et al.

In a Human-in-the-Loop paradigm, a robotic agent is able to act mostly autonomously in solving a task, but can request help from an external expert when needed. However, knowing when to request such assistance is critical: too few requests can lead to the robot making mistakes, but too many requests can overload the expert. In this paper, we present a Reinforcement Learning based approach to this problem, where a semi-autonomous agent asks for external assistance when it has low confidence in the eventual success of the task. The confidence level is computed by estimating the variance of the return from the current state. We show that this estimate can be iteratively improved during training using a Bellman-like recursion. On discrete navigation problems with both fully- and partially-observable state information, we show that our method makes effective use of a limited budget of expert calls at run-time, despite having no access to the expert at training time.

Yulong Li, Andy Zeng, Shuran Song

Most successes in autonomous robotic assembly have been restricted to single target or category. We propose to investigate general part assembly, the task of creating novel target assemblies with unseen part shapes. As a fundamental step to a general part assembly system, we tackle the task of determining the precise poses of the parts in the target assembly, which we we term ``rearrangement planning''. We present General Part Assembly Transformer (GPAT), a transformer-based model architecture that accurately predicts part poses by inferring how each part shape corresponds to the target shape. Our experiments on both 3D CAD models and real-world scans demonstrate GPAT's generalization abilities to novel and diverse target and part shapes.

Mengda Xu, Manuela Veloso, Shuran Song

We introduce ASPiRe (Adaptive Skill Prior for RL), a new approach that leverages prior experience to accelerate reinforcement learning. Unlike existing methods that learn a single skill prior from a large and diverse dataset, our framework learns a library of different distinction skill priors (i.e., behavior priors) from a collection of specialized datasets, and learns how to combine them to solve a new task. This formulation allows the algorithm to acquire a set of specialized skill priors that are more reusable for downstream tasks; however, it also brings up additional challenges of how to effectively combine these unstructured sets of skill priors to form a new prior for new tasks. Specifically, it requires the agent not only to identify which skill prior(s) to use but also how to combine them (either sequentially or concurrently) to form a new prior. To achieve this goal, ASPiRe includes Adaptive Weight Module (AWM) that learns to infer an adaptive weight assignment between different skill priors and uses them to guide policy learning for downstream tasks via weighted Kullback-Leibler divergences. Our experiments demonstrate that ASPiRe can significantly accelerate the learning of new downstream tasks in the presence of multiple priors and show improvement on competitive baselines.

Maximilian Du, Jennifer Grannen, Shuran Song et al.

The physicality of exercise makes the role of athletic trainers unique. Their physical presence allows them to guide a student through a motion, demonstrate an exercise, and give intuitive feedback. Robot quadrupeds are also embodied agents with robust agility and athleticism. In our work, we investigate whether a robot quadruped can serve as an effective and enjoyable personal trainer device. We focus on a case study of interval training for runners: a repetitive, long-horizon task where precision and consistency are important. To meet this challenge, we propose SNOOPIE, an autonomous robot quadruped pacer capable of running interval training exercises tailored to challenge a user's personal abilities. We conduct a set of user experiments that compare the robot trainer to a wearable trainer device--the Apple Watch--to investigate the benefits of a physical embodiment in exercise-based interactions. We demonstrate 60.6% better adherence to a pace schedule and were 45.9% more consistent across their running speeds with the quadruped trainer. Subjective results also showed that participants strongly preferred training with the robot over wearable devices across many qualitative axes, including its ease of use (+56.7%), enjoyability of the interaction (+60.6%), and helpfulness (+39.1%). Additional videos and visualizations can be found on our website: https://sites.google.com/view/snoopie

Mengda Xu, Zhenjia Xu, Yinghao Xu et al.

We present Im2Flow2Act, a scalable learning framework that enables robots to acquire real-world manipulation skills without the need of real-world robot training data. The key idea behind Im2Flow2Act is to use object flow as the manipulation interface, bridging domain gaps between different embodiments (i.e., human and robot) and training environments (i.e., real-world and simulated). Im2Flow2Act comprises two components: a flow generation network and a flow-conditioned policy. The flow generation network, trained on human demonstration videos, generates object flow from the initial scene image, conditioned on the task description. The flow-conditioned policy, trained on simulated robot play data, maps the generated object flow to robot actions to realize the desired object movements. By using flow as input, this policy can be directly deployed in the real world with a minimal sim-to-real gap. By leveraging real-world human videos and simulated robot play data, we bypass the challenges of teleoperating physical robots in the real world, resulting in a scalable system for diverse tasks. We demonstrate Im2Flow2Act's capabilities in a variety of real-world tasks, including the manipulation of rigid, articulated, and deformable objects.

Bardienus P. Duisterhof, Zhao Mandi, Yunchao Yao et al.

Teaching robots to fold, drape, or reposition deformable objects such as cloth will unlock a variety of automation applications. While remarkable progress has been made for rigid object manipulation, manipulating deformable objects poses unique challenges, including frequent occlusions, infinite-dimensional state spaces and complex dynamics. Just as object pose estimation and tracking have aided robots for rigid manipulation, dense 3D tracking (scene flow) of highly deformable objects will enable new applications in robotics while aiding existing approaches, such as imitation learning or creating digital twins with real2sim transfer. We propose DeformGS, an approach to recover scene flow in highly deformable scenes, using simultaneous video captures of a dynamic scene from multiple cameras. DeformGS builds on recent advances in Gaussian splatting, a method that learns the properties of a large number of Gaussians for state-of-the-art and fast novel-view synthesis. DeformGS learns a deformation function to project a set of Gaussians with canonical properties into world space. The deformation function uses a neural-voxel encoding and a multilayer perceptron (MLP) to infer Gaussian position, rotation, and a shadow scalar. We enforce physics-inspired regularization terms based on conservation of momentum and isometry, which leads to trajectories with smaller trajectory errors. We also leverage existing foundation models SAM and XMEM to produce noisy masks, and learn a per-Gaussian mask for better physics-inspired regularization. DeformGS achieves high-quality 3D tracking on highly deformable scenes with shadows and occlusions. In experiments, DeformGS improves 3D tracking by an average of 55.8% compared to the state-of-the-art. With sufficient texture, DeformGS achieves a median tracking error of 3.3 mm on a cloth of 1.5 x 1.5 m in area. Website: https://deformgs.github.io

Zeyi Liu, Zhenjia Xu, Shuran Song

We introduce BusyBoard, a toy-inspired robot learning environment that leverages a diverse set of articulated objects and inter-object functional relations to provide rich visual feedback for robot interactions. Based on this environment, we introduce a learning framework, BusyBot, which allows an agent to jointly acquire three fundamental capabilities (interaction, reasoning, and planning) in an integrated and self-supervised manner. With the rich sensory feedback provided by BusyBoard, BusyBot first learns a policy to efficiently interact with the environment; then with data collected using the policy, BusyBot reasons the inter-object functional relations through a causal discovery network; and finally by combining the learned interaction policy and relation reasoning skill, the agent is able to perform goal-conditioned manipulation tasks. We evaluate BusyBot in both simulated and real-world environments, and validate its generalizability to unseen objects and relations. Video is available at https://youtu.be/EJ98xBJZ9ek.

Zoey Chen, Zhao Mandi, Homanga Bharadhwaj et al.

Generalization to unseen real-world scenarios for robot manipulation requires exposure to diverse datasets during training. However, collecting large real-world datasets is intractable due to high operational costs. For robot learning to generalize despite these challenges, it is essential to leverage sources of data or priors beyond the robot's direct experience. In this work, we posit that image-text generative models, which are pre-trained on large corpora of web-scraped data, can serve as such a data source. These generative models encompass a broad range of real-world scenarios beyond a robot's direct experience and can synthesize novel synthetic experiences that expose robotic agents to additional world priors aiding real-world generalization at no extra cost. In particular, our approach leverages pre-trained generative models as an effective tool for data augmentation. We propose a generative augmentation framework for semantically controllable augmentations and rapidly multiplying robot datasets while inducing rich variations that enable real-world generalization. Based on diverse augmentations of robot data, we show how scalable robot manipulation policies can be trained and deployed both in simulation and in unseen real-world environments such as kitchens and table-tops. By demonstrating the effectiveness of image-text generative models in diverse real-world robotic applications, our generative augmentation framework provides a scalable and efficient path for boosting generalization in robot learning at no extra human cost.

Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani et al.

We introduce Structure from Action (SfA), a framework to discover 3D part geometry and joint parameters of unseen articulated objects via a sequence of inferred interactions. Our key insight is that 3D interaction and perception should be considered in conjunction to construct 3D articulated CAD models, especially for categories not seen during training. By selecting informative interactions, SfA discovers parts and reveals occluded surfaces, like the inside of a closed drawer. By aggregating visual observations in 3D, SfA accurately segments multiple parts, reconstructs part geometry, and infers all joint parameters in a canonical coordinate frame. Our experiments demonstrate that a SfA model trained in simulation can generalize to many unseen object categories with diverse structures and to real-world objects. Empirically, SfA outperforms a pipeline of state-of-the-art components by 25.4 3D IoU percentage points on unseen categories, while matching already performant joint estimation baselines.

Ryan Punamiya, Simar Kareer, Zeyi Liu et al.

Robot learning increasingly depends on large and diverse data, yet robot data collection remains expensive and difficult to scale. Egocentric human data offer a promising alternative by capturing rich manipulation behavior across everyday environments. However, existing human datasets are often limited in scope, difficult to extend, and fragmented across institutions. We introduce EgoVerse, a collaborative platform for human data-driven robot learning that unifies data collection, processing, and access under a shared framework, enabling contributions from individual researchers, academic labs, and industry partners. The current release includes 1,362 hours (80k episodes) of human demonstrations spanning 1,965 tasks, 240 scenes, and 2,087 unique demonstrators, with standardized formats, manipulation-relevant annotations, and tooling for downstream learning. Beyond the dataset, we conduct a large-scale study of human-to-robot transfer with experiments replicated across multiple labs, tasks, and robot embodiments under shared protocols. We find that policy performance generally improves with increased human data, but that effective scaling depends on alignment between human data and robot learning objectives. Together, the dataset, platform, and study establish a foundation for reproducible progress in human data-driven robot learning. Videos and additional information can be found at https://egoverse.ai/

Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin et al.

Recent video generation models demonstrate remarkable ability to capture complex physical interactions and scene evolution over time. To leverage their spatiotemporal priors, robotics works have adapted video models for policy learning but introduce complexity by requiring multiple stages of post-training and new architectural components for action generation. In this work, we introduce Cosmos Policy, a simple approach for adapting a large pretrained video model (Cosmos-Predict2) into an effective robot policy through a single stage of post-training on the robot demonstration data collected on the target platform, with no architectural modifications. Cosmos Policy learns to directly generate robot actions encoded as latent frames within the video model's latent diffusion process, harnessing the model's pretrained priors and core learning algorithm to capture complex action distributions. Additionally, Cosmos Policy generates future state images and values (expected cumulative rewards), which are similarly encoded as latent frames, enabling test-time planning of action trajectories with higher likelihood of success. In our evaluations, Cosmos Policy achieves state-of-the-art performance on the LIBERO and RoboCasa simulation benchmarks (98.5% and 67.1% average success rates, respectively) and the highest average score in challenging real-world bimanual manipulation tasks, outperforming strong diffusion policies trained from scratch, video model-based policies, and state-of-the-art vision-language-action models fine-tuned on the same robot demonstrations. Furthermore, given policy rollout data, Cosmos Policy can learn from experience to refine its world model and value function and leverage model-based planning to achieve even higher success rates in challenging tasks. We release code, models, and training data at https://research.nvidia.com/labs/dir/cosmos-policy/

Samir Yitzhak Gadre, Georgios Smyrnis, Vaishaal Shankar et al. · allen-ai, cmu

Scaling laws are useful guides for derisking expensive training runs, as they predict performance of large models using cheaper, small-scale experiments. However, there remain gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., "Chinchilla optimal" regime). In contrast, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but models are usually compared on downstream task performance. To address both shortcomings, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we fit scaling laws that extrapolate in both the amount of over-training and the number of model parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32$\times$ over-trained) and a 6.9B parameter, 138B token run (i.e., a compute-optimal run)$\unicode{x2014}$each from experiments that take 300$\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance by proposing a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models, using experiments that take 20$\times$ less compute. Our experiments are available at https://github.com/mlfoundations/scaling.

Yihuai Gao, Jinyun Liu, Shuang Li et al.

Robotic manipulation tasks exhibit varying memory requirements, ranging from Markovian tasks that require no memory to non-Markovian tasks that depend on historical information spanning single or multiple interaction trials. Surprisingly, simply extending observation histories of a visuomotor policy often leads to a significant performance drop due to distribution shift and overfitting. To address these issues, we propose Gated Memory Policy (GMP), a visuomotor policy that learns both when to recall memory and what to recall. To learn when to recall memory, GMP employs a learned memory gate mechanism that selectively activates history context only when necessary, improving robustness and reactivity. To learn what to recall efficiently, GMP introduces a lightweight cross-attention module that constructs effective latent memory representations. To further enhance robustness, GMP injects diffusion noise into historical actions, mitigating sensitivity to noisy or inaccurate histories during both training and inference. On our proposed non-Markovian benchmark MemMimic, GMP achieves a 30.1% average success rate improvement over long-history baselines, while maintaining competitive performance on Markovian tasks in RoboMimic. All code, data and in-the-wild deployment instructions are available on our project website https://gated-memory-policy.github.io/.

Yunfan Jiang, Ruohan Zhang, Josiah Wong et al. · stanford

Real-world household tasks present significant challenges for mobile manipulation robots. An analysis of existing robotics benchmarks reveals that successful task performance hinges on three key whole-body control capabilities: bimanual coordination, stable and precise navigation, and extensive end-effector reachability. Achieving these capabilities requires careful hardware design, but the resulting system complexity further complicates visuomotor policy learning. To address these challenges, we introduce the BEHAVIOR Robot Suite (BRS), a comprehensive framework for whole-body manipulation in diverse household tasks. Built on a bimanual, wheeled robot with a 4-DoF torso, BRS integrates a cost-effective whole-body teleoperation interface for data collection and a novel algorithm for learning whole-body visuomotor policies. We evaluate BRS on five challenging household tasks that not only emphasize the three core capabilities but also introduce additional complexities, such as long-range navigation, interaction with articulated and deformable objects, and manipulation in confined spaces. We believe that BRS's integrated robotic embodiment, data collection interface, and learning framework mark a significant step toward enabling real-world whole-body manipulation for everyday household tasks. BRS is open-sourced at https://behavior-robot-suite.github.io/

Hongchi Xia, Xuan Li, Zhaoshuo Li et al.

Real-world data collection for embodied agents remains costly and unsafe, calling for scalable, realistic, and simulator-ready 3D environments. However, existing scene-generation systems often rely on rule-based or task-specific pipelines, yielding artifacts and physically invalid scenes. We present SAGE, an agentic framework that, given a user-specified embodied task (e.g., "pick up a bowl and place it on the table"), understands the intent and automatically generates simulation-ready environments at scale. The agent couples multiple generators for layout and object composition with critics that evaluate semantic plausibility, visual realism, and physical stability. Through iterative reasoning and adaptive tool selection, it self-refines the scenes until meeting user intent and physical validity. The resulting environments are realistic, diverse, and directly deployable in modern simulators for policy training. Policies trained purely on this data exhibit clear scaling trends and generalize to unseen objects and layouts, demonstrating the promise of simulation-driven scaling for embodied AI. Code, demos, and the SAGE-10k dataset can be found on the project page here: https://nvlabs.github.io/sage.

Jimmy Wu, William Chong, Robert Holmberg et al.

Exploiting the promise of recent advances in imitation learning for mobile manipulation will require the collection of large numbers of human-guided demonstrations. This paper proposes an open-source design for an inexpensive, robust, and flexible mobile manipulator that can support arbitrary arms, enabling a wide range of real-world household mobile manipulation tasks. Crucially, our design uses powered casters to enable the mobile base to be fully holonomic, able to control all planar degrees of freedom independently and simultaneously. This feature makes the base more maneuverable and simplifies many mobile manipulation tasks, eliminating the kinematic constraints that create complex and time-consuming motions in nonholonomic bases. We equip our robot with an intuitive mobile phone teleoperation interface to enable easy data acquisition for imitation learning. In our experiments, we use this interface to collect data and show that the resulting learned policies can successfully perform a variety of common household mobile manipulation tasks.

Harsh Gupta, Xiaofeng Guo, Huy Ha et al.

We introduce UMI-on-Air, a framework for embodiment-aware deployment of embodiment-agnostic manipulation policies. Our approach leverages diverse, unconstrained human demonstrations collected with a handheld gripper (UMI) to train generalizable visuomotor policies. A central challenge in transferring these policies to constrained robotic embodiments-such as aerial manipulators-is the mismatch in control and robot dynamics, which often leads to out-of-distribution behaviors and poor execution. To address this, we propose Embodiment-Aware Diffusion Policy (EADP), which couples a high-level UMI policy with a low-level embodiment-specific controller at inference time. By integrating gradient feedback from the controller's tracking cost into the diffusion sampling process, our method steers trajectory generation towards dynamically feasible modes tailored to the deployment embodiment. This enables plug-and-play, embodiment-aware trajectory adaptation at test time. We validate our approach on multiple long-horizon and high-precision aerial manipulation tasks, showing improved success rates, efficiency, and robustness under disturbances compared to unguided diffusion baselines. Finally, we demonstrate deployment in previously unseen environments, using UMI demonstrations collected in the wild, highlighting a practical pathway for scaling generalizable manipulation skills across diverse-and even highly constrained-embodiments. All code, data, checkpoints, and result videos can be found at umi-on-air.github.io.

Xiaomeng Xu, Jisang Park, Han Zhang et al.

We present Whole-Body Mobile Manipulation Interface (HoMMI), a data collection and policy learning framework that learns whole-body mobile manipulation directly from robot-free human demonstrations. We augment UMI interfaces with egocentric sensing to capture the global context required for mobile manipulation, enabling portable, robot-free, and scalable data collection. However, naively incorporating egocentric sensing introduces a larger human-to-robot embodiment gap in both observation and action spaces, making policy transfer difficult. We explicitly bridge this gap with a cross-embodiment hand-eye policy design, including an embodiment agnostic visual representation; a relaxed head action representation; and a whole-body controller that realizes hand-eye trajectories through coordinated whole-body motion under robot-specific physical constraints. Together, these enable long-horizon mobile manipulation tasks requiring bimanual and whole-body coordination, navigation, and active perception. Results are best viewed on: https://hommi-robot.github.io

Zhanyi Sun, Shuran Song

We introduce Distribution Contractive Reinforcement Learning (DICE-RL), a framework that uses reinforcement learning (RL) as a "distribution contraction" operator to refine pretrained generative robot policies. DICE-RL turns a pretrained behavior prior into a high-performing "pro" policy by amplifying high-success behaviors from online feedback. We pretrain a diffusion- or flow-based policy for broad behavioral coverage, then finetune it with a stable, sample-efficient residual off-policy RL framework that combines selective behavior regularization with value-guided action selection. Extensive experiments and analyses show that DICE-RL reliably improves performance with strong stability and sample efficiency. It enables mastery of complex long-horizon manipulation skills directly from high-dimensional pixel inputs, both in simulation and on a real robot. Project website: https://zhanyisun.github.io/dice.rl.2026/.

Hao Li, Long Yin Chung, Jack Goler et al.

Underwater robotic grasping is difficult due to degraded, highly variable imagery and the expense of collecting diverse underwater demonstrations. We introduce a system that (i) autonomously collects successful underwater grasp demonstrations via a self-supervised data collection pipeline and (ii) transfers grasp knowledge from on-land human demonstrations through a depth-based affordance representation that bridges the on-land-to-underwater domain gap and is robust to lighting and color shift. An affordance model trained on on-land handheld demonstrations is deployed underwater zero-shot via geometric alignment, and an affordance-conditioned diffusion policy is then trained on underwater demonstrations to generate control actions. In pool experiments, our approach improves grasping performance and robustness to background shifts, and enables generalization to objects seen only in on-land data, outperforming RGB-only baselines. Code, videos, and additional results are available at https://umi-under-water.github.io.

Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song et al.

We propose Continuous Scene Representations (CSR), a scene representation constructed by an embodied agent navigating within a space, where objects and their relationships are modeled by continuous valued embeddings. Our method captures feature relationships between objects, composes them into a graph structure on-the-fly, and situates an embodied agent within the representation. Our key insight is to embed pair-wise relationships between objects in a latent space. This allows for a richer representation compared to discrete relations (e.g., [support], [next-to]) commonly used for building scene representations. CSR can track objects as the agent moves in a scene, update the representation accordingly, and detect changes in room configurations. Using CSR, we outperform state-of-the-art approaches for the challenging downstream task of visual room rearrangement, without any task specific training. Moreover, we show the learned embeddings capture salient spatial details of the scene and show applicability to real world data. A summery video and code is available at https://prior.allenai.org/projects/csr.

Shuang Li, Yihuai Gao, Dorsa Sadigh et al.

A unified video and action model holds significant promise for robotics, where videos provide rich scene information for action prediction, and actions provide dynamics information for video prediction. However, effectively combining video generation and action prediction remains challenging, and current video generation-based methods struggle to match the performance of direct policy learning in action accuracy and inference speed. To bridge this gap, we introduce the Unified Video Action model (UVA), which jointly optimizes video and action predictions to achieve both high accuracy and efficient action inference. The key lies in learning a joint video-action latent representation and decoupling video-action decoding. The joint latent representation bridges the visual and action domains, effectively modeling the relationship between video and action sequences. Meanwhile, the decoupled decoding, powered by two lightweight diffusion heads, enables high-speed action inference by bypassing video generation during inference. Such a unified framework further enables versatile functionality through masked input training. By selectively masking actions or videos, a single model can tackle diverse tasks beyond policy learning, such as forward and inverse dynamics modeling and video generation. Via an extensive set of experiments, we demonstrate that UVA can serve as a general-purpose solution for a wide range of robotics tasks, such as policy learning, forward/inverse dynamics and video observation prediction, without compromising performance compared to methods tailored for specific applications. Results are best viewed on https://unified-video-action-model.github.io/.

Ruoshi Liu, Alper Canberk, Shuran Song et al.

Vision foundation models trained on massive amounts of visual data have shown unprecedented reasoning and planning skills in open-world settings. A key challenge in applying them to robotic tasks is the modality gap between visual data and action data. We introduce differentiable robot rendering, a method allowing the visual appearance of a robot body to be directly differentiable with respect to its control parameters. Our model integrates a kinematics-aware deformable model and Gaussians Splatting and is compatible with any robot form factors and degrees of freedom. We demonstrate its capability and usage in applications including reconstruction of robot poses from images and controlling robots through vision language models. Quantitative and qualitative results show that our differentiable rendering model provides effective gradients for robotic control directly from pixels, setting the foundation for the future applications of vision foundation models in robotics.

Dominik Bauer, Zhenjia Xu, Shuran Song

Manipulation of elastoplastic objects like dough often involves topological changes such as splitting and merging. The ability to accurately predict these topological changes that a specific action might incur is critical for planning interactions with elastoplastic objects. We present DoughNet, a Transformer-based architecture for handling these challenges, consisting of two components. First, a denoising autoencoder represents deformable objects of varying topology as sets of latent codes. Second, a visual predictive model performs autoregressive set prediction to determine long-horizon geometrical deformation and topological changes purely in latent space. Given a partial initial state and desired manipulation trajectories, it infers all resulting object geometries and topologies at each step. DoughNet thereby allows to plan robotic manipulation; selecting a suited tool, its pose and opening width to recreate robot- or human-made goals. Our experiments in simulated and real environments show that DoughNet is able to significantly outperform related approaches that consider deformation only as geometrical change.

Jingxi Xu, Yinsen Jia, Dongxiao Yang et al.

We introduce GEOTACT, a robotic manipulation method capable of retrieving objects buried in granular media. This is a challenging task due to the need to interact with granular media, and doing so based exclusively on tactile feedback, since a buried object can be completely hidden from vision. Tactile feedback is in itself challenging in this context, due to ubiquitous contact with the surrounding media, and the inherent noise level induced by the tactile readings. To address these challenges, we use a learning method trained end-to-end with simulated sensor noise. We show that our problem formulation leads to the natural emergence of learned pushing behaviors that the manipulator uses to reduce uncertainty and funnel the object to a stable grasp despite spurious and noisy tactile readings. We also introduce a training curriculum that enables learning these behaviors in simulation, followed by zero-shot transfer to real hardware. To the best of our knowledge, GEOTACT is the first method to reliably retrieve a number of different objects from a granular environment, doing so on real hardware and with integrated tactile sensing. Videos and additional information can be found at https://jxu.ai/geotact.

Jiaxiang Tang, Ruijie Lu, Zhaoshuo Li et al.

Recent progress in 3D object generation has greatly improved both the quality and efficiency. However, most existing methods generate a single mesh with all parts fused together, which limits the ability to edit or manipulate individual parts. A key challenge is that different objects may have a varying number of parts. To address this, we propose a new end-to-end framework for part-level 3D object generation. Given a single input image, our method generates high-quality 3D objects with an arbitrary number of complete and semantically meaningful parts. We introduce a dual volume packing strategy that organizes all parts into two complementary volumes, allowing for the creation of complete and interleaved parts that assemble into the final object. Experiments show that our model achieves better quality, diversity, and generalization than previous image-based part-level generation methods.

Zhao Mandi, Yifan Hou, Dieter Fox et al.

We study the problem of functional retargeting: learning dexterous manipulation policies to track object states from human hand-object demonstrations. We focus on long-horizon, bimanual tasks with articulated objects, which is challenging due to large action space, spatiotemporal discontinuities, and embodiment gap between human and robot hands. We propose DexMachina, a novel curriculum-based algorithm: the key idea is to use virtual object controllers with decaying strength: an object is first driven automatically towards its target states, such that the policy can gradually learn to take over under motion and contact guidance. We release a simulation benchmark with a diverse set of tasks and dexterous hands, and show that DexMachina significantly outperforms baseline methods. Our algorithm and benchmark enable a functional comparison for hardware designs, and we present key findings informed by quantitative and qualitative results. With the recent surge in dexterous hand development, we hope this work will provide a useful platform for identifying desirable hardware capabilities and lower the barrier for contributing to future research. Videos and more at https://project-dexmachina.github.io/

Zixi Wang, Zeyi Liu, Nicolas Ouporov et al.

Robot-to-human object handover is an important step in many human robot collaboration tasks. A successful handover requires the robot to maintain a stable grasp on the object while making sure the human receives the object in a natural and easy-to-use manner. We propose ContactHandover, a robot to human handover system that consists of two phases: a contact-guided grasping phase and an object delivery phase. During the grasping phase, ContactHandover predicts both 6-DoF robot grasp poses and a 3D affordance map of human contact points on the object. The robot grasp poses are re-ranked by penalizing those that block human contact points, and the robot executes the highest ranking grasp. During the delivery phase, the robot end effector pose is computed by maximizing human contact points close to the human while minimizing the human arm joint torques and displacements. We evaluate our system on 27 diverse household objects and show that our system achieves better visibility and reachability of human contacts to the receiver compared to several baselines. More results can be found on https://clairezixiwang.github.io/ContactHandover.github.io

Xiaomeng Xu, Huy Ha, Shuran Song

We present Dynamics-Guided Diffusion Model (DGDM), a data-driven framework for generating task-specific manipulator designs without task-specific training. Given object shapes and task specifications, DGDM generates sensor-less manipulator designs that can blindly manipulate objects towards desired motions and poses using an open-loop parallel motion. This framework 1) flexibly represents manipulation tasks as interaction profiles, 2) represents the design space using a geometric diffusion model, and 3) efficiently searches this design space using the gradients provided by a dynamics network trained without any task information. We evaluate DGDM on various manipulation tasks ranging from shifting/rotating objects to converging objects to a specific pose. Our generated designs outperform optimization-based and unguided diffusion baselines relatively by 31.5% and 45.3% on average success rate. With the ability to generate a new design within 0.8s, DGDM facilitates rapid design iteration and enhances the adoption of data-driven approaches for robot mechanism design. Qualitative results are best viewed on our project website https://dgdm-robot.github.io/.

Zeyi Liu, Shuang Li, Eric Cousineau et al.

Understanding and predicting the dynamics of the physical world can enhance a robot's ability to plan and interact effectively in complex environments. While recent video generation models have shown strong potential in modeling dynamic scenes, generating videos that are both temporally coherent and geometrically consistent across camera views remains a significant challenge. To address this, we propose a 4D video generation model that enforces multi-view 3D consistency of videos by supervising the model with cross-view pointmap alignment during training. This geometric supervision enables the model to learn a shared 3D representation of the scene, allowing it to predict future video sequences from novel viewpoints based solely on the given RGB-D observations, without requiring camera poses as inputs. Compared to existing baselines, our method produces more visually stable and spatially aligned predictions across multiple simulated and real-world robotic datasets. We further show that the predicted 4D videos can be used to recover robot end-effector trajectories using an off-the-shelf 6DoF pose tracker, supporting robust robot manipulation and generalization to novel camera viewpoints.

Xiaomeng Xu, Yifan Hou, Zeyi Liu et al.

We address key challenges in Dataset Aggregation (DAgger) for real-world contact-rich manipulation: how to collect informative human correction data and how to effectively update policies with this new data. We introduce Compliant Residual DAgger (CR-DAgger), which contains two novel components: 1) a Compliant Intervention Interface that leverages compliance control, allowing humans to provide gentle, accurate delta action corrections without interrupting the ongoing robot policy execution; and 2) a Compliant Residual Policy formulation that learns from human corrections while incorporating force feedback and force control. Our system significantly enhances performance on precise contact-rich manipulation tasks using minimal correction data, improving base policy success rates by over 50\% on two challenging tasks (book flipping and belt assembly) while outperforming both retraining-from-scratch and finetuning approaches. Through extensive real-world experiments, we provide practical guidance for implementing effective DAgger in real-world robot learning tasks. Result videos are available at: https://compliant-residual-dagger.github.io/

Tao Sun, Liyuan Zhu, Shengyu Huang et al. · stanford

We introduce Rectified Point Flow, a unified parameterization that formulates pairwise point cloud registration and multi-part shape assembly as a single conditional generative problem. Given unposed point clouds, our method learns a continuous point-wise velocity field that transports noisy points toward their target positions, from which part poses are recovered. In contrast to prior work that regresses part-wise poses with ad-hoc symmetry handling, our method intrinsically learns assembly symmetries without symmetry labels. Together with a self-supervised encoder focused on overlapping points, our method achieves a new state-of-the-art performance on six benchmarks spanning pairwise registration and shape assembly. Notably, our unified formulation enables effective joint training on diverse datasets, facilitating the learning of shared geometric priors and consequently boosting accuracy. Project page: https://rectified-pointflow.github.io/.

Sedrick Keh, Jean Mercat, Samir Yitzhak Gadre et al. · cmu

Pre-trained LLMs that are further trained with image data perform well on vision-language tasks. While adding images during a second training phase effectively unlocks this capability, it is unclear how much of a gain or loss this two-step pipeline gives over VLMs which integrate images earlier into the training process. To investigate this, we train models spanning various datasets, scales, image-text ratios, and amount of pre-training done before introducing vision tokens. We then fine-tune these models and evaluate their downstream performance on a suite of vision-language and text-only tasks. We find that pre-training with a mixture of image and text data allows models to perform better on vision-language tasks while maintaining strong performance on text-only evaluations. On an average of 6 diverse tasks, we find that for a 1B model, introducing visual tokens 80% of the way through pre-training results in a 2% average improvement over introducing visual tokens to a fully pre-trained model.