Kuo-Hao Zeng

Matt Deitke, Christopher Clark, Sangho Lee et al. · allen-ai

Today's most advanced vision-language models (VLMs) remain proprietary. The strongest open-weight models rely heavily on synthetic data from proprietary VLMs to achieve good performance, effectively distilling these closed VLMs into open ones. As a result, the community has been missing foundational knowledge about how to build performant VLMs from scratch. We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. Our key contribution is a collection of new datasets called PixMo, including a dataset of highly detailed image captions for pre-training, a free-form image Q&A dataset for fine-tuning, and an innovative 2D pointing dataset, all collected without the use of external VLMs. The success of our approach relies on careful modeling choices, a well-tuned training pipeline, and, most critically, the quality of our newly collected datasets. Our best-in-class 72B model not only outperforms others in the class of open weight and data models, but also outperforms larger proprietary models including Claude 3.5 Sonnet, and Gemini 1.5 Pro and Flash, second only to GPT-4o based on both academic benchmarks and on a large human evaluation. Our model weights, new datasets, and source code are available at https://molmo.allenai.org/blog.

Ainaz Eftekhar, Kuo-Hao Zeng, Jiafei Duan et al. · uw

Embodied AI models often employ off the shelf vision backbones like CLIP to encode their visual observations. Although such general purpose representations encode rich syntactic and semantic information about the scene, much of this information is often irrelevant to the specific task at hand. This introduces noise within the learning process and distracts the agent's focus from task-relevant visual cues. Inspired by selective attention in humans-the process through which people filter their perception based on their experiences, knowledge, and the task at hand-we introduce a parameter-efficient approach to filter visual stimuli for embodied AI. Our approach induces a task-conditioned bottleneck using a small learnable codebook module. This codebook is trained jointly to optimize task reward and acts as a task-conditioned selective filter over the visual observation. Our experiments showcase state-of-the-art performance for object goal navigation and object displacement across 5 benchmarks, ProcTHOR, ArchitecTHOR, RoboTHOR, AI2-iTHOR, and ManipulaTHOR. The filtered representations produced by the codebook are also able generalize better and converge faster when adapted to other simulation environments such as Habitat. Our qualitative analyses show that agents explore their environments more effectively and their representations retain task-relevant information like target object recognition while ignoring superfluous information about other objects. Code and pretrained models are available at our project website: https://embodied-codebook.github.io.

Kuo-Hao Zeng, Luca Weihs, Roozbeh Mottaghi et al. · allen-ai

A common assumption when training embodied agents is that the impact of taking an action is stable; for instance, executing the "move ahead" action will always move the agent forward by a fixed distance, perhaps with some small amount of actuator-induced noise. This assumption is limiting; an agent may encounter settings that dramatically alter the impact of actions: a move ahead action on a wet floor may send the agent twice as far as it expects and using the same action with a broken wheel might transform the expected translation into a rotation. Instead of relying that the impact of an action stably reflects its pre-defined semantic meaning, we propose to model the impact of actions on-the-fly using latent embeddings. By combining these latent action embeddings with a novel, transformer-based, policy head, we design an Action Adaptive Policy (AAP). We evaluate our AAP on two challenging visual navigation tasks in the AI2-THOR and Habitat environments and show that our AAP is highly performant even when faced, at inference-time with missing actions and, previously unseen, perturbed action space. Moreover, we observe significant improvement in robustness against these actions when evaluating in real-world scenarios.

Jiaheng Hu, Rose Hendrix, Ali Farhadi et al.

In recent years, the Robotics field has initiated several efforts toward building generalist robot policies through large-scale multi-task Behavior Cloning. However, direct deployments of these policies have led to unsatisfactory performance, where the policy struggles with unseen states and tasks. How can we break through the performance plateau of these models and elevate their capabilities to new heights? In this paper, we propose FLaRe, a large-scale Reinforcement Learning fine-tuning framework that integrates robust pre-trained representations, large-scale training, and gradient stabilization techniques. Our method aligns pre-trained policies towards task completion, achieving state-of-the-art (SoTA) performance both on previously demonstrated and on entirely novel tasks and embodiments. Specifically, on a set of long-horizon mobile manipulation tasks, FLaRe achieves an average success rate of 79.5% in unseen environments, with absolute improvements of +23.6% in simulation and +30.7% on real robots over prior SoTA methods. By utilizing only sparse rewards, our approach can enable generalizing to new capabilities beyond the pretraining data with minimal human effort. Moreover, we demonstrate rapid adaptation to new embodiments and behaviors with less than a day of fine-tuning. Videos can be found on the project website at https://robot-flare.github.io/

Albert J. Zhai, Kuo-Hao Zeng, Jiasen Lu et al.

The ability to learn manipulation skills by watching videos of humans has the potential to unlock a new source of highly scalable data for robot learning. Here, we tackle prehensile manipulation, in which tasks involve grasping an object before performing various post-grasp motions. Human videos offer strong signals for learning the post-grasp motions, but they are less useful for learning the prerequisite grasping behaviors, especially for robots without human-like hands. A promising way forward is to use a modular policy design, leveraging a dedicated grasp generator to produce stable grasps. However, arbitrary stable grasps are often not task-compatible, hindering the robot's ability to perform the desired downstream motion. To address this challenge, we present Perceive-Simulate-Imitate (PSI), a framework for training a modular manipulation policy using human video motion data processed by paired grasp-trajectory filtering in simulation. This simulation step extends the trajectory data with grasp suitability labels, which allows for supervised learning of task-oriented grasping capabilities. We show through real-world experiments that our framework can be used to learn precise manipulation skills efficiently without any robot data, resulting in significantly more robust performance than using a grasp generator naively.

Kuo-Hao Zeng, Roozbeh Mottaghi, Luca Weihs et al.

In this paper we address the problem of visual reaction: the task of interacting with dynamic environments where the changes in the environment are not necessarily caused by the agent itself. Visual reaction entails predicting the future changes in a visual environment and planning accordingly. We study the problem of visual reaction in the context of playing catch with a drone in visually rich synthetic environments. This is a challenging problem since the agent is required to learn (1) how objects with different physical properties and shapes move, (2) what sequence of actions should be taken according to the prediction, (3) how to adjust the actions based on the visual feedback from the dynamic environment (e.g., when objects bouncing off a wall), and (4) how to reason and act with an unexpected state change in a timely manner. We propose a new dataset for this task, which includes 30K throws of 20 types of objects in different directions with different forces. Our results show that our model that integrates a forecaster with a planner outperforms a set of strong baselines that are based on tracking as well as pure model-based and model-free RL baselines. The code and dataset are available at github.com/KuoHaoZeng/Visual_Reaction.

Kiana Ehsani, Tanmay Gupta, Rose Hendrix et al. · allen-ai

Reinforcement learning (RL) with dense rewards and imitation learning (IL) with human-generated trajectories are the most widely used approaches for training modern embodied agents. RL requires extensive reward shaping and auxiliary losses and is often too slow and ineffective for long-horizon tasks. While IL with human supervision is effective, collecting human trajectories at scale is extremely expensive. In this work, we show that imitating shortest-path planners in simulation produces agents that, given a language instruction, can proficiently navigate, explore, and manipulate objects in both simulation and in the real world using only RGB sensors (no depth map or GPS coordinates). This surprising result is enabled by our end-to-end, transformer-based, SPOC architecture, powerful visual encoders paired with extensive image augmentation, and the dramatic scale and diversity of our training data: millions of frames of shortest-path-expert trajectories collected inside approximately 200,000 procedurally generated houses containing 40,000 unique 3D assets. Our models, data, training code, and newly proposed 10-task benchmarking suite CHORES are available in https://spoc-robot.github.io.

Arijit Ray, Jiafei Duan, Ellis Brown et al. · uw

Reasoning about motion and space is a fundamental cognitive capability that is required by multiple real-world applications. While many studies highlight that large multimodal language models (MLMs) struggle to reason about space, they only focus on static spatial relationships, and not dynamic awareness of motion and space, i.e., reasoning about the effect of egocentric and object motions on spatial relationships. Manually annotating such object and camera movements is expensive. Hence, we introduce SAT, a simulated spatial aptitude training dataset comprising both static and dynamic spatial reasoning across 175K question-answer (QA) pairs and 20K scenes. Complementing this, we also construct a small (150 image-QAs) yet challenging dynamic spatial test set using real-world images. Leveraging our SAT datasets and 6 existing static spatial benchmarks, we systematically investigate what improves both static and dynamic spatial awareness. Our results reveal that simulations are surprisingly effective at imparting spatial aptitude to MLMs that translate to real images. We show that perfect annotations in simulation are more effective than existing approaches of pseudo-annotating real images. For instance, SAT training improves a LLaVA-13B model by an average 11% and a LLaVA-Video-7B model by an average 8% on multiple spatial benchmarks, including our real-image dynamic test set and spatial reasoning on long videos -- even outperforming some large proprietary models. While reasoning over static relationships improves with synthetic training data, there is still considerable room for improvement for dynamic reasoning questions.

Ainaz Eftekhar, Rose Hendrix, Luca Weihs et al. · allen-ai

Modern robots vary significantly in shape, size, and sensor configurations used to perceive and interact with their environments. However, most navigation policies are embodiment-specific--a policy trained on one robot typically fails to generalize to another, even with minor changes in body size or camera viewpoint. As custom hardware becomes increasingly common, there is a growing need for a single policy that generalizes across embodiments, eliminating the need to retrain for each specific robot. In this paper, we introduce RING (Robotic Indoor Navigation Generalist), an embodiment-agnostic policy that turns any mobile robot into an effective indoor semantic navigator. Trained entirely in simulation, RING leverages large-scale randomization over robot embodiments to enable robust generalization to many real-world platforms. To support this, we augment the AI2-THOR simulator to instantiate robots with controllable configurations, varying in body size, rotation pivot point, and camera parameters. On the visual object-goal navigation task, RING achieves strong cross-embodiment (XE) generalization--72.1% average success rate across five simulated embodiments (a 16.7% absolute improvement on the Chores-S benchmark) and 78.9% across four real-world platforms, including Stretch RE-1, LoCoBot, and Unitree Go1--matching or even surpassing embodiment-specific policies. We further deploy RING on the RB-Y1 wheeled humanoid in a real-world kitchen environment, showcasing its out-of-the-box potential for mobile manipulation platforms. (Project website: https://one-ring-policy.allen.ai)

Ram Ramrakhya, Aniruddha Kembhavi, Dhruv Batra et al. · allen-ai

Computer vision tasks typically involve describing what is present in an image (e.g. classification, detection, segmentation, and captioning). We study a visual common sense task that requires understanding what is not present. Specifically, given an image (e.g. of a living room) and name of an object ("cushion"), a vision system is asked to predict semantically-meaningful regions (masks or bounding boxes) in the image where that object could be placed or is likely be placed by humans (e.g. on the sofa). We call this task: Semantic Placement (SP) and believe that such common-sense visual understanding is critical for assitive robots (tidying a house), and AR devices (automatically rendering an object in the user's space). Studying the invisible is hard. Datasets for image description are typically constructed by curating relevant images and asking humans to annotate the contents of the image; neither of those two steps are straightforward for objects not present in the image. We overcome this challenge by operating in the opposite direction: we start with an image of an object in context from web, and then remove that object from the image via inpainting. This automated pipeline converts unstructured web data into a dataset comprising pairs of images with/without the object. Using this, we collect a novel dataset, with ${\sim}1.3$M images across $9$ object categories, and train a SP prediction model called CLIP-UNet. CLIP-UNet outperforms existing VLMs and baselines that combine semantic priors with object detectors on real-world and simulated images. In our user studies, we find that the SP masks predicted by CLIP-UNet are favored $43.7\%$ and $31.3\%$ times when comparing against the $4$ SP baselines on real and simulated images. In addition, we demonstrate leveraging SP mask predictions from CLIP-UNet enables downstream applications like building tidying robots in indoor environments.

Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani et al.

We present PoliFormer (Policy Transformer), an RGB-only indoor navigation agent trained end-to-end with reinforcement learning at scale that generalizes to the real-world without adaptation despite being trained purely in simulation. PoliFormer uses a foundational vision transformer encoder with a causal transformer decoder enabling long-term memory and reasoning. It is trained for hundreds of millions of interactions across diverse environments, leveraging parallelized, multi-machine rollouts for efficient training with high throughput. PoliFormer is a masterful navigator, producing state-of-the-art results across two distinct embodiments, the LoCoBot and Stretch RE-1 robots, and four navigation benchmarks. It breaks through the plateaus of previous work, achieving an unprecedented 85.5% success rate in object goal navigation on the CHORES-S benchmark, a 28.5% absolute improvement. PoliFormer can also be trivially extended to a variety of downstream applications such as object tracking, multi-object navigation, and open-vocabulary navigation with no finetuning.

Kuo-Hao Zeng, Luca Weihs, Ali Farhadi et al.

We have observed significant progress in visual navigation for embodied agents. A common assumption in studying visual navigation is that the environments are static; this is a limiting assumption. Intelligent navigation may involve interacting with the environment beyond just moving forward/backward and turning left/right. Sometimes, the best way to navigate is to push something out of the way. In this paper, we study the problem of interactive navigation where agents learn to change the environment to navigate more efficiently to their goals. To this end, we introduce the Neural Interaction Engine (NIE) to explicitly predict the change in the environment caused by the agent's actions. By modeling the changes while planning, we find that agents exhibit significant improvements in their navigational capabilities. More specifically, we consider two downstream tasks in the physics-enabled, visually rich, AI2-THOR environment: (1) reaching a target while the path to the target is blocked (2) moving an object to a target location by pushing it. For both tasks, agents equipped with an NIE significantly outperform agents without the understanding of the effect of the actions indicating the benefits of our approach.

Luca Weihs, Jordi Salvador, Klemen Kotar et al.

The domain of Embodied AI, in which agents learn to complete tasks through interaction with their environment from egocentric observations, has experienced substantial growth with the advent of deep reinforcement learning and increased interest from the computer vision, NLP, and robotics communities. This growth has been facilitated by the creation of a large number of simulated environments (such as AI2-THOR, Habitat and CARLA), tasks (like point navigation, instruction following, and embodied question answering), and associated leaderboards. While this diversity has been beneficial and organic, it has also fragmented the community: a huge amount of effort is required to do something as simple as taking a model trained in one environment and testing it in another. This discourages good science. We introduce AllenAct, a modular and flexible learning framework designed with a focus on the unique requirements of Embodied AI research. AllenAct provides first-class support for a growing collection of embodied environments, tasks and algorithms, provides reproductions of state-of-the-art models and includes extensive documentation, tutorials, start-up code, and pre-trained models. We hope that our framework makes Embodied AI more accessible and encourages new researchers to join this exciting area. The framework can be accessed at: https://allenact.org/

Kuo-Hao Zeng, Mohammad Shoeybi, Ming-Yu Liu

We introduce a language generative model framework for generating a styled paragraph based on a context sentence and a style reference example. The framework consists of a style encoder and a texts decoder. The style encoder extracts a style code from the reference example, and the text decoder generates texts based on the style code and the context. We propose a novel objective function to train our framework. We also investigate different network design choices. We conduct extensive experimental validation with comparison to strong baselines to validate the effectiveness of the proposed framework using a newly collected dataset with diverse text styles. Both code and dataset will be released upon publication.

Tsun-Hsuan Wang, Hung-Jui Huang, Juan-Ting Lin et al.

$ $Visual place recognition is challenging, especially when only a few place exemplars are given. To mitigate the challenge, we consider place recognition method using omnidirectional cameras and propose a novel Omnidirectional Convolutional Neural Network (O-CNN) to handle severe camera pose variation. Given a visual input, the task of the O-CNN is not to retrieve the matched place exemplar, but to retrieve the closest place exemplar and estimate the relative distance between the input and the closest place. With the ability to estimate relative distance, a heuristic policy is proposed to navigate a robot to the retrieved closest place. Note that the network is designed to take advantage of the omnidirectional view by incorporating circular padding and rotation invariance. To train a powerful O-CNN, we build a virtual world for training on a large scale. We also propose a continuous lifted structured feature embedding loss to learn the concept of distance efficiently. Finally, our experimental results confirm that our method achieves state-of-the-art accuracy and speed with both the virtual world and real-world datasets.

Shih-Han Chou, Yi-Chun Chen, Kuo-Hao Zeng et al.

Narrated 360° videos are typically provided in many touring scenarios to mimic real-world experience. However, previous work has shown that smart assistance (i.e., providing visual guidance) can significantly help users to follow the Normal Field of View (NFoV) corresponding to the narrative. In this project, we aim at automatically grounding the NFoVs of a 360° video given subtitles of the narrative (referred to as "NFoV-grounding"). We propose a novel Visual Grounding Model (VGM) to implicitly and efficiently predict the NFoVs given the video content and subtitles. Specifically, at each frame, we efficiently encode the panorama into feature map of candidate NFoVs using a Convolutional Neural Network (CNN) and the subtitles to the same hidden space using an RNN with Gated Recurrent Units (GRU). Then, we apply soft-attention on candidate NFoVs to trigger sentence decoder aiming to minimize the reconstruct loss between the generated and given sentence. Finally, we obtain the NFoV as the candidate NFoV with the maximum attention without any human supervision. To train VGM more robustly, we also generate a reverse sentence conditioning on one minus the soft-attention such that the attention focuses on candidate NFoVs less relevant to the given sentence. The negative log reconstruction loss of the reverse sentence (referred to as "irrelevant loss") is jointly minimized to encourage the reverse sentence to be different from the given sentence. To evaluate our method, we collect the first narrated 360° videos dataset and achieve state-of-the-art NFoV-grounding performance.

Kuo-Hao Zeng, William B. Shen, De-An Huang et al.

We introduce a general framework for visual forecasting, which directly imitates visual sequences without additional supervision. As a result, our model can be applied at several semantic levels and does not require any domain knowledge or handcrafted features. We achieve this by formulating visual forecasting as an inverse reinforcement learning (IRL) problem, and directly imitate the dynamics in natural sequences from their raw pixel values. The key challenge is the high-dimensional and continuous state-action space that prohibits the application of previous IRL algorithms. We address this computational bottleneck by extending recent progress in model-free imitation with trainable deep feature representations, which (1) bypasses the exhaustive state-action pair visits in dynamic programming by using a dual formulation and (2) avoids explicit state sampling at gradient computation using a deep feature reparametrization. This allows us to apply IRL at scale and directly imitate the dynamics in high-dimensional continuous visual sequences from the raw pixel values. We evaluate our approach at three different level-of-abstraction, from low level pixels to higher level semantics: future frame generation, action anticipation, visual story forecasting. At all levels, our approach outperforms existing methods.

Kuo-Hao Zeng, Shih-Han Chou, Fu-Hsiang Chan et al.

For survival, a living agent must have the ability to assess risk (1) by temporally anticipating accidents before they occur, and (2) by spatially localizing risky regions in the environment to move away from threats. In this paper, we take an agent-centric approach to study the accident anticipation and risky region localization tasks. We propose a novel soft-attention Recurrent Neural Network (RNN) which explicitly models both spatial and appearance-wise non-linear interaction between the agent triggering the event and another agent or static-region involved. In order to test our proposed method, we introduce the Epic Fail (EF) dataset consisting of 3000 viral videos capturing various accidents. In the experiments, we evaluate the risk assessment accuracy both in the temporal domain (accident anticipation) and spatial domain (risky region localization) on our EF dataset and the Street Accident (SA) dataset. Our method consistently outperforms other baselines on both datasets.

Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang et al.

We propose a scalable approach to learn video-based question answering (QA): answer a "free-form natural language question" about a video content. Our approach automatically harvests a large number of videos and descriptions freely available online. Then, a large number of candidate QA pairs are automatically generated from descriptions rather than manually annotated. Next, we use these candidate QA pairs to train a number of video-based QA methods extended fromMN (Sukhbaatar et al. 2015), VQA (Antol et al. 2015), SA (Yao et al. 2015), SS (Venugopalan et al. 2015). In order to handle non-perfect candidate QA pairs, we propose a self-paced learning procedure to iteratively identify them and mitigate their effects in training. Finally, we evaluate performance on manually generated video-based QA pairs. The results show that our self-paced learning procedure is effective, and the extended SS model outperforms various baselines.

Kuo-Hao Zeng, Tseng-Hung Chen, Juan Carlos Niebles et al.

A great video title describes the most salient event compactly and captures the viewer's attention. In contrast, video captioning tends to generate sentences that describe the video as a whole. Although generating a video title automatically is a very useful task, it is much less addressed than video captioning. We address video title generation for the first time by proposing two methods that extend state-of-the-art video captioners to this new task. First, we make video captioners highlight sensitive by priming them with a highlight detector. Our framework allows for jointly training a model for title generation and video highlight localization. Second, we induce high sentence diversity in video captioners, so that the generated titles are also diverse and catchy. This means that a large number of sentences might be required to learn the sentence structure of titles. Hence, we propose a novel sentence augmentation method to train a captioner with additional sentence-only examples that come without corresponding videos. We collected a large-scale Video Titles in the Wild (VTW) dataset of 18100 automatically crawled user-generated videos and titles. On VTW, our methods consistently improve title prediction accuracy, and achieve the best performance in both automatic and human evaluation. Finally, our sentence augmentation method also outperforms the baselines on the M-VAD dataset.