Jitendra Malik

Kristen Grauman, Andrew Westbury, Lorenzo Torresani et al. · cmu, gatech

We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured egocentric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions -- including a novel "expert commentary" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled human activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community. Project page: http://ego-exo4d-data.org/

Arjun Majumdar, Karmesh Yadav, Sergio Arnaud et al. · meta-ai

We present the largest and most comprehensive empirical study of pre-trained visual representations (PVRs) or visual 'foundation models' for Embodied AI. First, we curate CortexBench, consisting of 17 different tasks spanning locomotion, navigation, dexterous, and mobile manipulation. Next, we systematically evaluate existing PVRs and find that none are universally dominant. To study the effect of pre-training data size and diversity, we combine over 4,000 hours of egocentric videos from 7 different sources (over 4.3M images) and ImageNet to train different-sized vision transformers using Masked Auto-Encoding (MAE) on slices of this data. Contrary to inferences from prior work, we find that scaling dataset size and diversity does not improve performance universally (but does so on average). Our largest model, named VC-1, outperforms all prior PVRs on average but does not universally dominate either. Next, we show that task- or domain-specific adaptation of VC-1 leads to substantial gains, with VC-1 (adapted) achieving competitive or superior performance than the best known results on all of the benchmarks in CortexBench. Finally, we present real-world hardware experiments, in which VC-1 and VC-1 (adapted) outperform the strongest pre-existing PVR. Overall, this paper presents no new techniques but a rigorous systematic evaluation, a broad set of findings about PVRs (that in some cases, refute those made in narrow domains in prior work), and open-sourced code and models (that required over 10,000 GPU-hours to train) for the benefit of the research community.

Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya et al. · meta-ai

Modern hierarchical vision transformers have added several vision-specific components in the pursuit of supervised classification performance. While these components lead to effective accuracies and attractive FLOP counts, the added complexity actually makes these transformers slower than their vanilla ViT counterparts. In this paper, we argue that this additional bulk is unnecessary. By pretraining with a strong visual pretext task (MAE), we can strip out all the bells-and-whistles from a state-of-the-art multi-stage vision transformer without losing accuracy. In the process, we create Hiera, an extremely simple hierarchical vision transformer that is more accurate than previous models while being significantly faster both at inference and during training. We evaluate Hiera on a variety of tasks for image and video recognition. Our code and models are available at https://github.com/facebookresearch/hiera.

Sehoon Kim, Amir Gholami, Albert Shaw et al. · gatech

The recently proposed Conformer model has become the de facto backbone model for various downstream speech tasks based on its hybrid attention-convolution architecture that captures both local and global features. However, through a series of systematic studies, we find that the Conformer architecture's design choices are not optimal. After re-examining the design choices for both the macro and micro-architecture of Conformer, we propose Squeezeformer which consistently outperforms the state-of-the-art ASR models under the same training schemes. In particular, for the macro-architecture, Squeezeformer incorporates (i) the Temporal U-Net structure which reduces the cost of the multi-head attention modules on long sequences, and (ii) a simpler block structure of multi-head attention or convolution modules followed up by feed-forward module instead of the Macaron structure proposed in Conformer. Furthermore, for the micro-architecture, Squeezeformer (i) simplifies the activations in the convolutional block, (ii) removes redundant Layer Normalization operations, and (iii) incorporates an efficient depthwise down-sampling layer to efficiently sub-sample the input signal. Squeezeformer achieves state-of-the-art results of 7.5%, 6.5%, and 6.0% word-error-rate (WER) on LibriSpeech test-other without external language models, which are 3.1%, 1.4%, and 0.6% better than Conformer-CTC with the same number of FLOPs. Our code is open-sourced and available online.

Ananye Agarwal, Ashish Kumar, Jitendra Malik et al. · berkeley

Animals are capable of precise and agile locomotion using vision. Replicating this ability has been a long-standing goal in robotics. The traditional approach has been to decompose this problem into elevation mapping and foothold planning phases. The elevation mapping, however, is susceptible to failure and large noise artifacts, requires specialized hardware, and is biologically implausible. In this paper, we present the first end-to-end locomotion system capable of traversing stairs, curbs, stepping stones, and gaps. We show this result on a medium-sized quadruped robot using a single front-facing depth camera. The small size of the robot necessitates discovering specialized gait patterns not seen elsewhere. The egocentric camera requires the policy to remember past information to estimate the terrain under its hind feet. We train our policy in simulation. Training has two phases - first, we train a policy using reinforcement learning with a cheap-to-compute variant of depth image and then in phase 2 distill it into the final policy that uses depth using supervised learning. The resulting policy transfers to the real world and is able to run in real-time on the limited compute of the robot. It can traverse a large variety of terrain while being robust to perturbations like pushes, slippery surfaces, and rocky terrain. Videos are at https://vision-locomotion.github.io

Karttikeya Mangalam, Raiymbek Akshulakov, Jitendra Malik

We introduce EgoSchema, a very long-form video question-answering dataset, and benchmark to evaluate long video understanding capabilities of modern vision and language systems. Derived from Ego4D, EgoSchema consists of over 5000 human curated multiple choice question answer pairs, spanning over 250 hours of real video data, covering a very broad range of natural human activity and behavior. For each question, EgoSchema requires the correct answer to be selected between five given options based on a three-minute-long video clip. While some prior works have proposed video datasets with long clip lengths, we posit that merely the length of the video clip does not truly capture the temporal difficulty of the video task that is being considered. To remedy this, we introduce temporal certificate sets, a general notion for capturing the intrinsic temporal understanding length associated with a broad range of video understanding tasks & datasets. Based on this metric, we find EgoSchema to have intrinsic temporal lengths over 5.7x longer than the second closest dataset and 10x to 100x longer than any other video understanding dataset. Further, our evaluation of several current state-of-the-art video and language models shows them to be severely lacking in long-term video understanding capabilities. Even models with several billions of parameters achieve QA accuracy less than 33% (random is 20%) on the EgoSchema multi-choice question answering task, while humans achieve about 76% accuracy. We posit that \name{}{}, with its long intrinsic temporal structures and diverse complexity, would serve as a valuable evaluation probe for developing effective long-term video understanding systems in the future. Data and Zero-shot model evaluation code are open-sourced for both public and commercial use under the Ego4D license at http://egoschema.github.io

Sehoon Kim, Karttikeya Mangalam, Suhong Moon et al.

The recent emergence of Large Language Models based on the Transformer architecture has enabled dramatic advancements in the field of Natural Language Processing. However, these models have long inference latency, which limits their deployment and makes them prohibitively expensive for various real-time applications. The inference latency is further exacerbated by autoregressive generative tasks, as models need to run iteratively to generate tokens sequentially without leveraging token-level parallelization. To address this, we propose Big Little Decoder (BiLD), a framework that can improve inference efficiency and latency for a wide range of text generation applications. The BiLD framework contains two models with different sizes that collaboratively generate text. The small model runs autoregressively to generate text with a low inference cost, and the large model is only invoked occasionally to refine the small model's inaccurate predictions in a non-autoregressive manner. To coordinate the small and large models, BiLD introduces two simple yet effective policies: (1) the fallback policy that determines when to hand control over to the large model; and (2) the rollback policy that determines when the large model needs to correct the small model's inaccurate predictions. To evaluate our framework across different tasks and models, we apply BiLD to various text generation scenarios encompassing machine translation on IWSLT 2017 De-En and WMT 2014 De-En, and summarization on XSUM and CNN/DailyMail. On an NVIDIA T4 GPU, our framework achieves a speedup of up to 2.12x speedup with minimal generation quality degradation. Furthermore, our framework is fully plug-and-play and can be applied without any modifications in the training process or model architecture. Our code is open-sourced

Ashish Kumar, Zhongyu Li, Jun Zeng et al. · berkeley

Recent advances in legged locomotion have enabled quadrupeds to walk on challenging terrains. However, bipedal robots are inherently more unstable and hence it's harder to design walking controllers for them. In this work, we leverage recent advances in rapid adaptation for locomotion control, and extend them to work on bipedal robots. Similar to existing works, we start with a base policy which produces actions while taking as input an estimated extrinsics vector from an adaptation module. This extrinsics vector contains information about the environment and enables the walking controller to rapidly adapt online. However, the extrinsics estimator could be imperfect, which might lead to poor performance of the base policy which expects a perfect estimator. In this paper, we propose A-RMA (Adapting RMA), which additionally adapts the base policy for the imperfect extrinsics estimator by finetuning it using model-free RL. We demonstrate that A-RMA outperforms a number of RL-based baseline controllers and model-based controllers in simulation, and show zero-shot deployment of a single A-RMA policy to enable a bipedal robot, Cassie, to walk in a variety of different scenarios in the real world beyond what it has seen during training. Videos and results at https://ashish-kmr.github.io/a-rma/

William Peebles, Ilija Radosavovic, Tim Brooks et al. · berkeley

We explore a data-driven approach for learning to optimize neural networks. We construct a dataset of neural network checkpoints and train a generative model on the parameters. In particular, our model is a conditional diffusion transformer that, given an initial input parameter vector and a prompted loss, error, or return, predicts the distribution over parameter updates that achieve the desired metric. At test time, it can optimize neural networks with unseen parameters for downstream tasks in just one update. We find that our approach successfully generates parameters for a wide range of loss prompts. Moreover, it can sample multimodal parameter solutions and has favorable scaling properties. We apply our method to different neural network architectures and tasks in supervised and reinforcement learning.

Haozhi Qi, Ashish Kumar, Roberto Calandra et al. · berkeley

Generalized in-hand manipulation has long been an unsolved challenge of robotics. As a small step towards this grand goal, we demonstrate how to design and learn a simple adaptive controller to achieve in-hand object rotation using only fingertips. The controller is trained entirely in simulation on only cylindrical objects, which then - without any fine-tuning - can be directly deployed to a real robot hand to rotate dozens of objects with diverse sizes, shapes, and weights over the z-axis. This is achieved via rapid online adaptation of the controller to the object properties using only proprioception history. Furthermore, natural and stable finger gaits automatically emerge from training the control policy via reinforcement learning. Code and more videos are available at https://haozhi.io/hora

Dingqi Zhang, Antonio Loquercio, Xiangyu Wu et al. · berkeley

This paper proposes an adaptive near-hover position controller for quadcopters, which can be deployed to quadcopters of very different mass, size and motor constants, and also shows rapid adaptation to unknown disturbances during runtime. The core algorithmic idea is to learn a single policy that can adapt online at test time not only to the disturbances applied to the drone, but also to the robot dynamics and hardware in the same framework. We achieve this by training a neural network to estimate a latent representation of the robot and environment parameters, which is used to condition the behaviour of the controller, also represented as a neural network. We train both networks exclusively in simulation with the goal of flying the quadcopters to goal positions and avoiding crashes to the ground. We directly deploy the same controller trained in the simulation without any modifications on two quadcopters in the real world with differences in mass, size, motors, and propellers with mass differing by 4.5 times. In addition, we show rapid adaptation to sudden and large disturbances up to one-third of the mass of the quadcopters. We perform an extensive evaluation in both simulation and the physical world, where we outperform a state-of-the-art learning-based adaptive controller and a traditional PID controller specifically tuned to each platform individually. Video results can be found at https://youtu.be/U-c-LbTfvoA.

Po-Yao Huang, Vasu Sharma, Hu Xu et al. · meta-ai, mit

We present Masked Audio-Video Learners (MAViL) to train audio-visual representations. Our approach learns with three complementary forms of self-supervision: (1) reconstruction of masked audio and video input data, (2) intra- and inter-modal contrastive learning with masking, and (3) self-training by reconstructing joint audio-video contextualized features learned from the first two objectives. Pre-training with MAViL not only enables the model to perform well in audio-visual classification and retrieval tasks but also improves representations of each modality in isolation, without using information from the other modality for fine-tuning or inference. Empirically, MAViL sets a new state-of-the-art on AudioSet (53.1 mAP) and VGGSound (67.1% accuracy). For the first time, a self-supervised audio-visual model outperforms ones that use external supervision on these benchmarks.

Karttikeya Mangalam, Haoqi Fan, Yanghao Li et al.

We present Reversible Vision Transformers, a memory efficient architecture design for visual recognition. By decoupling the GPU memory requirement from the depth of the model, Reversible Vision Transformers enable scaling up architectures with efficient memory usage. We adapt two popular models, namely Vision Transformer and Multiscale Vision Transformers, to reversible variants and benchmark extensively across both model sizes and tasks of image classification, object detection and video classification. Reversible Vision Transformers achieve a reduced memory footprint of up to 15.5x at roughly identical model complexity, parameters and accuracy, demonstrating the promise of reversible vision transformers as an efficient backbone for hardware resource limited training regimes. Finally, we find that the additional computational burden of recomputing activations is more than overcome for deeper models, where throughput can increase up to 2.3x over their non-reversible counterparts. Full code and trained models are available at https://github.com/facebookresearch/slowfast. A simpler, easy to understand and modify version is also available at https://github.com/karttikeya/minREV

Weiyao Wang, Matt Feiszli, Heng Wang et al.

Open-world instance segmentation is the task of grouping pixels into object instances without any pre-determined taxonomy. This is challenging, as state-of-the-art methods rely on explicit class semantics obtained from large labeled datasets, and out-of-domain evaluation performance drops significantly. Here we propose a novel approach for mask proposals, Generic Grouping Networks (GGNs), constructed without semantic supervision. Our approach combines a local measure of pixel affinity with instance-level mask supervision, producing a training regimen designed to make the model as generic as the data diversity allows. We introduce a method for predicting Pairwise Affinities (PA), a learned local relationship between pairs of pixels. PA generalizes very well to unseen categories. From PA we construct a large set of pseudo-ground-truth instance masks; combined with human-annotated instance masks we train GGNs and significantly outperform the SOTA on open-world instance segmentation on various benchmarks including COCO, LVIS, ADE20K, and UVO. Code is available on project website: https://sites.google.com/view/generic-grouping/.

Yikai Tang, Haoran Geng, Jindou Jia et al.

Imitation learning has emerged as a crucial approach for acquiring visuomotor skills from demonstrations, where designing effective observation encoders is essential for policy generalization. However, existing methods tend to struggle once test-time conditions differ from the demonstrations, such as changes in lighting, texture, viewpoint, object placement, or object identity. To address this challenge, we propose DIffusion POlicy with compLementarity Encoders (DIPOLE), a visuomotor policy that learns to fuse complementary modalities through a training-time mechanism rather than a specialized fusion architecture. A modality-wise dropout masks one branch at each training step, encouraging each modality to remain individually informative. A lightweight cross-attention layer then exchanges complementary cues between the two. This design endows DIPOLE with five core strengths: stable high performance across diverse tasks, robustness to visual changes, spatial generalization at sub-centimeter precision, emergent capability beyond either modality, and zero-shot transfer to unseen objects. Across 18 simulated and 4 real-world tasks, DIPOLE outperforms six baselines by 39.1% on average, with gains of 41.5% under unseen visual distractors and 15.2% under randomized object placement.

Sandeep Bajamahal, Lawrence Yunliang Chen, Toru Lin et al.

Bimanual coordination is essential for many real-world manipulation tasks, yet learning bimanual robot policies is limited by the scarcity of bimanual robots and datasets. Single-arm robots, however, are widely available in research labs. Can we leverage them to train bimanual robot policies? We present MonoDuo, a framework for learning bimanual manipulation policies using single-arm robot demonstrations paired with human collaboration. MonoDuo collects data by teleoperating a single-arm robot to perform one side of a bimanual task while a human performs the other, then swapping roles to cover both sides. RGB-D observations from a wrist-mounted and fixed camera are augmented into synthetic demonstrations for target bimanual robots using state-of-the-art hand pose estimation, image and point cloud segmentation, and inpainting. These synthetic demonstrations, grounded in real robot kinematics, are used to train bimanual policies. We evaluate MonoDuo on five tasks: box lifting, backpack packing, cloth folding, jacket zipping, and plate handover. Compared to approaches relying solely on human bimanual videos, MonoDuo enables zero-shot deployment on unseen bimanual robot configurations, achieving success rates up to 70%. With only 25 target robot demonstrations, few-shot finetuning further boosts success rates by 65-70% over training from scratch, demonstrating MonoDuo's effectiveness in efficiently transferring knowledge from single-arm robot data to bimanual robot policies.

Haozhi Qi, Brent Yi, Sudharshan Suresh et al. · berkeley

We introduce RotateIt, a system that enables fingertip-based object rotation along multiple axes by leveraging multimodal sensory inputs. Our system is trained in simulation, where it has access to ground-truth object shapes and physical properties. Then we distill it to operate on realistic yet noisy simulated visuotactile and proprioceptive sensory inputs. These multimodal inputs are fused via a visuotactile transformer, enabling online inference of object shapes and physical properties during deployment. We show significant performance improvements over prior methods and the importance of visual and tactile sensing.

Ilija Radosavovic, Tete Xiao, Stephen James et al.

In this work, we explore self-supervised visual pre-training on images from diverse, in-the-wild videos for real-world robotic tasks. Like prior work, our visual representations are pre-trained via a masked autoencoder (MAE), frozen, and then passed into a learnable control module. Unlike prior work, we show that the pre-trained representations are effective across a range of real-world robotic tasks and embodiments. We find that our encoder consistently outperforms CLIP (up to 75%), supervised ImageNet pre-training (up to 81%), and training from scratch (up to 81%). Finally, we train a 307M parameter vision transformer on a massive collection of 4.5M images from the Internet and egocentric videos, and demonstrate clearly the benefits of scaling visual pre-training for robot learning.

Jasmine Collins, Anqi Liang, Jitendra Malik et al. · berkeley

We present a neural network approach to transfer the motion from a single image of an articulated object to a rest-state (i.e., unarticulated) 3D model. Our network learns to predict the object's pose, part segmentation, and corresponding motion parameters to reproduce the articulation shown in the input image. The network is composed of three distinct branches that take a shared joint image-shape embedding and is trained end-to-end. Unlike previous methods, our approach is independent of the topology of the object and can work with objects from arbitrary categories. Our method, trained with only synthetic data, can be used to automatically animate a mesh, infer motion from real images, and transfer articulation to functionally similar but geometrically distinct 3D models at test time.

Boyi Li, Rodolfo Corona, Karttikeya Mangalam et al. · berkeley

Are multimodal inputs necessary for grammar induction? Recent work has shown that multimodal training inputs can improve grammar induction. However, these improvements are based on comparisons to weak text-only baselines that were trained on relatively little textual data. To determine whether multimodal inputs are needed in regimes with large amounts of textual training data, we design a stronger text-only baseline, which we refer to as LC-PCFG. LC-PCFG is a C-PFCG that incorporates em-beddings from text-only large language models (LLMs). We use a fixed grammar family to directly compare LC-PCFG to various multi-modal grammar induction methods. We compare performance on four benchmark datasets. LC-PCFG provides an up to 17% relative improvement in Corpus-F1 compared to state-of-the-art multimodal grammar induction methods. LC-PCFG is also more computationally efficient, providing an up to 85% reduction in parameter count and 8.8x reduction in training time compared to multimodal approaches. These results suggest that multimodal inputs may not be necessary for grammar induction, and emphasize the importance of strong vision-free baselines for evaluating the benefit of multimodal approaches.

Tete Xiao, Ilija Radosavovic, Trevor Darrell et al.

This paper shows that self-supervised visual pre-training from real-world images is effective for learning motor control tasks from pixels. We first train the visual representations by masked modeling of natural images. We then freeze the visual encoder and train neural network controllers on top with reinforcement learning. We do not perform any task-specific fine-tuning of the encoder; the same visual representations are used for all motor control tasks. To the best of our knowledge, this is the first self-supervised model to exploit real-world images at scale for motor control. To accelerate progress in learning from pixels, we contribute a benchmark suite of hand-designed tasks varying in movements, scenes, and robots. Without relying on labels, state-estimation, or expert demonstrations, we consistently outperform supervised encoders by up to 80% absolute success rate, sometimes even matching the oracle state performance. We also find that in-the-wild images, e.g., from YouTube or Egocentric videos, lead to better visual representations for various manipulation tasks than ImageNet images.

Jordan Lekeufack, Anastasios N. Angelopoulos, Andrea Bajcsy et al. · berkeley

We introduce Conformal Decision Theory, a framework for producing safe autonomous decisions despite imperfect machine learning predictions. Examples of such decisions are ubiquitous, from robot planning algorithms that rely on pedestrian predictions, to calibrating autonomous manufacturing to exhibit high throughput and low error, to the choice of trusting a nominal policy versus switching to a safe backup policy at run-time. The decisions produced by our algorithms are safe in the sense that they come with provable statistical guarantees of having low risk without any assumptions on the world model whatsoever; the observations need not be I.I.D. and can even be adversarial. The theory extends results from conformal prediction to calibrate decisions directly, without requiring the construction of prediction sets. Experiments demonstrate the utility of our approach in robot motion planning around humans, automated stock trading, and robot manufacturing.

Xavier Puig, Eric Undersander, Andrew Szot et al.

We present Habitat 3.0: a simulation platform for studying collaborative human-robot tasks in home environments. Habitat 3.0 offers contributions across three dimensions: (1) Accurate humanoid simulation: addressing challenges in modeling complex deformable bodies and diversity in appearance and motion, all while ensuring high simulation speed. (2) Human-in-the-loop infrastructure: enabling real human interaction with simulated robots via mouse/keyboard or a VR interface, facilitating evaluation of robot policies with human input. (3) Collaborative tasks: studying two collaborative tasks, Social Navigation and Social Rearrangement. Social Navigation investigates a robot's ability to locate and follow humanoid avatars in unseen environments, whereas Social Rearrangement addresses collaboration between a humanoid and robot while rearranging a scene. These contributions allow us to study end-to-end learned and heuristic baselines for human-robot collaboration in-depth, as well as evaluate them with humans in the loop. Our experiments demonstrate that learned robot policies lead to efficient task completion when collaborating with unseen humanoid agents and human partners that might exhibit behaviors that the robot has not seen before. Additionally, we observe emergent behaviors during collaborative task execution, such as the robot yielding space when obstructing a humanoid agent, thereby allowing the effective completion of the task by the humanoid agent. Furthermore, our experiments using the human-in-the-loop tool demonstrate that our automated evaluation with humanoids can provide an indication of the relative ordering of different policies when evaluated with real human collaborators. Habitat 3.0 unlocks interesting new features in simulators for Embodied AI, and we hope it paves the way for a new frontier of embodied human-AI interaction capabilities.

Theophile Gervet, Soumith Chintala, Dhruv Batra et al.

Semantic navigation is necessary to deploy mobile robots in uncontrolled environments like our homes, schools, and hospitals. Many learning-based approaches have been proposed in response to the lack of semantic understanding of the classical pipeline for spatial navigation, which builds a geometric map using depth sensors and plans to reach point goals. Broadly, end-to-end learning approaches reactively map sensor inputs to actions with deep neural networks, while modular learning approaches enrich the classical pipeline with learning-based semantic sensing and exploration. But learned visual navigation policies have predominantly been evaluated in simulation. How well do different classes of methods work on a robot? We present a large-scale empirical study of semantic visual navigation methods comparing representative methods from classical, modular, and end-to-end learning approaches across six homes with no prior experience, maps, or instrumentation. We find that modular learning works well in the real world, attaining a 90% success rate. In contrast, end-to-end learning does not, dropping from 77% simulation to 23% real-world success rate due to a large image domain gap between simulation and reality. For practitioners, we show that modular learning is a reliable approach to navigate to objects: modularity and abstraction in policy design enable Sim-to-Real transfer. For researchers, we identify two key issues that prevent today's simulators from being reliable evaluation benchmarks - (A) a large Sim-to-Real gap in images and (B) a disconnect between simulation and real-world error modes - and propose concrete steps forward.

Ilija Radosavovic, Tete Xiao, Bike Zhang et al.

Humanoid robots that can autonomously operate in diverse environments have the potential to help address labour shortages in factories, assist elderly at homes, and colonize new planets. While classical controllers for humanoid robots have shown impressive results in a number of settings, they are challenging to generalize and adapt to new environments. Here, we present a fully learning-based approach for real-world humanoid locomotion. Our controller is a causal transformer that takes the history of proprioceptive observations and actions as input and predicts the next action. We hypothesize that the observation-action history contains useful information about the world that a powerful transformer model can use to adapt its behavior in-context, without updating its weights. We train our model with large-scale model-free reinforcement learning on an ensemble of randomized environments in simulation and deploy it to the real world zero-shot. Our controller can walk over various outdoor terrains, is robust to external disturbances, and can adapt in context.

Ilija Radosavovic, Baifeng Shi, Letian Fu et al.

We present a self-supervised sensorimotor pre-training approach for robotics. Our model, called RPT, is a Transformer that operates on sequences of sensorimotor tokens. Given a sequence of camera images, proprioceptive robot states, and actions, we encode the sequence into tokens, mask out a subset, and train a model to predict the missing content from the rest. We hypothesize that if a robot can predict the masked-out content it will have acquired a good model of the physical world that can enable it to act. RPT is designed to operate on latent visual representations which makes prediction tractable, enables scaling to larger models, and allows fast inference on a real robot. To evaluate our approach, we collected a dataset of 20,000 real-world trajectories over 9 months using a combination of motion planning and grasping algorithms. We find that sensorimotor pre-training consistently outperforms training from scratch, has favorable scaling properties, and enables transfer across different tasks, environments, and robots.

Jacob Krantz, Theophile Gervet, Karmesh Yadav et al.

Images are a convenient way to specify which particular object instance an embodied agent should navigate to. Solving this task requires semantic visual reasoning and exploration of unknown environments. We present a system that can perform this task in both simulation and the real world. Our modular method solves sub-tasks of exploration, goal instance re-identification, goal localization, and local navigation. We re-identify the goal instance in egocentric vision using feature-matching and localize the goal instance by projecting matched features to a map. Each sub-task is solved using off-the-shelf components requiring zero fine-tuning. On the HM3D InstanceImageNav benchmark, this system outperforms a baseline end-to-end RL policy 7x and a state-of-the-art ImageNav model 2.3x (56% vs 25% success). We deploy this system to a mobile robot platform and demonstrate effective real-world performance, achieving an 88% success rate across a home and an office environment.

Vickie Ye, Georgios Pavlakos, Jitendra Malik et al.

We propose a method to reconstruct global human trajectories from videos in the wild. Our optimization method decouples the camera and human motion, which allows us to place people in the same world coordinate frame. Most existing methods do not model the camera motion; methods that rely on the background pixels to infer 3D human motion usually require a full scene reconstruction, which is often not possible for in-the-wild videos. However, even when existing SLAM systems cannot recover accurate scene reconstructions, the background pixel motion still provides enough signal to constrain the camera motion. We show that relative camera estimates along with data-driven human motion priors can resolve the scene scale ambiguity and recover global human trajectories. Our method robustly recovers the global 3D trajectories of people in challenging in-the-wild videos, such as PoseTrack. We quantify our improvement over existing methods on 3D human dataset Egobody. We further demonstrate that our recovered camera scale allows us to reason about motion of multiple people in a shared coordinate frame, which improves performance of downstream tracking in PoseTrack. Code and video results can be found at https://vye16.github.io/slahmr.

Jacob Krantz, Stefan Lee, Jitendra Malik et al.

We consider the problem of embodied visual navigation given an image-goal (ImageNav) where an agent is initialized in an unfamiliar environment and tasked with navigating to a location 'described' by an image. Unlike related navigation tasks, ImageNav does not have a standardized task definition which makes comparison across methods difficult. Further, existing formulations have two problematic properties; (1) image-goals are sampled from random locations which can lead to ambiguity (e.g., looking at walls), and (2) image-goals match the camera specification and embodiment of the agent; this rigidity is limiting when considering user-driven downstream applications. We present the Instance-specific ImageNav task (InstanceImageNav) to address these limitations. Specifically, the goal image is 'focused' on some particular object instance in the scene and is taken with camera parameters independent of the agent. We instantiate InstanceImageNav in the Habitat Simulator using scenes from the Habitat-Matterport3D dataset (HM3D) and release a standardized benchmark to measure community progress.

Chao-Yuan Wu, Justin Johnson, Jitendra Malik et al.

A central goal of visual recognition is to understand objects and scenes from a single image. 2D recognition has witnessed tremendous progress thanks to large-scale learning and general-purpose representations. Comparatively, 3D poses new challenges stemming from occlusions not depicted in the image. Prior works try to overcome these by inferring from multiple views or rely on scarce CAD models and category-specific priors which hinder scaling to novel settings. In this work, we explore single-view 3D reconstruction by learning generalizable representations inspired by advances in self-supervised learning. We introduce a simple framework that operates on 3D points of single objects or whole scenes coupled with category-agnostic large-scale training from diverse RGB-D videos. Our model, Multiview Compressive Coding (MCC), learns to compress the input appearance and geometry to predict the 3D structure by querying a 3D-aware decoder. MCC's generality and efficiency allow it to learn from large-scale and diverse data sources with strong generalization to novel objects imagined by DALL$\cdot$E 2 or captured in-the-wild with an iPhone.

Jiayuan Gu, Devendra Singh Chaplot, Hao Su et al.

We study a modular approach to tackle long-horizon mobile manipulation tasks for object rearrangement, which decomposes a full task into a sequence of subtasks. To tackle the entire task, prior work chains multiple stationary manipulation skills with a point-goal navigation skill, which are learned individually on subtasks. Although more effective than monolithic end-to-end RL policies, this framework suffers from compounding errors in skill chaining, e.g., navigating to a bad location where a stationary manipulation skill can not reach its target to manipulate. To this end, we propose that the manipulation skills should include mobility to have flexibility in interacting with the target object from multiple locations and at the same time the navigation skill could have multiple end points which lead to successful manipulation. We operationalize these ideas by implementing mobile manipulation skills rather than stationary ones and training a navigation skill trained with region goal instead of point goal. We evaluate our multi-skill mobile manipulation method M3 on 3 challenging long-horizon mobile manipulation tasks in the Home Assistant Benchmark (HAB), and show superior performance as compared to the baselines.

Antonio Loquercio, Ashish Kumar, Jitendra Malik

In this work, we show how to learn a visual walking policy that only uses a monocular RGB camera and proprioception. Since simulating RGB is hard, we necessarily have to learn vision in the real world. We start with a blind walking policy trained in simulation. This policy can traverse some terrains in the real world but often struggles since it lacks knowledge of the upcoming geometry. This can be resolved with the use of vision. We train a visual module in the real world to predict the upcoming terrain with our proposed algorithm Cross-Modal Supervision (CMS). CMS uses time-shifted proprioception to supervise vision and allows the policy to continually improve with more real-world experience. We evaluate our vision-based walking policy over a diverse set of terrains including stairs (up to 19cm high), slippery slopes (inclination of 35 degrees), curbs and tall steps (up to 20cm), and complex discrete terrains. We achieve this performance with less than 30 minutes of real-world data. Finally, we show that our policy can adapt to shifts in the visual field with a limited amount of real-world experience. Video results and code at https://antonilo.github.io/vision_locomotion/.

Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa et al.

In this work we study the benefits of using tracking and 3D poses for action recognition. To achieve this, we take the Lagrangian view on analysing actions over a trajectory of human motion rather than at a fixed point in space. Taking this stand allows us to use the tracklets of people to predict their actions. In this spirit, first we show the benefits of using 3D pose to infer actions, and study person-person interactions. Subsequently, we propose a Lagrangian Action Recognition model by fusing 3D pose and contextualized appearance over tracklets. To this end, our method achieves state-of-the-art performance on the AVA v2.2 dataset on both pose only settings and on standard benchmark settings. When reasoning about the action using only pose cues, our pose model achieves +10.0 mAP gain over the corresponding state-of-the-art while our fused model has a gain of +2.8 mAP over the best state-of-the-art model. Code and results are available at: https://brjathu.github.io/LART

Austin Patel, Andrew Wang, Ilija Radosavovic et al.

We study the problem of imitating object interactions from Internet videos. This requires understanding the hand-object interactions in 4D, spatially in 3D and over time, which is challenging due to mutual hand-object occlusions. In this paper we make two main contributions: (1) a novel reconstruction technique RHOV (Reconstructing Hands and Objects from Videos), which reconstructs 4D trajectories of both the hand and the object using 2D image cues and temporal smoothness constraints; (2) a system for imitating object interactions in a physics simulator with reinforcement learning. We apply our reconstruction technique to 100 challenging Internet videos. We further show that we can successfully imitate a range of different object interactions in a physics simulator. Our object-centric approach is not limited to human-like end-effectors and can learn to imitate object interactions using different embodiments, like a robotic arm with a parallel jaw gripper.

Yulu Gan, Ligeng Zhu, Dandan Shan et al.

Motion understanding is fundamental to physical reasoning, enabling models to infer dynamics and predict future states. However, state-of-the-art models still struggle on recent motion benchmarks, primarily due to the scarcity of large-scale, fine-grained motion datasets. Existing motion datasets are often constructed from costly manual annotation, severely limiting scalability. To address this challenge, we introduce FoundationMotion, a fully automated data curation pipeline that constructs large-scale motion datasets. Our approach first detects and tracks objects in videos to extract their trajectories, then leverages these trajectories and video frames with Large Language Models (LLMs) to generate fine-grained captions and diverse question-answer pairs about motion and spatial reasoning. Using datasets produced by this pipeline, we fine-tune open-source models including NVILA-Video-15B and Qwen2.5-7B, achieving substantial improvements in motion understanding without compromising performance on other tasks. Notably, our models outperform strong closed-source baselines like Gemini-2.5 Flash and large open-source models such as Qwen2.5-VL-72B across diverse motion understanding datasets and benchmarks. FoundationMotion thus provides a scalable solution for curating fine-grained motion datasets that enable effective fine-tuning of diverse models to enhance motion understanding and spatial reasoning capabilities.

Andrea Bajcsy, Antonio Loquercio, Ashish Kumar et al.

Learning strategic robot behavior -- like that required in pursuit-evasion interactions -- under real-world constraints is extremely challenging. It requires exploiting the dynamics of the interaction, and planning through both physical state and latent intent uncertainty. In this paper, we transform this intractable problem into a supervised learning problem, where a fully-observable robot policy generates supervision for a partially-observable one. We find that the quality of the supervision signal for the partially-observable pursuer policy depends on two key factors: the balance of diversity and optimality of the evader's behavior and the strength of the modeling assumptions in the fully-observable policy. We deploy our policy on a physical quadruped robot with an RGB-D camera on pursuit-evasion interactions in the wild. Despite all the challenges, the sensing constraints bring about creativity: the robot is pushed to gather information when uncertain, predict intent from noisy measurements, and anticipate in order to intercept. Project webpage: https://abajcsy.github.io/vision-based-pursuit/

Liang Heng, Haoran Geng, Kaifeng Zhang et al.

Dexterous manipulation is a cornerstone capability for robotic systems aiming to interact with the physical world in a human-like manner. Although vision-based methods have advanced rapidly, tactile sensing remains crucial for fine-grained control, particularly in unstructured or visually occluded settings. We present ViTacFormer, a representation-learning approach that couples a cross-attention encoder to fuse high-resolution vision and touch with an autoregressive tactile prediction head that anticipates future contact signals. Building on this architecture, we devise an easy-to-challenging curriculum that steadily refines the visual-tactile latent space, boosting both accuracy and robustness. The learned cross-modal representation drives imitation learning for multi-fingered hands, enabling precise and adaptive manipulation. Across a suite of challenging real-world benchmarks, our method achieves approximately 50% higher success rates than prior state-of-the-art systems. To our knowledge, it is also the first to autonomously complete long-horizon dexterous manipulation tasks that demand highly precise control with an anthropomorphic hand, successfully executing up to 11 sequential stages and sustaining continuous operation for 2.5 minutes.

Jessica Yin, Haozhi Qi, Youngsun Wi et al.

Human video demonstrations provide abundant training data for learning robot policies, but video alone cannot capture the rich contact signals critical for mastering manipulation. We introduce OSMO, an open-source wearable tactile glove designed for human-to-robot skill transfer. The glove features 12 three-axis tactile sensors across the fingertips and palm and is designed to be compatible with state-of-the-art hand-tracking methods for in-the-wild data collection. We demonstrate that a robot policy trained exclusively on human demonstrations collected with OSMO, without any real robot data, is capable of executing a challenging contact-rich manipulation task. By equipping both the human and the robot with the same glove, OSMO minimizes the visual and tactile embodiment gap, enabling the transfer of continuous shear and normal force feedback while avoiding the need for image inpainting or other vision-based force inference. On a real-world wiping task requiring sustained contact pressure, our tactile-aware policy achieves a 72% success rate, outperforming vision-only baselines by eliminating contact-related failure modes. We release complete hardware designs, firmware, and assembly instructions to support community adoption.

Vongani Maluleke, Lea Müller, Jathushan Rajasegaran et al.

This paper asks to what extent social interaction influences one's behavior. We study this in the setting of two dancers dancing as a couple. We first consider a baseline in which we predict a dancer's future moves conditioned only on their past motion without regard to their partner. We then investigate the advantage of taking social information into account by conditioning also on the motion of their dancing partner. We focus our analysis on Swing, a dance genre with tight physical coupling for which we present an in-the-wild video dataset. We demonstrate that single-person future motion prediction in this context is challenging. Instead, we observe that prediction greatly benefits from considering the interaction partners' behavior, resulting in surprisingly compelling couple dance synthesis results (see supp. video). Our contributions are a demonstration of the advantages of socially conditioned future motion prediction and an in-the-wild, couple dance video dataset to enable future research in this direction. Video results are available on the project website: https://von31.github.io/synNsync

Tanish Baranwal, Himanshu Gaurav Singh, Jathushan Rajasegaran et al.

We propose Video Gaussian Masked Autoencoders (Video-GMAE), a self-supervised approach for representation learning that encodes a sequence of images into a set of Gaussian splats moving over time. Representing a video as a set of Gaussians enforces a reasonable inductive bias: that 2-D videos are often consistent projections of a dynamic 3-D scene. We find that tracking emerges when pretraining a network with this architecture. Mapping the trajectory of the learnt Gaussians onto the image plane gives zero-shot tracking performance comparable to state-of-the-art. With small-scale finetuning, our models achieve 34.6% improvement on Kinetics, and 13.1% on Kubric datasets, surpassing existing self-supervised video approaches. The project page and code are publicly available at https://videogmae.org/ and https://github.com/tekotan/video-gmae.

Boyuan Chen, Tianyuan Zhang, Haoran Geng et al. · mit

General-purpose robots require decision-making models that generalize across diverse tasks and environments. Recent works build robot foundation models by extending multimodal large language models (MLLMs) with action outputs, creating vision-language-action (VLA) systems. These efforts are motivated by the intuition that MLLMs' large-scale language and image pretraining can be effectively transferred to the action output modality. In this work, we explore an alternative paradigm of using large-scale video pretraining as a primary modality for building robot foundation models. Unlike static images and language, videos capture spatio-temporal sequences of states and actions in the physical world that are naturally aligned with robotic behavior. We curate an internet-scale video dataset of human activities and task demonstrations, and train, for the first time at a foundation-model scale, an open video model for generative robotics planning. The model produces zero-shot video plans for novel scenes and tasks, which we post-process to extract executable robot actions. We evaluate task-level generalization through third-party selected tasks in the wild and real-robot experiments, demonstrating successful physical execution. Together, these results show robust instruction following, strong generalization, and real-world feasibility. We release both the model and dataset to support open, reproducible video-based robot learning. Our website is available at https://www.boyuan.space/large-video-planner/.

Himanshu Gaurav Singh, Antonio Loquercio, Carmelo Sferrazza et al.

We present an approach to learn general robot manipulation priors from 3D hand-object interaction trajectories. We build a framework to use in-the-wild videos to generate sensorimotor robot trajectories. We do so by lifting both the human hand and the manipulated object in a shared 3D space and retargeting human motions to robot actions. Generative modeling on this data gives us a task-agnostic base policy. This policy captures a general yet flexible manipulation prior. We empirically demonstrate that finetuning this policy, with both reinforcement learning (RL) and behavior cloning (BC), enables sample-efficient adaptation to downstream tasks and simultaneously improves robustness and generalizability compared to prior approaches. Qualitative experiments are available at: \url{https://hgaurav2k.github.io/hop/}.

Haozhi Qi, Yen-Jen Wang, Toru Lin et al.

Humanoid robots hold great promise for operating in human-centric environments, yet achieving robust whole-body coordination across the head, hands, and legs remains a major challenge. We present a system that combines a modular teleoperation interface with a scalable learning framework to address this problem. Our teleoperation design decomposes humanoid control into intuitive submodules, which include hand-eye coordination, grasp primitives, arm end-effector tracking, and locomotion. This modularity allows us to collect high-quality demonstrations efficiently. Building on this, we introduce Choice Policy, an imitation learning approach that generates multiple candidate actions and learns to score them. This architecture enables both fast inference and effective modeling of multimodal behaviors. We validate our approach on two real-world tasks: dishwasher loading and whole-body loco-manipulation for whiteboard wiping. Experiments show that Choice Policy significantly outperforms diffusion policies and standard behavior cloning. Furthermore, our results indicate that hand-eye coordination is critical for success in long-horizon tasks. Our work demonstrates a practical path toward scalable data collection and learning for coordinated humanoid manipulation in unstructured environments.

Jiahe Pan, Stelian Coros, Jitendra Malik et al.

A primary bottleneck in contact-rich manipulation is the difficulty of collecting real-world data. Sim-to-real reinforcement learning offers a scalable alternative, but the simulation-reality gap prevents information-dense modalities like touch from being effectively used. Existing sim-to-real methods often mitigate this gap by simplifying tactile data into coarse low-dimensional features -- sacrificing the richness required for complex manipulation. In this work, we introduce Center-of-Pressure (CoP), an effective tactile representation grounded in physical principles that preserves dense contact information while maintaining robustness for sim-to-real transfer. To support this representation, we propose a sensor calibration scheme based on differentiable dynamics, enabling the estimation of taxel orientations without requiring ground-truth force measurements. We evaluate CoP on two blind, challenging contact-rich manipulation tasks: peg-in-hole insertion and ball balancing. Across both tasks, policies conditioned on CoP achieve zero-shot sim-to-real transfer on a multi-fingered hand, and outperform both coarse binary-contact and raw-taxel baselines. Analysis of learned policy states further suggests that CoP-conditioned policies encode task-relevant physical properties, such as object mass, as an emergent byproduct of control.

Ran Tian, Chenfeng Xu, Masayoshi Tomizuka et al.

When operating in service of people, robots need to optimize rewards aligned with end-user preferences. Since robots will rely on raw perceptual inputs like RGB images, their rewards will inevitably use visual representations. Recently there has been excitement in using representations from pre-trained visual models, but key to making these work in robotics is fine-tuning, which is typically done via proxy tasks like dynamics prediction or enforcing temporal cycle-consistency. However, all these proxy tasks bypass the human's input on what matters to them, exacerbating spurious correlations and ultimately leading to robot behaviors that are misaligned with user preferences. In this work, we propose that robots should leverage human feedback to align their visual representations with the end-user and disentangle what matters for the task. We propose Representation-Aligned Preference-based Learning (RAPL), a method for solving the visual representation alignment problem and visual reward learning problem through the lens of preference-based learning and optimal transport. Across experiments in X-MAGICAL and in robotic manipulation, we find that RAPL's reward consistently generates preferred robot behaviors with high sample efficiency, and shows strong zero-shot generalization when the visual representation is learned from a different embodiment than the robot's.

Rosy Chen, Mustafa Mukadam, Michael Kaess et al.

Tactile dexterous manipulation is essential to automating complex household tasks, yet learning effective control policies remains a challenge. While recent work has relied on imitation learning, obtaining high quality demonstrations for multi-fingered hands via robot teleoperation or kinesthetic teaching is prohibitive. Alternatively, with reinforcement we can learn skills in simulation, but fast and realistic simulation of tactile observations is challenging. To bridge this gap, we introduce PTLD: sim-to-real Privileged Tactile Latent Distillation, a novel approach to learning tactile manipulation skills without requiring tactile simulation. Instead of simulating tactile sensors or relying purely on proprioceptive policies to transfer zero-shot sim-to-real, our key idea is to leverage privileged sensors in the real world to collect real-world tactile policy data. This data is then used to distill a robust state estimator that operates on tactile input. We demonstrate from our experiments that PTLD can be used to improve proprioceptive manipulation policies trained in simulation significantly by incorporating tactile sensing. On the benchmark in-hand rotation task, PTLD achieves a 182% improvement over a proprioception only policy. We also show that PTLD enables learning the challenging task of tactile in-hand reorientation where we see a 57% improvement in the number of goals reached over using proprioception alone. Website: https://akashsharma02.github.io/ptld-website/.

Boyi Li, Ligeng Zhu, Ran Tian et al.

We propose Wolf, a WOrLd summarization Framework for accurate video captioning. Wolf is an automated captioning framework that adopts a mixture-of-experts approach, leveraging complementary strengths of Vision Language Models (VLMs). By utilizing both image and video models, our framework captures different levels of information and summarizes them efficiently. Our approach can be applied to enhance video understanding, auto-labeling, and captioning. To evaluate caption quality, we introduce CapScore, an LLM-based metric to assess the similarity and quality of generated captions compared to the ground truth captions. We further build four human-annotated datasets in three domains: autonomous driving, general scenes, and robotics, to facilitate comprehensive comparisons. We show that Wolf achieves superior captioning performance compared to state-of-the-art approaches from the research community (VILA1.5, CogAgent) and commercial solutions (Gemini-Pro-1.5, GPT-4V). For instance, in comparison with GPT-4V, Wolf improves CapScore both quality-wise by 55.6% and similarity-wise by 77.4% on challenging driving videos. Finally, we establish a benchmark for video captioning and introduce a leaderboard, aiming to accelerate advancements in video understanding, captioning, and data alignment. Webpage: https://wolfv0.github.io/.

Jun Wang, Ying Yuan, Haichuan Che et al.

In-hand manipulation of pen-like objects is an important skill in our daily lives, as many tools such as hammers and screwdrivers are similarly shaped. However, current learning-based methods struggle with this task due to a lack of high-quality demonstrations and the significant gap between simulation and the real world. In this work, we push the boundaries of learning-based in-hand manipulation systems by demonstrating the capability to spin pen-like objects. We first use reinforcement learning to train an oracle policy with privileged information and generate a high-fidelity trajectory dataset in simulation. This serves two purposes: 1) pre-training a sensorimotor policy in simulation; 2) conducting open-loop trajectory replay in the real world. We then fine-tune the sensorimotor policy using these real-world trajectories to adapt it to the real world dynamics. With less than 50 trajectories, our policy learns to rotate more than ten pen-like objects with different physical properties for multiple revolutions. We present a comprehensive analysis of our design choices and share the lessons learned during development.

Jessica Yin, Haozhi Qi, Jitendra Malik et al.

Recent progress in reinforcement learning (RL) and tactile sensing has significantly advanced dexterous manipulation. However, these methods often utilize simplified tactile signals due to the gap between tactile simulation and the real world. We introduce a sensor model for tactile skin that enables zero-shot sim-to-real transfer of ternary shear and binary normal forces. Using this model, we develop an RL policy that leverages sliding contact for dexterous in-hand translation. We conduct extensive real-world experiments to assess how tactile sensing facilitates policy adaptation to various unseen object properties and robot hand orientations. We demonstrate that our 3-axis tactile policies consistently outperform baselines that use only shear forces, only normal forces, or only proprioception. Website: https://jessicayin.github.io/tactile-skin-rl/

Du Tran, Jitendra Malik

We propose a new task of space-time semantic correspondence prediction in videos. Given a source video, a target video, and a set of space-time key-points in the source video, the task requires predicting a set of keypoints in the target video that are the semantic correspondences of the provided source keypoints. We believe that this task is important for fine-grain video understanding, potentially enabling applications such as activity coaching, sports analysis, robot imitation learning, and more. Our contributions in this paper are: (i) proposing a new task and providing annotations for space-time semantic correspondences on two existing benchmarks: Penn Action and Pouring; and (ii) presenting a comprehensive set of baselines and experiments to gain insights about the new problem. Our main finding is that the space-time semantic correspondence prediction problem is best approached jointly in space and time rather than in their decomposed sub-problems: time alignment and spatial correspondences.