Xiaolong Wang

Hanzhe Hu, Yinbo Chen, Jiarui Xu et al.

Integrating high-level context information with low-level details is of central importance in semantic segmentation. Towards this end, most existing segmentation models apply bilinear up-sampling and convolutions to feature maps of different scales, and then align them at the same resolution. However, bilinear up-sampling blurs the precise information learned in these feature maps and convolutions incur extra computation costs. To address these issues, we propose the Implicit Feature Alignment function (IFA). Our method is inspired by the rapidly expanding topic of implicit neural representations, where coordinate-based neural networks are used to designate fields of signals. In IFA, feature vectors are viewed as representing a 2D field of information. Given a query coordinate, nearby feature vectors with their relative coordinates are taken from the multi-level feature maps and then fed into an MLP to generate the corresponding output. As such, IFA implicitly aligns the feature maps at different levels and is capable of producing segmentation maps in arbitrary resolutions. We demonstrate the efficacy of IFA on multiple datasets, including Cityscapes, PASCAL Context, and ADE20K. Our method can be combined with improvement on various architectures, and it achieves state-of-the-art computation-accuracy trade-off on common benchmarks. Code will be made available at https://github.com/hzhupku/IFA.

Renhao Wang, Yu Sun, Arnuv Tandon et al. · berkeley, ibm-research

Prior work has established Test-Time Training (TTT) as a general framework to further improve a trained model at test time. Before making a prediction on each test instance, the model is first trained on the same instance using a self-supervised task such as reconstruction. We extend TTT to the streaming setting, where multiple test instances - video frames in our case - arrive in temporal order. Our extension is online TTT: The current model is initialized from the previous model, then trained on the current frame and a small window of frames immediately before. Online TTT significantly outperforms the fixed-model baseline for four tasks, on three real-world datasets. The improvements are more than 2.2x and 1.5x for instance and panoptic segmentation. Surprisingly, online TTT also outperforms its offline variant that accesses strictly more information, training on all frames from the entire test video regardless of temporal order. This finding challenges those in prior work using synthetic videos. We formalize a notion of locality as the advantage of online over offline TTT, and analyze its role with ablations and a theory based on bias-variance trade-off.

Jiarui Xu, Sifei Liu, Arash Vahdat et al.

We present ODISE: Open-vocabulary DIffusion-based panoptic SEgmentation, which unifies pre-trained text-image diffusion and discriminative models to perform open-vocabulary panoptic segmentation. Text-to-image diffusion models have the remarkable ability to generate high-quality images with diverse open-vocabulary language descriptions. This demonstrates that their internal representation space is highly correlated with open concepts in the real world. Text-image discriminative models like CLIP, on the other hand, are good at classifying images into open-vocabulary labels. We leverage the frozen internal representations of both these models to perform panoptic segmentation of any category in the wild. Our approach outperforms the previous state of the art by significant margins on both open-vocabulary panoptic and semantic segmentation tasks. In particular, with COCO training only, our method achieves 23.4 PQ and 30.0 mIoU on the ADE20K dataset, with 8.3 PQ and 7.9 mIoU absolute improvement over the previous state of the art. We open-source our code and models at https://github.com/NVlabs/ODISE .

Yuzhe Qin, Wei Yang, Binghao Huang et al. · nvidia

Vision-based teleoperation offers the possibility to endow robots with human-level intelligence to physically interact with the environment, while only requiring low-cost camera sensors. However, current vision-based teleoperation systems are designed and engineered towards a particular robot model and deploy environment, which scales poorly as the pool of the robot models expands and the variety of the operating environment increases. In this paper, we propose AnyTeleop, a unified and general teleoperation system to support multiple different arms, hands, realities, and camera configurations within a single system. Although being designed to provide great flexibility to the choice of simulators and real hardware, our system can still achieve great performance. For real-world experiments, AnyTeleop can outperform a previous system that was designed for a specific robot hardware with a higher success rate, using the same robot. For teleoperation in simulation, AnyTeleop leads to better imitation learning performance, compared with a previous system that is particularly designed for that simulator. Project page: https://yzqin.github.io/anyteleop/.

Yuzhuang Xu, Shuo Wang, Peng Li et al. · ibm-research, tsinghua

Communication games, which we refer to as incomplete information games that heavily depend on natural language communication, hold significant research value in fields such as economics, social science, and artificial intelligence. In this work, we explore the problem of how to engage large language models (LLMs) in communication games, and in response, propose a tuning-free framework. Our approach keeps LLMs frozen, and relies on the retrieval and reflection on past communications and experiences for improvement. An empirical study on the representative and widely-studied communication game, ``Werewolf'', demonstrates that our framework can effectively play Werewolf game without tuning the parameters of the LLMs. More importantly, strategic behaviors begin to emerge in our experiments, suggesting that it will be a fruitful journey to engage LLMs in communication games and associated domains.

Zeyuan Chen, Yinbo Chen, Jingwen Liu et al. · gatech, ibm-research

Videos typically record the streaming and continuous visual data as discrete consecutive frames. Since the storage cost is expensive for videos of high fidelity, most of them are stored in a relatively low resolution and frame rate. Recent works of Space-Time Video Super-Resolution (STVSR) are developed to incorporate temporal interpolation and spatial super-resolution in a unified framework. However, most of them only support a fixed up-sampling scale, which limits their flexibility and applications. In this work, instead of following the discrete representations, we propose Video Implicit Neural Representation (VideoINR), and we show its applications for STVSR. The learned implicit neural representation can be decoded to videos of arbitrary spatial resolution and frame rate. We show that VideoINR achieves competitive performances with state-of-the-art STVSR methods on common up-sampling scales and significantly outperforms prior works on continuous and out-of-training-distribution scales. Our project page is at http://zeyuan-chen.com/VideoINR/ .

Runsen Xu, Xiaolong Wang, Tai Wang et al.

The unprecedented advancements in Large Language Models (LLMs) have shown a profound impact on natural language processing but are yet to fully embrace the realm of 3D understanding. This paper introduces PointLLM, a preliminary effort to fill this gap, enabling LLMs to understand point clouds and offering a new avenue beyond 2D visual data. PointLLM understands colored object point clouds with human instructions and generates contextually appropriate responses, illustrating its grasp of point clouds and common sense. Specifically, it leverages a point cloud encoder with a powerful LLM to effectively fuse geometric, appearance, and linguistic information. We collect a novel dataset comprising 660K simple and 70K complex point-text instruction pairs to enable a two-stage training strategy: aligning latent spaces and subsequently instruction-tuning the unified model. To rigorously evaluate the perceptual and generalization capabilities of PointLLM, we establish two benchmarks: Generative 3D Object Classification and 3D Object Captioning, assessed through three different methods, including human evaluation, GPT-4/ChatGPT evaluation, and traditional metrics. Experimental results reveal PointLLM's superior performance over existing 2D and 3D baselines, with a notable achievement in human-evaluated object captioning tasks where it surpasses human annotators in over 50% of the samples. Codes, datasets, and benchmarks are available at https://github.com/OpenRobotLab/PointLLM .

Shaowei Liu, Subarna Tripathi, Somdeb Majumdar et al. · ibm-research

We propose to forecast future hand-object interactions given an egocentric video. Instead of predicting action labels or pixels, we directly predict the hand motion trajectory and the future contact points on the next active object (i.e., interaction hotspots). This relatively low-dimensional representation provides a concrete description of future interactions. To tackle this task, we first provide an automatic way to collect trajectory and hotspots labels on large-scale data. We then use this data to train an Object-Centric Transformer (OCT) model for prediction. Our model performs hand and object interaction reasoning via the self-attention mechanism in Transformers. OCT also provides a probabilistic framework to sample the future trajectory and hotspots to handle uncertainty in prediction. We perform experiments on the Epic-Kitchens-55, Epic-Kitchens-100, and EGTEA Gaze+ datasets, and show that OCT significantly outperforms state-of-the-art approaches by a large margin. Project page is available at https://stevenlsw.github.io/hoi-forecast .

Xuanchi Ren, Xiaolong Wang · ibm-research

Novel view synthesis from a single image has recently attracted a lot of attention, and it has been primarily advanced by 3D deep learning and rendering techniques. However, most work is still limited by synthesizing new views within relatively small camera motions. In this paper, we propose a novel approach to synthesize a consistent long-term video given a single scene image and a trajectory of large camera motions. Our approach utilizes an autoregressive Transformer to perform sequential modeling of multiple frames, which reasons the relations between multiple frames and the corresponding cameras to predict the next frame. To facilitate learning and ensure consistency among generated frames, we introduce a locality constraint based on the input cameras to guide self-attention among a large number of patches across space and time. Our method outperforms state-of-the-art view synthesis approaches by a large margin, especially when synthesizing long-term future in indoor 3D scenes. Project page at https://xrenaa.github.io/look-outside-room/.

Weitao Li, Hao Zhou, Xuanyu Lei et al.

Reinforcement learning with verifiable rewards (RLVR) has become central to LLM reasoning, but its outcome-level rewards can make models more willing to give confident answers when evidence or reasoning is unreliable. Existing SFT or RL methods mainly teach LLMs to refuse or express uncertainty at the response level, which can overfit abstention behavior rather than improve reasoning reliability. To address this limitation, we propose Cognitive Pairwise Training (CPT), a cognitive mid-training alignment stage that turns pairwise comparisons over reasoning traces into a reusable alignment signal. By learning to distinguish trustworthy from flawed reasoning, CPT encourages the model to internalize a reasoning-quality discrimination boundary rather than memorize surface refusal patterns. Across five model scales and three model families, CPT improves the reasoning--metacognition trade-off. At 14B, CPT+RL outperforms the standard SFT+RL pipeline by +2.2 math-average points and +5.2 abstention-F1 points. Further analyses show that CPT improves trace quality and exhibits strong robustness and scalability across evaluation and training settings. Code and models are released at https://github.com/Tsinghua-dhy/CPT.

Yang Fu, Ishan Misra, Xiaolong Wang · ibm-research

We propose a generalizable neural radiance fields - MonoNeRF, that can be trained on large-scale monocular videos of moving in static scenes without any ground-truth annotations of depth and camera poses. MonoNeRF follows an Autoencoder-based architecture, where the encoder estimates the monocular depth and the camera pose, and the decoder constructs a Multiplane NeRF representation based on the depth encoder feature, and renders the input frames with the estimated camera. The learning is supervised by the reconstruction error. Once the model is learned, it can be applied to multiple applications including depth estimation, camera pose estimation, and single-image novel view synthesis. More qualitative results are available at: https://oasisyang.github.io/mononerf .

Xuxin Cheng, Jialong Li, Shiqi Yang et al.

Teleoperation serves as a powerful method for collecting on-robot data essential for robot learning from demonstrations. The intuitiveness and ease of use of the teleoperation system are crucial for ensuring high-quality, diverse, and scalable data. To achieve this, we propose an immersive teleoperation system Open-TeleVision that allows operators to actively perceive the robot's surroundings in a stereoscopic manner. Additionally, the system mirrors the operator's arm and hand movements on the robot, creating an immersive experience as if the operator's mind is transmitted to a robot embodiment. We validate the effectiveness of our system by collecting data and training imitation learning policies on four long-horizon, precise tasks (Can Sorting, Can Insertion, Folding, and Unloading) for 2 different humanoid robots and deploy them in the real world. The system is open-sourced at: https://robot-tv.github.io/

Mert Yuksekgonul, Daniel Koceja, Xinhao Li et al. · stanford

How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average, and to solve this very problem rather than generalize to other problems. Therefore, our learning objective and search subroutine are designed to prioritize the most promising solutions. We call this method Test-Time Training to Discover (TTT-Discover). Following prior work, we focus on problems with continuous rewards. We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) Erdős' minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to $2\times$ faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers. All our results are achieved with an open model, OpenAI gpt-oss-120b, and can be reproduced with our publicly available code, in contrast to previous best results that required closed frontier models. Our test-time training runs are performed using Tinker, an API by Thinking Machines, with a cost of only a few hundred dollars per problem.

Nicklas Hansen, Xiaolong Wang, Hao Su

Data-driven model predictive control has two key advantages over model-free methods: a potential for improved sample efficiency through model learning, and better performance as computational budget for planning increases. However, it is both costly to plan over long horizons and challenging to obtain an accurate model of the environment. In this work, we combine the strengths of model-free and model-based methods. We use a learned task-oriented latent dynamics model for local trajectory optimization over a short horizon, and use a learned terminal value function to estimate long-term return, both of which are learned jointly by temporal difference learning. Our method, TD-MPC, achieves superior sample efficiency and asymptotic performance over prior work on both state and image-based continuous control tasks from DMControl and Meta-World. Code and video results are available at https://nicklashansen.github.io/td-mpc.

Adrian Remonda, Nicklas Hansen, Ayoub Raji et al.

Despite the availability of international prize-money competitions, scaled vehicles, and simulation environments, research on autonomous racing and the control of sports cars operating close to the limit of handling has been limited by the high costs of vehicle acquisition and management, as well as the limited physics accuracy of open-source simulators. In this paper, we propose a racing simulation platform based on the simulator Assetto Corsa to test, validate, and benchmark autonomous driving algorithms, including reinforcement learning (RL) and classical Model Predictive Control (MPC), in realistic and challenging scenarios. Our contributions include the development of this simulation platform, several state-of-the-art algorithms tailored to the racing environment, and a comprehensive dataset collected from human drivers. Additionally, we evaluate algorithms in the offline RL setting. All the necessary code (including environment and benchmarks), working examples, datasets, and videos are publicly released and can be found at: https://assetto-corsa-gym.github.io

Yu Zhao, Yunxin Li, Yuxiang Wu et al.

Medical dialogue generation is an important yet challenging task. Most previous works rely on the attention mechanism and large-scale pretrained language models. However, these methods often fail to acquire pivotal information from the long dialogue history to yield an accurate and informative response, due to the fact that the medical entities usually scatters throughout multiple utterances along with the complex relationships between them. To mitigate this problem, we propose a medical response generation model with Pivotal Information Recalling (MedPIR), which is built on two components, i.e., knowledge-aware dialogue graph encoder and recall-enhanced generator. The knowledge-aware dialogue graph encoder constructs a dialogue graph by exploiting the knowledge relationships between entities in the utterances, and encodes it with a graph attention network. Then, the recall-enhanced generator strengthens the usage of these pivotal information by generating a summary of the dialogue before producing the actual response. Experimental results on two large-scale medical dialogue datasets show that MedPIR outperforms the strong baselines in BLEU scores and medical entities F1 measure.

Yu Sun, Xinhao Li, Karan Dalal et al.

Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, but their performance in long context is limited by the expressive power of their hidden states. We present a practical framework for instantiating sequence modeling layers with linear complexity and expressive hidden states. The key idea is to make the hidden state a machine learning model itself, and the update rule a step of self-supervised learning. Since the hidden state is updated by training even on test sequences, our layers are called Test-Time Training (TTT) layers. We consider two instantiations: TTT-Linear and TTT-MLP, whose hidden state is a linear model and a two-layer MLP respectively. We evaluate our instantiations at the scale of 125M to 1.3B parameters, comparing with a strong Transformer and Mamba, a modern RNN. Similar to Transformer, TTT-Linear and TTT-MLP can keep reducing perplexity by conditioning on more tokens, while Mamba cannot after 16k context. TTT-MLP still faces challenges in memory I/O, but shows larger potential in long context, pointing to a promising direction for future research.

Jiteng Mu, Shen Sang, Nuno Vasconcelos et al.

While NeRF-based human representations have shown impressive novel view synthesis results, most methods still rely on a large number of images / views for training. In this work, we propose a novel animatable NeRF called ActorsNeRF. It is first pre-trained on diverse human subjects, and then adapted with few-shot monocular video frames for a new actor with unseen poses. Building on previous generalizable NeRFs with parameter sharing using a ConvNet encoder, ActorsNeRF further adopts two human priors to capture the large human appearance, shape, and pose variations. Specifically, in the encoded feature space, we will first align different human subjects in a category-level canonical space, and then align the same human from different frames in an instance-level canonical space for rendering. We quantitatively and qualitatively demonstrate that ActorsNeRF significantly outperforms the existing state-of-the-art on few-shot generalization to new people and poses on multiple datasets. Project Page: https://jitengmu.github.io/ActorsNeRF/

Kaifeng Zhang, Yang Fu, Shubhankar Borse et al.

While 6D object pose estimation has wide applications across computer vision and robotics, it remains far from being solved due to the lack of annotations. The problem becomes even more challenging when moving to category-level 6D pose, which requires generalization to unseen instances. Current approaches are restricted by leveraging annotations from simulation or collected from humans. In this paper, we overcome this barrier by introducing a self-supervised learning approach trained directly on large-scale real-world object videos for category-level 6D pose estimation in the wild. Our framework reconstructs the canonical 3D shape of an object category and learns dense correspondences between input images and the canonical shape via surface embedding. For training, we propose novel geometrical cycle-consistency losses which construct cycles across 2D-3D spaces, across different instances and different time steps. The learned correspondence can be applied for 6D pose estimation and other downstream tasks such as keypoint transfer. Surprisingly, our method, without any human annotations or simulators, can achieve on-par or even better performance than previous supervised or semi-supervised methods on in-the-wild images. Our project page is: https://kywind.github.io/self-pose .

Yanjie Ze, Ge Yan, Yueh-Hua Wu et al.

It is a long-standing problem in robotics to develop agents capable of executing diverse manipulation tasks from visual observations in unstructured real-world environments. To achieve this goal, the robot needs to have a comprehensive understanding of the 3D structure and semantics of the scene. In this work, we present $\textbf{GNFactor}$, a visual behavior cloning agent for multi-task robotic manipulation with $\textbf{G}$eneralizable $\textbf{N}$eural feature $\textbf{F}$ields. GNFactor jointly optimizes a generalizable neural field (GNF) as a reconstruction module and a Perceiver Transformer as a decision-making module, leveraging a shared deep 3D voxel representation. To incorporate semantics in 3D, the reconstruction module utilizes a vision-language foundation model ($\textit{e.g.}$, Stable Diffusion) to distill rich semantic information into the deep 3D voxel. We evaluate GNFactor on 3 real robot tasks and perform detailed ablations on 10 RLBench tasks with a limited number of demonstrations. We observe a substantial improvement of GNFactor over current state-of-the-art methods in seen and unseen tasks, demonstrating the strong generalization ability of GNFactor. Our project website is https://yanjieze.com/GNFactor/ .

Yuzhe Qin, Hao Su, Xiaolong Wang

We propose to perform imitation learning for dexterous manipulation with multi-finger robot hand from human demonstrations, and transfer the policy to the real robot hand. We introduce a novel single-camera teleoperation system to collect the 3D demonstrations efficiently with only an iPad and a computer. One key contribution of our system is that we construct a customized robot hand for each user in the physical simulator, which is a manipulator resembling the same kinematics structure and shape of the operator's hand. This provides an intuitive interface and avoid unstable human-robot hand retargeting for data collection, leading to large-scale and high quality data. Once the data is collected, the customized robot hand trajectories can be converted to different specified robot hands (models that are manufactured) to generate training demonstrations. With imitation learning using our data, we show large improvement over baselines with multiple complex manipulation tasks. Importantly, we show our learned policy is significantly more robust when transferring to the real robot. More videos can be found in the https://yzqin.github.io/dex-teleop-imitation .

Kehlani Fay, Darin Anthony Djapri, Anya Zorin et al.

Dexterous manipulation is limited by both control and design, without consensus as to what makes manipulators best for performing dexterous tasks. This raises a fundamental challenge: how should we design and control robot manipulators that are optimized for dexterity? We present a co-design framework that learns task-specific hand morphology and complementary dexterous control policies. The framework supports 1) an expansive morphology search space including joint, finger, and palm generation, 2) scalable evaluation across the wide design space via morphology-conditioned cross-embodied control, and 3) real-world fabrication with accessible components. We evaluate the approach across multiple dexterous tasks, including in-hand rotation with simulation and real deployment. Our framework enables an end-to-end pipeline that can design, train, fabricate, and deploy a new robotic hand in under 24 hours. The full framework will be open-sourced and available on our website.

Yutong Liang, Quanquan Peng, Ri-Zhao Qiu et al.

Human demonstrations provide strong priors for robot manipulation, yet it is non-trivial to transfer them to execute on real robots due to the kinematic gap. In dexterous manipulation, it remains challenging to track long-horizon, contact-rich sequences even in simulators: a reference-tracking policy must keep objects on their target trajectories while preserving demonstrated joint motion and contact timing. Existing approaches often rely on hand-crafted reward tuning that require per-sequence tuning and break under limited interaction budgets. We introduce ConTrack, a reinforcement learning (RL) framework that scales with tracking data. ConTrack treats object tracking as a constraint and allocates remaining control authority to motion fidelity, which allows it to adapt task--style trade-offs online using a dual-variable update. In addition, ConTrack also stabilizes long-horizon learning with an adaptive mid-trajectory reset library that reuses policy-reachable simulator states. Our qualitative and quantitative results in simulation tracking and real robot demonstrate that ConTrack improves success and object pose accuracy significantly over prior arts while preserving joint and contact fidelity. Website: https://www.lyt0112.com/projects/ConTrack.

Yuzhe Qin, Binghao Huang, Zhao-Heng Yin et al.

We propose a sim-to-real framework for dexterous manipulation which can generalize to new objects of the same category in the real world. The key of our framework is to train the manipulation policy with point cloud inputs and dexterous hands. We propose two new techniques to enable joint learning on multiple objects and sim-to-real generalization: (i) using imagined hand point clouds as augmented inputs; and (ii) designing novel contact-based rewards. We empirically evaluate our method using an Allegro Hand to grasp novel objects in both simulation and real world. To the best of our knowledge, this is the first policy learning-based framework that achieves such generalization results with dexterous hands. Our project page is available at https://yzqin.github.io/dexpoint

Yang Fu, Xiaolong Wang

6D object pose estimation is one of the fundamental problems in computer vision and robotics research. While a lot of recent efforts have been made on generalizing pose estimation to novel object instances within the same category, namely category-level 6D pose estimation, it is still restricted in constrained environments given the limited number of annotated data. In this paper, we collect Wild6D, a new unlabeled RGBD object video dataset with diverse instances and backgrounds. We utilize this data to generalize category-level 6D object pose estimation in the wild with semi-supervised learning. We propose a new model, called Rendering for Pose estimation network RePoNet, that is jointly trained using the free ground-truths with the synthetic data, and a silhouette matching objective function on the real-world data. Without using any 3D annotations on real data, our method outperforms state-of-the-art methods on the previous dataset and our Wild6D test set (with manual annotations for evaluation) by a large margin. Project page with Wild6D data: https://oasisyang.github.io/semi-pose .

Jianglong Ye, Yuntao Chen, Naiyan Wang et al.

Recent development of neural implicit function has shown tremendous success on high-quality 3D shape reconstruction. However, most works divide the space into inside and outside of the shape, which limits their representing power to single-layer and watertight shapes. This limitation leads to tedious data processing (converting non-watertight raw data to watertight) as well as the incapability of representing general object shapes in the real world. In this work, we propose a novel method to represent general shapes including non-watertight shapes and shapes with multi-layer surfaces. We introduce General Implicit Function for 3D Shape (GIFS), which models the relationships between every two points instead of the relationships between points and surfaces. Instead of dividing 3D space into predefined inside-outside regions, GIFS encodes whether two points are separated by any surface. Experiments on ShapeNet show that GIFS outperforms previous state-of-the-art methods in terms of reconstruction quality, rendering efficiency, and visual fidelity. Project page is available at https://jianglongye.com/gifs .

Yueh-Hua Wu, Jiashun Wang, Xiaolong Wang

Dexterous manipulation with a multi-finger hand is one of the most challenging problems in robotics. While recent progress in imitation learning has largely improved the sample efficiency compared to Reinforcement Learning, the learned policy can hardly generalize to manipulate novel objects, given limited expert demonstrations. In this paper, we propose to learn dexterous manipulation using large-scale demonstrations with diverse 3D objects in a category, which are generated from a human grasp affordance model. This generalizes the policy to novel object instances within the same category. To train the policy, we propose a novel imitation learning objective jointly with a geometric representation learning objective using our demonstrations. By experimenting with relocating diverse objects in simulation, we show that our approach outperforms baselines with a large margin when manipulating novel objects. We also ablate the importance on 3D object representation learning for manipulation. We include videos, code, and additional information on the project website - https://kristery.github.io/ILAD/ .

Nicklas Hansen, Zhecheng Yuan, Yanjie Ze et al.

In this paper, we examine the effectiveness of pre-training for visuo-motor control tasks. We revisit a simple Learning-from-Scratch (LfS) baseline that incorporates data augmentation and a shallow ConvNet, and find that this baseline is surprisingly competitive with recent approaches (PVR, MVP, R3M) that leverage frozen visual representations trained on large-scale vision datasets -- across a variety of algorithms, task domains, and metrics in simulation and on a real robot. Our results demonstrate that these methods are hindered by a significant domain gap between the pre-training datasets and current benchmarks for visuo-motor control, which is alleviated by finetuning. Based on our findings, we provide recommendations for future research in pre-training for control and hope that our simple yet strong baseline will aid in accurately benchmarking progress in this area.

Yanjie Ze, Nicklas Hansen, Yinbo Chen et al.

A prominent approach to visual Reinforcement Learning (RL) is to learn an internal state representation using self-supervised methods, which has the potential benefit of improved sample-efficiency and generalization through additional learning signal and inductive biases. However, while the real world is inherently 3D, prior efforts have largely been focused on leveraging 2D computer vision techniques as auxiliary self-supervision. In this work, we present a unified framework for self-supervised learning of 3D representations for motor control. Our proposed framework consists of two phases: a pretraining phase where a deep voxel-based 3D autoencoder is pretrained on a large object-centric dataset, and a finetuning phase where the representation is jointly finetuned together with RL on in-domain data. We empirically show that our method enjoys improved sample efficiency in simulated manipulation tasks compared to 2D representation learning methods. Additionally, our learned policies transfer zero-shot to a real robot setup with only approximate geometric correspondence, and successfully solve motor control tasks that involve grasping and lifting from a single, uncalibrated RGB camera. Code and videos are available at https://yanjieze.com/3d4rl/ .

Jianglong Ye, Naiyan Wang, Xiaolong Wang

Recent works on generalizable NeRFs have shown promising results on novel view synthesis from single or few images. However, such models have rarely been applied on other downstream tasks beyond synthesis such as semantic understanding and parsing. In this paper, we propose a novel framework named FeatureNeRF to learn generalizable NeRFs by distilling pre-trained vision foundation models (e.g., DINO, Latent Diffusion). FeatureNeRF leverages 2D pre-trained foundation models to 3D space via neural rendering, and then extract deep features for 3D query points from NeRF MLPs. Consequently, it allows to map 2D images to continuous 3D semantic feature volumes, which can be used for various downstream tasks. We evaluate FeatureNeRF on tasks of 2D/3D semantic keypoint transfer and 2D/3D object part segmentation. Our extensive experiments demonstrate the effectiveness of FeatureNeRF as a generalizable 3D semantic feature extractor. Our project page is available at https://jianglongye.com/featurenerf/ .

Nicklas Hansen, Yixin Lin, Hao Su et al.

Poor sample efficiency continues to be the primary challenge for deployment of deep Reinforcement Learning (RL) algorithms for real-world applications, and in particular for visuo-motor control. Model-based RL has the potential to be highly sample efficient by concurrently learning a world model and using synthetic rollouts for planning and policy improvement. However, in practice, sample-efficient learning with model-based RL is bottlenecked by the exploration challenge. In this work, we find that leveraging just a handful of demonstrations can dramatically improve the sample-efficiency of model-based RL. Simply appending demonstrations to the interaction dataset, however, does not suffice. We identify key ingredients for leveraging demonstrations in model learning -- policy pretraining, targeted exploration, and oversampling of demonstration data -- which forms the three phases of our model-based RL framework. We empirically study three complex visuo-motor control domains and find that our method is 150%-250% more successful in completing sparse reward tasks compared to prior approaches in the low data regime (100K interaction steps, 5 demonstrations). Code and videos are available at: https://nicklashansen.github.io/modemrl

Sateesh Kumar, Jonathan Zamora, Nicklas Hansen et al.

Research on Inverse Reinforcement Learning (IRL) from third-person videos has shown encouraging results on removing the need for manual reward design for robotic tasks. However, most prior works are still limited by training from a relatively restricted domain of videos. In this paper, we argue that the true potential of third-person IRL lies in increasing the diversity of videos for better scaling. To learn a reward function from diverse videos, we propose to perform graph abstraction on the videos followed by temporal matching in the graph space to measure the task progress. Our insight is that a task can be described by entity interactions that form a graph, and this graph abstraction can help remove irrelevant information such as textures, resulting in more robust reward functions. We evaluate our approach, GraphIRL, on cross-embodiment learning in X-MAGICAL and learning from human demonstrations for real-robot manipulation. We show significant improvements in robustness to diverse video demonstrations over previous approaches, and even achieve better results than manual reward design on a real robot pushing task. Videos are available at https://sateeshkumar21.github.io/GraphIRL .

Nicklas Hansen, Hao Su, Xiaolong Wang

TD-MPC is a model-based reinforcement learning (RL) algorithm that performs local trajectory optimization in the latent space of a learned implicit (decoder-free) world model. In this work, we present TD-MPC2: a series of improvements upon the TD-MPC algorithm. We demonstrate that TD-MPC2 improves significantly over baselines across 104 online RL tasks spanning 4 diverse task domains, achieving consistently strong results with a single set of hyperparameters. We further show that agent capabilities increase with model and data size, and successfully train a single 317M parameter agent to perform 80 tasks across multiple task domains, embodiments, and action spaces. We conclude with an account of lessons, opportunities, and risks associated with large TD-MPC2 agents. Explore videos, models, data, code, and more at https://tdmpc2.com

Chenhongyi Yang, Jiarui Xu, Shalini De Mello et al.

We present the Group Propagation Vision Transformer (GPViT): a novel nonhierarchical (i.e. non-pyramidal) transformer model designed for general visual recognition with high-resolution features. High-resolution features (or tokens) are a natural fit for tasks that involve perceiving fine-grained details such as detection and segmentation, but exchanging global information between these features is expensive in memory and computation because of the way self-attention scales. We provide a highly efficient alternative Group Propagation Block (GP Block) to exchange global information. In each GP Block, features are first grouped together by a fixed number of learnable group tokens; we then perform Group Propagation where global information is exchanged between the grouped features; finally, global information in the updated grouped features is returned back to the image features through a transformer decoder. We evaluate GPViT on a variety of visual recognition tasks including image classification, semantic segmentation, object detection, and instance segmentation. Our method achieves significant performance gains over previous works across all tasks, especially on tasks that require highresolution outputs, for example, our GPViT-L3 outperforms Swin Transformer-B by 2.0 mIoU on ADE20K semantic segmentation with only half as many parameters. Project page: chenhongyiyang.com/projects/GPViT/GPViT

Yuying Ge, Annabella Macaluso, Li Erran Li et al.

Recent progress on vision-language foundation models have brought significant advancement to building general-purpose robots. By using the pre-trained models to encode the scene and instructions as inputs for decision making, the instruction-conditioned policy can generalize across different objects and tasks. While this is encouraging, the policy still fails in most cases given an unseen task or environment. In this work, we propose Policy Adaptation from Foundation model Feedback (PAFF). When deploying the trained policy to a new task or a new environment, we first let the policy play with randomly generated instructions to record the demonstrations. While the execution could be wrong, we can use the pre-trained foundation models to provide feedback to relabel the demonstrations. This automatically provides new pairs of demonstration-instruction data for policy fine-tuning. We evaluate our method on a broad range of experiments with the focus on generalization on unseen objects, unseen tasks, unseen environments, and sim-to-real transfer. We show PAFF improves baselines by a large margin in all cases. Our project page is available at https://geyuying.github.io/PAFF/

Lirui Wang, Yiyang Ling, Zhecheng Yuan et al.

Collecting large amounts of real-world interaction data to train general robotic policies is often prohibitively expensive, thus motivating the use of simulation data. However, existing methods for data generation have generally focused on scene-level diversity (e.g., object instances and poses) rather than task-level diversity, due to the human effort required to come up with and verify novel tasks. This has made it challenging for policies trained on simulation data to demonstrate significant task-level generalization. In this paper, we propose to automatically generate rich simulation environments and expert demonstrations by exploiting a large language models' (LLM) grounding and coding ability. Our approach, dubbed GenSim, has two modes: goal-directed generation, wherein a target task is given to the LLM and the LLM proposes a task curriculum to solve the target task, and exploratory generation, wherein the LLM bootstraps from previous tasks and iteratively proposes novel tasks that would be helpful in solving more complex tasks. We use GPT4 to expand the existing benchmark by ten times to over 100 tasks, on which we conduct supervised finetuning and evaluate several LLMs including finetuned GPTs and Code Llama on code generation for robotic simulation tasks. Furthermore, we observe that LLMs-generated simulation programs can enhance task-level generalization significantly when used for multitask policy training. We further find that with minimal sim-to-real adaptation, the multitask policies pretrained on GPT4-generated simulation tasks exhibit stronger transfer to unseen long-horizon tasks in the real world and outperform baselines by 25%. See the project website (https://liruiw.github.io/gensim) for code, demos, and videos.

Zhao-Heng Yin, Binghao Huang, Yuzhe Qin et al.

Tactile information plays a critical role in human dexterity. It reveals useful contact information that may not be inferred directly from vision. In fact, humans can even perform in-hand dexterous manipulation without using vision. Can we enable the same ability for the multi-finger robot hand? In this paper, we present Touch Dexterity, a new system that can perform in-hand object rotation using only touching without seeing the object. Instead of relying on precise tactile sensing in a small region, we introduce a new system design using dense binary force sensors (touch or no touch) overlaying one side of the whole robot hand (palm, finger links, fingertips). Such a design is low-cost, giving a larger coverage of the object, and minimizing the Sim2Real gap at the same time. We train an in-hand rotation policy using Reinforcement Learning on diverse objects in simulation. Relying on touch-only sensing, we can directly deploy the policy in a real robot hand and rotate novel objects that are not presented in training. Extensive ablations are performed on how tactile information help in-hand manipulation.Our project is available at https://touchdexterity.github.io.

Runyu Ding, Yuzhe Qin, Jiyue Zhu et al.

Teleoperation is a crucial tool for collecting human demonstrations, but controlling robots with bimanual dexterous hands remains a challenge. Existing teleoperation systems struggle to handle the complexity of coordinating two hands for intricate manipulations. We introduce Bunny-VisionPro, a real-time bimanual dexterous teleoperation system that leverages a VR headset. Unlike previous vision-based teleoperation systems, we design novel low-cost devices to provide haptic feedback to the operator, enhancing immersion. Our system prioritizes safety by incorporating collision and singularity avoidance while maintaining real-time performance through innovative designs. Bunny-VisionPro outperforms prior systems on a standard task suite, achieving higher success rates and reduced task completion times. Moreover, the high-quality teleoperation demonstrations improve downstream imitation learning performance, leading to better generalizability. Notably, Bunny-VisionPro enables imitation learning with challenging multi-stage, long-horizon dexterous manipulation tasks, which have rarely been addressed in previous work. Our system's ability to handle bimanual manipulations while prioritizing safety and real-time performance makes it a powerful tool for advancing dexterous manipulation and imitation learning.

Yuzhuang Xu, Shuo Wang, Peng Li et al. · tsinghua

Although neural machine translation (NMT) models perform well in the general domain, it remains rather challenging to control their generation behavior to satisfy the requirement of different users. Given the expensive training cost and the data scarcity challenge of learning a new model from scratch for each user requirement, we propose a memory-augmented adapter to steer pretrained NMT models in a pluggable manner. Specifically, we construct a multi-granular memory based on the user-provided text samples and propose a new adapter architecture to combine the model representations and the retrieved results. We also propose a training strategy using memory dropout to reduce spurious dependencies between the NMT model and the memory. We validate our approach on both style- and domain-specific experiments and the results indicate that our method can outperform several representative pluggable baselines.

Jianglong Ye, Jiashun Wang, Binghao Huang et al.

We propose to learn to generate grasping motion for manipulation with a dexterous hand using implicit functions. With continuous time inputs, the model can generate a continuous and smooth grasping plan. We name the proposed model Continuous Grasping Function (CGF). CGF is learned via generative modeling with a Conditional Variational Autoencoder using 3D human demonstrations. We will first convert the large-scale human-object interaction trajectories to robot demonstrations via motion retargeting, and then use these demonstrations to train CGF. During inference, we perform sampling with CGF to generate different grasping plans in the simulator and select the successful ones to transfer to the real robot. By training on diverse human data, our CGF allows generalization to manipulate multiple objects. Compared to previous planning algorithms, CGF is more efficient and achieves significant improvement on success rate when transferred to grasping with the real Allegro Hand. Our project page is available at https://jianglongye.com/cgf .

Binghao Huang, Yuanpei Chen, Tianyu Wang et al.

Humans throw and catch objects all the time. However, such a seemingly common skill introduces a lot of challenges for robots to achieve: The robots need to operate such dynamic actions at high-speed, collaborate precisely, and interact with diverse objects. In this paper, we design a system with two multi-finger hands attached to robot arms to solve this problem. We train our system using Multi-Agent Reinforcement Learning in simulation and perform Sim2Real transfer to deploy on the real robots. To overcome the Sim2Real gap, we provide multiple novel algorithm designs including learning a trajectory prediction model for the object. Such a model can help the robot catcher has a real-time estimation of where the object will be heading, and then react accordingly. We conduct our experiments with multiple objects in the real-world system, and show significant improvements over multiple baselines. Our project page is available at \url{https://binghao-huang.github.io/dynamic_handover/}.

Shiqi Yang, Minghuan Liu, Yuzhe Qin et al.

Learning from demonstrations has shown to be an effective approach to robotic manipulation, especially with the recently collected large-scale robot data with teleoperation systems. Building an efficient teleoperation system across diverse robot platforms has become more crucial than ever. However, there is a notable lack of cost-effective and user-friendly teleoperation systems for different end-effectors, e.g., anthropomorphic robot hands and grippers, that can operate across multiple platforms. To address this issue, we develop ACE, a cross-platform visual-exoskeleton system for low-cost dexterous teleoperation. Our system utilizes a hand-facing camera to capture 3D hand poses and an exoskeleton mounted on a portable base, enabling accurate real-time capture of both finger and wrist poses. Compared to previous systems, which often require hardware customization according to different robots, our single system can generalize to humanoid hands, arm-hands, arm-gripper, and quadruped-gripper systems with high-precision teleoperation. This enables imitation learning for complex manipulation tasks on diverse platforms.

Xinyue Lou, Jinan Xu, Jingyi Yin et al.

As Multimodal Large Language Models (MLLMs) become an indispensable assistant in human life, the unsafe content generated by MLLMs poses a danger to human behavior, perpetually overhanging human society like a sword of Damocles. To investigate and evaluate the safety impact of MLLMs responses on human behavior in daily life, we introduce SaLAD, a multimodal safety benchmark which contains 2,013 real-world image-text samples across 10 common categories, with a balanced design covering both unsafe scenarios and cases of oversensitivity. It emphasizes realistic risk exposure, authentic visual inputs, and fine-grained cross-modal reasoning, ensuring that safety risks cannot be inferred from text alone. We further propose a safety-warning-based evaluation framework that encourages models to provide clear and informative safety warnings, rather than generic refusals. Results on 18 MLLMs demonstrate that the top-performing models achieve a safe response rate of only 57.2% on unsafe queries. Moreover, even popular safety alignment methods limit effectiveness of the models in our scenario, revealing the vulnerabilities of current MLLMs in identifying dangerous behaviors in daily life. Our dataset is available at https://github.com/xinyuelou/SaLAD.

Sicen Liu, Xiaolong Wang, JIngcheng Du et al.

Effectively medication recommendation with complex multimorbidity conditions is a critical task in healthcare. Most existing works predicted medications based on longitudinal records, which assumed the information transmitted patterns of learning longitudinal sequence data are stable and intra-visit medical events are serialized. However, the following conditions may have been ignored: 1) A more compact encoder for intra-relationship in the intra-visit medical event is urgent; 2) Strategies for learning accurate representations of the variable longitudinal sequences of patients are different. In this paper, we proposed a novel Sample-adaptive Hierarchical medicAtion Prediction nEtwork, termed SHAPE, to tackle the above challenges in the medication recommendation task. Specifically, we design a compact intra-visit set encoder to encode the relationship in the medical event for obtaining visit-level representation and then develop an inter-visit longitudinal encoder to learn the patient-level longitudinal representation efficiently. To endow the model with the capability of modeling the variable visit length, we introduce a soft curriculum learning method to assign the difficulty of each sample automatically by the visit length. Extensive experiments on a benchmark dataset verify the superiority of our model compared with several state-of-the-art baselines.

Sicen Liu, Xiaolong Wang, Yang Xiang et al.

Medical event prediction (MEP) is a fundamental task in the medical domain, which needs to predict medical events, including medications, diagnosis codes, laboratory tests, procedures, outcomes, and so on, according to historical medical records. The task is challenging as medical data is a type of complex time series data with heterogeneous and temporal irregular characteristics. Many machine learning methods that consider the two characteristics have been proposed for medical event prediction. However, most of them consider the two characteristics separately and ignore the correlations among different types of medical events, especially relations between historical medical events and target medical events. In this paper, we propose a novel neural network based on attention mechanism, called cross-event attention-based time-aware network (CATNet), for medical event prediction. It is a time-aware, event-aware and task-adaptive method with the following advantages: 1) modeling heterogeneous information and temporal information in a unified way and considering temporal irregular characteristics locally and globally respectively, 2) taking full advantage of correlations among different types of events via cross-event attention. Experiments on two public datasets (MIMIC-III and eICU) show CATNet can be adaptive with different MEP tasks and outperforms other state-of-the-art methods on various MEP tasks. The source code of CATNet will be released after this manuscript is accepted.

Yu Sun, Xinhao Li, Karan Dalal et al.

We reformulate the problem of supervised learning as learning to learn with two nested loops (i.e. learning problems). The inner loop learns on each individual instance with self-supervision before final prediction. The outer loop learns the self-supervised task used by the inner loop, such that its final prediction improves. Our inner loop turns out to be equivalent to linear attention when the inner-loop learner is only a linear model, and to self-attention when it is a kernel estimator. For practical comparison with linear or self-attention layers, we replace each of them in a transformer with an inner loop, so our outer loop is equivalent to training the architecture. When each inner-loop learner is a neural network, our approach vastly outperforms transformers with linear attention on ImageNet from 224 x 224 raw pixels in both accuracy and FLOPs, while (regular) transformers cannot run.

Tianshuo Yuan, Yuxiang Lin, Jue Wang et al.

Combining Vision Large Language Models (VLLMs) with diffusion models offers a powerful method for executing image editing tasks based on human language instructions. However, language instructions alone often fall short in accurately conveying user requirements, particularly when users want to add, replace elements in specific areas of an image. Luckily, masks can effectively indicate the exact locations or elements to be edited, while they require users to precisely draw the shapes at the desired locations, which is highly user-unfriendly. To address this, we propose FlexEdit, an end-to-end image editing method that leverages both free-shape masks and language instructions for Flexible Editing. Our approach employs a VLLM in comprehending the image content, mask, and user instructions. Additionally, we introduce the Mask Enhance Adapter (MEA) that fuses the embeddings of the VLLM with the image data, ensuring a seamless integration of mask information and model output embeddings. Furthermore, we construct FSMI-Edit, a benchmark specifically tailored for free-shape mask, including 8 types of free-shape mask. Extensive experiments show that our method achieves state-of-the-art (SOTA) performance in LLM-based image editing, and our simple prompting technique stands out in its effectiveness. The code and data can be found at https://github.com/A-new-b/flex_edit.

Jie Xu, Yazhou Ren, Xiaolong Wang et al.

Multi-view clustering (MVC) aims at exploring category structures among multi-view data in self-supervised manners. Multiple views provide more information than single views and thus existing MVC methods can achieve satisfactory performance. However, their performance might seriously degenerate when the views are noisy in practical multi-view scenarios. In this paper, we formally investigate the drawback of noisy views and then propose a theoretically grounded deep MVC method (namely MVCAN) to address this issue. Specifically, we propose a novel MVC objective that enables un-shared parameters and inconsistent clustering predictions across multiple views to reduce the side effects of noisy views. Furthermore, a two-level multi-view iterative optimization is designed to generate robust learning targets for refining individual views' representation learning. Theoretical analysis reveals that MVCAN works by achieving the multi-view consistency, complementarity, and noise robustness. Finally, experiments on extensive public datasets demonstrate that MVCAN outperforms state-of-the-art methods and is robust against the existence of noisy views.

Ruihan Yang, Ge Yang, Xiaolong Wang

Legged robots have the potential to expand the reach of autonomy beyond paved roads. In this work, we consider the difficult problem of locomotion on challenging terrains using a single forward-facing depth camera. Due to the partial observability of the problem, the robot has to rely on past observations to infer the terrain currently beneath it. To solve this problem, we follow the paradigm in computer vision that explicitly models the 3D geometry of the scene and propose Neural Volumetric Memory (NVM), a geometric memory architecture that explicitly accounts for the SE(3) equivariance of the 3D world. NVM aggregates feature volumes from multiple camera views by first bringing them back to the ego-centric frame of the robot. We test the learned visual-locomotion policy on a physical robot and show that our approach, which explicitly introduces geometric priors during training, offers superior performance than more naïve methods. We also include ablation studies and show that the representations stored in the neural volumetric memory capture sufficient geometric information to reconstruct the scene. Our project page with videos is https://rchalyang.github.io/NVM .

Yang Fu, Shalini De Mello, Xueting Li et al.

High-fidelity 3D scene reconstruction has been substantially advanced by recent progress in neural fields. However, most existing methods train a separate network from scratch for each individual scene. This is not scalable, inefficient, and unable to yield good results given limited views. While learning-based multi-view stereo methods alleviate this issue to some extent, their multi-view setting makes it less flexible to scale up and to broad applications. Instead, we introduce training generalizable Neural Fields incorporating scene Priors (NFPs). The NFP network maps any single-view RGB-D image into signed distance and radiance values. A complete scene can be reconstructed by merging individual frames in the volumetric space WITHOUT a fusion module, which provides better flexibility. The scene priors can be trained on large-scale datasets, allowing for fast adaptation to the reconstruction of a new scene with fewer views. NFP not only demonstrates SOTA scene reconstruction performance and efficiency, but it also supports single-image novel-view synthesis, which is underexplored in neural fields. More qualitative results are available at: https://oasisyang.github.io/neural-prior