Yang Jin

Yang Jin, Kun Xu, Kun Xu et al. · pku

Recently, the remarkable advance of the Large Language Model (LLM) has inspired researchers to transfer its extraordinary reasoning capability to both vision and language data. However, the prevailing approaches primarily regard the visual input as a prompt and focus exclusively on optimizing the text generation process conditioned upon vision content by a frozen LLM. Such an inequitable treatment of vision and language heavily constrains the model's potential. In this paper, we break through this limitation by representing both vision and language in a unified form. Specifically, we introduce a well-designed visual tokenizer to translate the non-linguistic image into a sequence of discrete tokens like a foreign language that LLM can read. The resulting visual tokens encompass high-level semantics worthy of a word and also support dynamic sequence length varying from the image. Coped with this tokenizer, the presented foundation model called LaVIT can handle both image and text indiscriminately under the same generative learning paradigm. This unification empowers LaVIT to serve as an impressive generalist interface to understand and generate multi-modal content simultaneously. Extensive experiments further showcase that it outperforms the existing models by a large margin on massive vision-language tasks. Our code and models are available at https://github.com/jy0205/LaVIT.

Yang Jin, Yongzhi Li, Zehuan Yuan et al.

Spatio-Temporal video grounding (STVG) focuses on retrieving the spatio-temporal tube of a specific object depicted by a free-form textual expression. Existing approaches mainly treat this complicated task as a parallel frame-grounding problem and thus suffer from two types of inconsistency drawbacks: feature alignment inconsistency and prediction inconsistency. In this paper, we present an end-to-end one-stage framework, termed Spatio-Temporal Consistency-Aware Transformer (STCAT), to alleviate these issues. Specially, we introduce a novel multi-modal template as the global objective to address this task, which explicitly constricts the grounding region and associates the predictions among all video frames. Moreover, to generate the above template under sufficient video-textual perception, an encoder-decoder architecture is proposed for effective global context modeling. Thanks to these critical designs, STCAT enjoys more consistent cross-modal feature alignment and tube prediction without reliance on any pre-trained object detectors. Extensive experiments show that our method outperforms previous state-of-the-arts with clear margins on two challenging video benchmarks (VidSTG and HC-STVG), illustrating the superiority of the proposed framework to better understanding the association between vision and natural language. Code is publicly available at https://github.com/jy0205/STCAT.

Yang Jin, Yongzhi Li, Zehuan Yuan et al.

This paper aims to establish a generic multi-modal foundation model that has the scalable capability to massive downstream applications in E-commerce. Recently, large-scale vision-language pretraining approaches have achieved remarkable advances in the general domain. However, due to the significant differences between natural and product images, directly applying these frameworks for modeling image-level representations to E-commerce will be inevitably sub-optimal. To this end, we propose an instance-centric multi-modal pretraining paradigm called ECLIP in this work. In detail, we craft a decoder architecture that introduces a set of learnable instance queries to explicitly aggregate instance-level semantics. Moreover, to enable the model to focus on the desired product instance without reliance on expensive manual annotations, two specially configured pretext tasks are further proposed. Pretrained on the 100 million E-commerce-related data, ECLIP successfully extracts more generic, semantic-rich, and robust representations. Extensive experimental results show that, without further fine-tuning, ECLIP surpasses existing methods by a large margin on a broad range of downstream tasks, demonstrating the strong transferability to real-world E-commerce applications.

Yang Jin, Zhicheng Sun, Kun Xu et al. · pku

In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.

Quzhe Huang, Zhenwei An, Nan Zhuang et al. · pku

In this paper, we introduce a novel dynamic expert selection framework for Mixture of Experts (MoE) models, aiming to enhance computational efficiency and model performance by adjusting the number of activated experts based on input difficulty. Unlike traditional MoE approaches that rely on fixed Top-K routing, which activates a predetermined number of experts regardless of the input's complexity, our method dynamically selects experts based on the confidence level in expert selection for each input. This allows for a more efficient utilization of computational resources, activating more experts for complex tasks requiring advanced reasoning and fewer for simpler tasks. Through extensive evaluations, our dynamic routing method demonstrates substantial improvements over conventional Top-2 routing across various benchmarks, achieving an average improvement of 0.7% with less than 90% activated parameters. Further analysis shows our model dispatches more experts to tasks requiring complex reasoning skills, like BBH, confirming its ability to dynamically allocate computational resources in alignment with the input's complexity. Our findings also highlight a variation in the number of experts needed across different layers of the transformer model, offering insights into the potential for designing heterogeneous MoE frameworks. The code and models are available at https://github.com/ZhenweiAn/Dynamic_MoE.

Zhicheng Sun, Zhenhao Yang, Yang Jin et al.

Customizing diffusion models to generate identity-preserving images from user-provided reference images is an intriguing new problem. The prevalent approaches typically require training on extensive domain-specific images to achieve identity preservation, which lacks flexibility across different use cases. To address this issue, we exploit classifier guidance, a training-free technique that steers diffusion models using an existing classifier, for personalized image generation. Our study shows that based on a recent rectified flow framework, the major limitation of vanilla classifier guidance in requiring a special classifier can be resolved with a simple fixed-point solution, allowing flexible personalization with off-the-shelf image discriminators. Moreover, its solving procedure proves to be stable when anchored to a reference flow trajectory, with a convergence guarantee. The derived method is implemented on rectified flow with different off-the-shelf image discriminators, delivering advantageous personalization results for human faces, live subjects, and certain objects. Code is available at https://github.com/feifeiobama/RectifID.

Yang Jin, Lei Zhang, Shi Yan et al.

Gaze object prediction (GOP) aims to predict the category and location of the object that a human is looking at. Previous methods utilized box-level supervision to identify the object that a person is looking at, but struggled with semantic ambiguity, ie, a single box may contain several items since objects are close together. The Vision foundation model (VFM) has improved in object segmentation using box prompts, which can reduce confusion by more precisely locating objects, offering advantages for fine-grained prediction of gaze objects. This paper presents a more challenging gaze object segmentation (GOS) task, which involves inferring the pixel-level mask corresponding to the object captured by human gaze behavior. In particular, we propose that the pixel-level supervision provided by VFM can be integrated into gaze object prediction to mitigate semantic ambiguity. This leads to our gaze object detection and segmentation framework that enables accurate pixel-level predictions. Different from previous methods that require additional head input or ignore head features, we propose to automatically obtain head features from scene features to ensure the model's inference efficiency and flexibility in the real world. Moreover, rather than directly fuse features to predict gaze heatmap as in existing methods, which may overlook spatial location and subtle details of the object, we develop a space-to-object gaze regression method to facilitate human-object gaze interaction. Specifically, it first constructs an initial human-object spatial connection, then refines this connection by interacting with semantically clear features in the segmentation branch, ultimately predicting a gaze heatmap for precise localization. Extensive experiments on GOO-Synth and GOO-Real datasets demonstrate the effectiveness of our method.

Binglu Wang, Chenxi Guo, Yang Jin et al.

Gaze object prediction aims to predict the location and category of the object that is watched by a human. Previous gaze object prediction works use CNN-based object detectors to predict the object's location. However, we find that Transformer-based object detectors can predict more accurate object location for dense objects in retail scenarios. Moreover, the long-distance modeling capability of the Transformer can help to build relationships between the human head and the gaze object, which is important for the GOP task. To this end, this paper introduces Transformer into the fields of gaze object prediction and proposes an end-to-end Transformer-based gaze object prediction method named TransGOP. Specifically, TransGOP uses an off-the-shelf Transformer-based object detector to detect the location of objects and designs a Transformer-based gaze autoencoder in the gaze regressor to establish long-distance gaze relationships. Moreover, to improve gaze heatmap regression, we propose an object-to-gaze cross-attention mechanism to let the queries of the gaze autoencoder learn the global-memory position knowledge from the object detector. Finally, to make the whole framework end-to-end trained, we propose a Gaze Box loss to jointly optimize the object detector and gaze regressor by enhancing the gaze heatmap energy in the box of the gaze object. Extensive experiments on the GOO-Synth and GOO-Real datasets demonstrate that our TransGOP achieves state-of-the-art performance on all tracks, i.e., object detection, gaze estimation, and gaze object prediction. Our code will be available at https://github.com/chenxi-Guo/TransGOP.git.

Yang Jin, Jun Lv, Wenye Yu et al.

Self-improvement requires robotic systems to initially learn from human-provided data and then gradually enhance their capabilities through interaction with the environment. This is similar to how humans improve their skills through continuous practice. However, achieving effective self-improvement is challenging, primarily because robots tend to repeat their existing abilities during interactions, often failing to generate new, valuable data for learning. In this paper, we identify the key to successful self-improvement: modal-level exploration and data selection. By incorporating a modal-level exploration mechanism during policy execution, the robot can produce more diverse and multi-modal interactions. At the same time, we select the most valuable trials and high-quality segments from these interactions for learning. We successfully demonstrate effective robot self-improvement on both simulation benchmarks and real-world experiments. The capability for self-improvement will enable us to develop more robust and high-success-rate robotic control strategies at a lower cost. Our code and experiment scripts are available at https://ericjin2002.github.io/SIME/

Jinghan Li, Yang Jin, Hao Jiang et al.

Recent advances in pretraining general foundation models have significantly improved performance across diverse downstream tasks. While autoregressive (AR) generative models like GPT have revolutionized NLP, most visual generative pretraining methods still rely on BERT-style masked modeling, which often disregards the temporal information essential for video analysis. The few existing autoregressive visual pretraining methods suffer from issues such as inaccurate semantic localization and poor generation quality, leading to poor semantics. In this work, we propose NExT-Vid, a novel autoregressive visual generative pretraining framework that utilizes masked next-frame prediction to jointly model images and videos. NExT-Vid introduces a context-isolated autoregressive predictor to decouple semantic representation from target decoding, and a conditioned flow-matching decoder to enhance generation quality and diversity. Through context-isolated flow-matching pretraining, our approach achieves strong representations. Extensive experiments on large-scale pretrained models demonstrate that our proposed method consistently outperforms previous generative pretraining methods for visual representation learning via attentive probing in downstream classification.

Wendi Chen, Han Xue, Yi Wang et al.

Human-level contact-rich manipulation relies on the distinct roles of two key modalities: vision provides spatially rich but temporally slow global context, while force sensing captures rapid, high-frequency local contact dynamics. Integrating these signals is challenging due to their fundamental frequency and informational disparities. In this work, we propose ImplicitRDP, a unified end-to-end visual-force diffusion policy that integrates visual planning and reactive force control within a single network. We introduce Structural Slow-Fast Learning, a mechanism utilizing causal attention to simultaneously process asynchronous visual and force tokens, allowing the policy to perform closed-loop adjustments at the force frequency while maintaining the temporal coherence of action chunks. Furthermore, to mitigate modality collapse where end-to-end models fail to adjust the weights across different modalities, we propose Virtual-target-based Representation Regularization. This auxiliary objective maps force feedback into the same space as the action, providing a stronger, physics-grounded learning signal than raw force prediction. Extensive experiments on contact-rich tasks demonstrate that ImplicitRDP significantly outperforms both vision-only and hierarchical baselines, achieving superior reactivity and success rates with a streamlined training pipeline. Code and videos will be publicly available at https://implicit-rdp.github.io.

Yang Jin, Jun Lv, Shuqiang Jiang et al.

Generating robot demonstrations through simulation is widely recognized as an effective way to scale up robot data. Previous work often trained reinforcement learning agents to generate expert policies, but this approach lacks sample efficiency. Recently, a line of work has attempted to generate robot demonstrations via differentiable simulation, which is promising but heavily relies on reward design, a labor-intensive process. In this paper, we propose DiffGen, a novel framework that integrates differentiable physics simulation, differentiable rendering, and a vision-language model to enable automatic and efficient generation of robot demonstrations. Given a simulated robot manipulation scenario and a natural language instruction, DiffGen can generate realistic robot demonstrations by minimizing the distance between the embedding of the language instruction and the embedding of the simulated observation after manipulation. The embeddings are obtained from the vision-language model, and the optimization is achieved by calculating and descending gradients through the differentiable simulation, differentiable rendering, and vision-language model components, thereby accomplishing the specified task. Experiments demonstrate that with DiffGen, we could efficiently and effectively generate robot data with minimal human effort or training time.

Yang Jin, Guangyu Guo, Binglu Wang

Gaze object detection and gaze following are fundamental tasks for interpreting human gaze behavior or intent. However, most previous methods usually solve these two tasks separately, and their prediction of gaze objects and gaze following typically depend on head-related prior knowledge during both the training phase and real-world deployment. This dependency necessitates an auxiliary network to extract head location, thus precluding joint optimization across the entire system and constraining the practical applicability. To this end, we propose GaTector+, a unified framework for gaze object detection and gaze following, which eliminates the dependence on the head-related priors during inference. Specifically, GaTector+ uses an expanded specific-general-specific feature extractor that leverages a shared backbone, which extracts general features for gaze following and object detection using the shared backbone while using specific blocks before and after the shared backbone to better consider the specificity of each sub-task. To obtain head-related knowledge without prior information, we first embed a head detection branch to predict the head of each person. Then, before regressing the gaze point, a head-based attention mechanism is proposed to fuse the sense feature and gaze feature with the help of head location. Since the suboptimization of the gaze point heatmap leads to the performance bottleneck, we propose an attention supervision mechanism to accelerate the learning of the gaze heatmap. Finally, we propose a novel evaluation metric, mean Similarity over Candidates (mSoC), for gaze object detection, which is more sensitive to variations between bounding boxes. The experimental results on multiple benchmark datasets demonstrate the effectiveness of our model in both gaze object detection and gaze following tasks.

Yang Jin, Jun Lv, Han Xue et al.

Intelligent agents progress by continually refining their capabilities through actively exploring environments. Yet robot policies often lack sufficient exploration capability due to action mode collapse. Existing methods that encourage exploration typically rely on random perturbations, which are unsafe and induce unstable, erratic behaviors, thereby limiting their effectiveness. We propose Self-Improvement via On-Manifold Exploration (SOE), a framework that enhances policy exploration and improvement in robotic manipulation. SOE learns a compact latent representation of task-relevant factors and constrains exploration to the manifold of valid actions, ensuring safety, diversity, and effectiveness. It can be seamlessly integrated with arbitrary policy models as a plug-in module, augmenting exploration without degrading the base policy performance. Moreover, the structured latent space enables human-guided exploration, further improving efficiency and controllability. Extensive experiments in both simulation and real-world tasks demonstrate that SOE consistently outperforms prior methods, achieving higher task success rates, smoother and safer exploration, and superior sample efficiency. These results establish on-manifold exploration as a principled approach to sample-efficient policy self-improvement. Project website: https://ericjin2002.github.io/SOE

Peng Gang, Wang Zhihao, Yang Jin et al.

The traditional kinematic calibration method for manipulators requires precise three-dimensional measuring instruments to measure the end pose, which is not only expensive due to the high cost of the measuring instruments but also not applicable to all manipulators. Another calibration method uses a camera, but the system error caused by the camera's parameters affects the calibration accuracy of the kinematics of the robot arm. Therefore, this paper proposes a method for calibrating the geometric parameters of a kinematic model of a manipulator based on monocular vision. Firstly, the classic Denavit-Hartenberg(D-H) modeling method is used to establish the kinematic parameters of the manipulator. Secondly, nonlinear optimization and parameter compensation are performed. The three-dimensional positions of the feature points of the calibration plate under each manipulator attitude corresponding to the actual kinematic model and the classic D-H kinematic model are mapped into the pixel coordinate system, and the sum of Euclidean distance errors of the pixel coordinates of the two is used as the objective function to be optimized. The experimental results show that the pixel deviation of the end pose corresponding to the optimized D-H kinematic model proposed in this paper and the end pose corresponding to the actual kinematic model in the pixel coordinate system is 0.99 pixels. Compared with the 7.9 deviation pixels between the pixel coordinates calculated by the classic D-H kinematic model and the actual pixel coordinates, the deviation is reduced by nearly 7 pixels for an 87% reduction in error. Therefore, the proposed method can effectively avoid system errors caused by camera parameters in visual calibration, can improve the absolute positioning accuracy of the end of the robotic arm, and has good economy and universality.

Edgar Xi, Selina Bing, Yang Jin

In recent years, convolutional neural networks (CNN) have played an important role in the field of deep learning. Variants of CNN's have proven to be very successful in classification tasks across different domains. However, there are two big drawbacks to CNN's: their failure to take into account of important spatial hierarchies between features, and their lack of rotational invariance. As long as certain key features of an object are present in the test data, CNN's classify the test data as the object, disregarding features' relative spatial orientation to each other. This causes false positives. The lack of rotational invariance in CNN's would cause the network to incorrectly assign the object another label, causing false negatives. To address this concern, Hinton et al. propose a novel type of neural network using the concept of capsules in a recent paper. With the use of dynamic routing and reconstruction regularization, the capsule network model would be both rotation invariant and spatially aware. The capsule network has shown its potential by achieving a state-of-the-art result of 0.25% test error on MNIST without data augmentation such as rotation and scaling, better than the previous baseline of 0.39%. To further test out the application of capsule networks on data with higher dimensionality, we attempt to find the best set of configurations that yield the optimal test error on CIFAR10 dataset.