arXiv subjects

Yongyi Su

Publications and source records attributed to Yongyi Su.

At least 19 recordsLinked to original sources

arXiv2026

PointPilot: Unified Robotic Manipulation via Object-Centric 3D Point Flow Prediction

Generalization remains a central challenge in robotics, calling for policies that address domain variations from multiple dimensions such as diverse embodiments, camera poses, and/or scene layouts. In order to approach this problem, we introduce PointPilot, a task-focused and 3D aligned manipulation framework built on a 3D space point-flow prediction. Rather than modeling the global scene, PointPilot only represents task-focused objects (robot grippers treated as objects), avoiding the prediction of potentially redundant scene parts unrelated to the manipulation tasks. Moreover, PointPilot aligns 3D spatial representation in one unified coordinate and supports closed-loop and fully automatic action execution across different manipulation modes from object point flow without a separately trained action expert, while existing works typically need an extra action head or are (semi-)automatic and single-arm mode specific. Given sensor observations and language instructions, our model jointly learns subtask planning, task-relevant object grounding, and 3D point flow prediction in an end-to-end way. By processing heterogeneous data in a shared 3D point flow space, PointPilot is able to absorb demonstrations from different embodiments. Extensive simulative and real world experiments validate our method. Notably, we demonstrate zero-shot embodiment generalization with PointPilot. A single trained policy can be deployed on previously unseen embodiments and successfully complete manipulation tasks without any additional fine-tuning.

cs.RO↗

arXiv2026

TaskAnchor: Grounding Task State in Reactive VLAs for Long-Horizon Manipulation

Reactive vision-language-action (VLA) policies suffer from task-state aliasing in long-horizon manipulation, where identical multimodal inputs call for distinct, context-dependent actions. Given that pretrained VLAs already possess rich control primitives to express diverse behaviors, we hypothesize that the execution bottleneck lies not in policy capacity, but in input ambiguity. In this paper, we propose TaskAnchor, a lightweight adapter that grounds task state by injecting execution context into the VLA's native input space. During post-training, TaskAnchor learns to represent the semantic execution stage as a milestone-supervised coordinate prepended to the language instruction, while incorporating fine-grained historical evidence via a residual update to the current visual tokens. This formulation avoids generating complex subtask instructions and leaves the backbone architecture unchanged. Across long-horizon benchmarks, TaskAnchor delivers substantial gains, achieving approximately 6 times the average success rate of the pi0.5 and X-VLA baselines on RMBench and more than doubling the task success rate of pi0.5 on RoboMemArena. Real-robot experiments further validate reliable multi-stage execution, with the same policy adapting its subsequent behaviors using earlier human interactions as in-context cues. Our project website is available at https://taskanchor.netlify.app/.

cs.RO↗

arXiv2026

UNITAS: A 3D-Native World Action Model for Embodied Manipulation

World action models (WAMs) aim to answer a coupled physical question: given a task instruction, what motion should the robot execute, and how will that motion change the surrounding world? Most existing WAMs build on pretrained video generators and represent world evolution through images or visual latents. Robotic interaction, however, takes place in metric three-dimensional space, while images are view-dependent projections whose pixel distances do not directly encode physical distances. We introduce UNITAS, to our knowledge the first 3D-native world action model that unifies observations, actions, and scene dynamics in a shared metric 3D frame within each interaction, using a common representation across robot embodiments and human hands. Action flow represents human hands and robot grippers as 3D point trajectories, while scene flow describes scene-point displacements conditioned on these trajectories. World-aligned 3D positional embeddings ground visual tokens with or without depth input, and a physical-time trajectory tokenizer encodes each point trajectory as one token anchored at its current 3D position. This interface supports both direct action execution and action-conditioned scene prediction. With 1.7B parameters, UNITAS achieves the best action-conditioned scene prediction on RoboTwin among the compared methods, with up to 49% lower displacement errors than PointWorld, and state-of-the-art manipulation success, including 99.8% on LIBERO and an average of 85% across real-world tasks. The code is available at https://github.com/DexForce/UNITAS.

cs.RO↗

arXiv2026

Aligning Thoughts with Answers: Probability Rewards to Tame Thinking Drift

This paper studies \textbf{thinking--answer consistency} in vision-language models. We focus on Visual Intention Grounding, where a model infers a target object based on a human intention query and predicts a bounding box. We reveal that previous IoU-based reinforcement learning (RL) frameworks suffer from ``thinking drift'', where the model produces a correct bounding box, despite having an incorrect reasoning process pointing to a different target object. Thus, we propose \textbf{Rita} (\textit{ReInforcing Thinking--Answer consistency}) as a novel RL paradigm to tame the drift. Specifically, Rita introduces two reasoning-label-free RL rewards, constructed from the conditional probability of reference answers: a \textbf{thinking reward} and a \textbf{consistency reward}. It also adopts a difficulty-aware \textbf{data filtering} strategy that selects informative easy-to-medium samples for RL using rollout error rate and reward variance. Extensive experiments on EgoIntention and the new RefEgo-Int benchmarks show that Rita performs consistently superior to the supervised finetuning approaches and vanilla RL-finetuned frameworks.

cs.CV↗

arXiv2026

AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization

While Multimodal Large Language Models (MLLMs) demonstrate remarkable capabilities across diverse domains, their application to specialized anomaly detection (AD) remains constrained by domain adaptation challenges. Existing Group Relative Policy Optimization (GRPO) based approaches suffer from two critical limitations: inadequate training data utilization when models produce uniform responses, and insufficient supervision over reasoning processes that encourage immediate binary decisions without deliberative analysis. We propose a comprehensive framework addressing these limitations through two synergistic innovations. First, we introduce a multi-stage deliberative reasoning process that guides models from region identification to focused examination, generating diverse response patterns essential for GRPO optimization while enabling structured supervision over analytical workflows. Second, we develop a fine-grained reward mechanism incorporating classification accuracy and localization supervision, transforming binary feedback into continuous signals that distinguish genuine analytical insight from spurious correctness. Comprehensive evaluation across multiple industrial datasets demonstrates substantial performance improvements in adapting general vision-language models to specialized anomaly detection. Our method achieves superior accuracy with efficient adaptation of existing annotations, effectively bridging the gap between general-purpose MLLM capabilities and the fine-grained visual discrimination required for detecting subtle manufacturing defects and structural irregularities.

cs.CV↗

arXiv2026

Dexterity-BEV: Aligning 3D World and Actions for Generalizable Robot Policies Learning

End-to-end manipulation policies, combined with web-scale pretrained Vision-Language Models (VLMs), show the promise for generalizable and dexterous robotic manipulation. However, they inherit two key limitations from 2D foundation models: 1) the reliance on 2D RGB inputs that ignores the intrinsically 3D nature of manipulation; and 2) the lack of spatial 3D alignment between input-output spaces as well as across diverse robot embodiments, camera setups, and trajectory datasets. In this paper, we present a series of contributions to address these issues. First, we introduce aligned vertex map and vertex spectrum -- a pixel-wise 3D representation that elevates 2D visual inputs to 3D, using camera calibration and optional depth. This novel input representation marries 3D awareness with the generalization of 2D large VLMs. Then, we propose to align the inputs and outputs of manipulation policies by expressing per-pixel 3D information of each camera view and robot actions to a shared coordinate. Based on this, we designate a canonical Bird's-Eye-View (BEV) alignment frame and innovatively propose to construct BEV images, producing a view-invariant representation robust to camera pose variations. To enable training and evaluation at scale, we develop a comprehensive data processing pipeline to perform such alignments; we also introduce a novel temporal alignment scheme for trajectories across diverse robots, human operators, and datasets. These contributions collectively mitigate input and output spatial-temporal misalignments, improving the consistency and generalization for real-world manipulation. Pretrained checkpoint, source code and data processing pipeline are available in https://hnuzhy.github.io/projects/Dex-BEV.

cs.RO↗

arXiv2025

Enhancing Generalization of Depth Estimation Foundation Model via Weakly-Supervised Adaptation with Regularization

The emergence of foundation models has substantially advanced zero-shot generalization in monocular depth estimation (MDE), as exemplified by the Depth Anything series. However, given access to some data from downstream tasks, a natural question arises: can the performance of these models be further improved? To this end, we propose WeSTAR, a parameter-efficient framework that performs Weakly supervised Self-Training Adaptation with Regularization, designed to enhance the robustness of MDE foundation models in unseen and diverse domains. We first adopt a dense self-training objective as the primary source of structural self-supervision. To further improve robustness, we introduce semantically-aware hierarchical normalization, which exploits instance-level segmentation maps to perform more stable and multi-scale structural normalization. Beyond dense supervision, we introduce a cost-efficient weak supervision in the form of pairwise ordinal depth annotations to further guide the adaptation process, which enforces informative ordinal constraints to mitigate local topological errors. Finally, a weight regularization loss is employed to anchor the LoRA updates, ensuring training stability and preserving the model's generalizable knowledge. Extensive experiments on both realistic and corrupted out-of-distribution datasets under diverse and challenging scenarios demonstrate that WeSTAR consistently improves generalization and achieves state-of-the-art performance across a wide range of benchmarks.

cs.CV↗

arXiv2025

Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs

Multimodal large language models (MLLMs) have advanced rapidly in recent years. However, existing approaches for vision tasks often rely on indirect representations, such as generating coordinates as text for detection, which limits performance and prevents dense prediction tasks like segmentation. To overcome these challenges, we introduce Patch-as-Decodable Token (PaDT), a unified paradigm that enables MLLMs to directly generate both textual and diverse visual outputs. Central to PaDT are Visual Reference Tokens (VRTs), derived from visual patch embeddings of query images and interleaved seamlessly with LLM's output textual tokens. A lightweight decoder then transforms LLM's outputs into detection, segmentation, and grounding predictions. Unlike prior methods, PaDT processes VRTs independently at each forward pass and dynamically expands the embedding table, thus improving localization and differentiation among similar objects. We further tailor a training strategy for PaDT by randomly selecting VRTs for supervised fine-tuning and introducing a robust per-token cross-entropy loss. Our empirical studies across four visual perception and understanding tasks suggest PaDT consistently achieving state-of-the-art performance, even compared with significantly larger MLLM models. The code is available at https://github.com/Gorilla-Lab-SCUT/PaDT.

cs.CV↗

arXiv2025

SODA: Out-of-Distribution Detection in Domain-Shifted Point Clouds via Neighborhood Propagation

As point cloud data increases in prevalence in a variety of applications, the ability to detect out-of-distribution (OOD) point cloud objects becomes critical for ensuring model safety and reliability. However, this problem remains under-explored in existing research. Inspired by success in the image domain, we propose to exploit advances in 3D vision-language models (3D VLMs) for OOD detection in point cloud objects. However, a major challenge is that point cloud datasets used to pre-train 3D VLMs are drastically smaller in size and object diversity than their image-based counterparts. Critically, they often contain exclusively computer-designed synthetic objects. This leads to a substantial domain shift when the model is transferred to practical tasks involving real objects scanned from the physical environment. In this paper, our empirical experiments show that synthetic-to-real domain shift significantly degrades the alignment of point cloud with their associated text embeddings in the 3D VLM latent space, hindering downstream performance. To address this, we propose a novel methodology called SODA which improves the detection of OOD point clouds through a neighborhood-based score propagation scheme. SODA is inference-based, requires no additional model training, and achieves state-of-the-art performance over existing approaches across datasets and problem settings.

cs.CV↗

arXiv2025

Towards Real-World Test-Time Adaptation: Tri-Net Self-Training with Balanced Normalization

Test-Time Adaptation aims to adapt source domain model to testing data at inference stage with success demonstrated in adapting to unseen corruptions. However, these attempts may fail under more challenging real-world scenarios. Existing works mainly consider real-world test-time adaptation under non-i.i.d. data stream and continual domain shift. In this work, we first complement the existing real-world TTA protocol with a globally class imbalanced testing set. We demonstrate that combining all settings together poses new challenges to existing methods. We argue the failure of state-of-the-art methods is first caused by indiscriminately adapting normalization layers to imbalanced testing data. To remedy this shortcoming, we propose a balanced batchnorm layer to swap out the regular batchnorm at inference stage. The new batchnorm layer is capable of adapting without biasing towards majority classes. We are further inspired by the success of self-training (ST) in learning from unlabeled data and adapt ST for test-time adaptation. However, ST alone is prone to over adaption which is responsible for the poor performance under continual domain shift. Hence, we propose to improve self-training under continual domain shift by regularizing model updates with an anchored loss. The final TTA model, termed as TRIBE, is built upon a tri-net architecture with balanced batchnorm layers. We evaluate TRIBE on four datasets representing real-world TTA settings. TRIBE consistently achieves the state-of-the-art performance across multiple evaluation protocols. The code is available at https://github.com/Gorilla-Lab-SCUT/TRIBE.

cs.LG↗

arXiv2025

Robust Distribution Alignment for Industrial Anomaly Detection under Distribution Shift

Anomaly detection plays a crucial role in quality control for industrial applications. However, ensuring robustness under unseen domain shifts such as lighting variations or sensor drift remains a significant challenge. Existing methods attempt to address domain shifts by training generalizable models but often rely on prior knowledge of target distributions and can hardly generalise to backbones designed for other data modalities. To overcome these limitations, we build upon memory-bank-based anomaly detection methods, optimizing a robust Sinkhorn distance on limited target training data to enhance generalization to unseen target domains. We evaluate the effectiveness on both 2D and 3D anomaly detection benchmarks with simulated distribution shifts. Our proposed method demonstrates superior results compared with state-of-the-art anomaly detection and domain adaptation methods.

cs.CV↗

arXiv2025

On the Adversarial Risk of Test Time Adaptation: An Investigation into Realistic Test-Time Data Poisoning

Test-time adaptation (TTA) updates the model weights during the inference stage using testing data to enhance generalization. However, this practice exposes TTA to adversarial risks. Existing studies have shown that when TTA is updated with crafted adversarial test samples, also known as test-time poisoned data, the performance on benign samples can deteriorate. Nonetheless, the perceived adversarial risk may be overstated if the poisoned data is generated under overly strong assumptions. In this work, we first review realistic assumptions for test-time data poisoning, including white-box versus grey-box attacks, access to benign data, attack order, and more. We then propose an effective and realistic attack method that better produces poisoned samples without access to benign samples, and derive an effective in-distribution attack objective. We also design two TTA-aware attack objectives. Our benchmarks of existing attack methods reveal that the TTA methods are more robust than previously believed. In addition, we analyze effective defense strategies to help develop adversarially robust TTA methods. The source code is available at https://github.com/Gorilla-Lab-SCUT/RTTDP.

cs.LG↗

arXiv2025

Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model

Vision-language models (VLMs) have revolutionized machine learning by leveraging large pre-trained models to tackle various downstream tasks. Although label, training, and data efficiency have improved, many state-of-the-art VLMs still require task-specific hyperparameter tuning and fail to fully exploit test samples. To overcome these challenges, we propose a graph-based approach for label-efficient adaptation and inference. Our method dynamically constructs a graph over text prompts, few-shot examples, and test samples, using label propagation for inference without task-specific tuning. Unlike existing zero-shot label propagation techniques, our approach requires no additional unlabeled support set and effectively leverages the test sample manifold through dynamic graph expansion. We further introduce a context-aware feature re-weighting mechanism to improve task adaptation accuracy. Additionally, our method supports efficient graph expansion, enabling real-time inductive inference. Extensive evaluations on downstream tasks, such as fine-grained categorization and out-of-distribution generalization, demonstrate the effectiveness of our approach. The source code is available at https://github.com/Yushu-Li/ECALP.

cs.CV↗

arXiv2025

PointSAM: Pointly-Supervised Segment Anything Model for Remote Sensing Images

Segment Anything Model (SAM) is an advanced foundational model for image segmentation, which is gradually being applied to remote sensing images (RSIs). Due to the domain gap between RSIs and natural images, traditional methods typically use SAM as a source pre-trained model and fine-tune it with fully supervised masks. Unlike these methods, our work focuses on fine-tuning SAM using more convenient and challenging point annotations. Leveraging SAM's zero-shot capabilities, we adopt a self-training framework that iteratively generates pseudo-labels for training. However, if the pseudo-labels contain noisy labels, there is a risk of error accumulation. To address this issue, we extract target prototypes from the target dataset and use the Hungarian algorithm to match them with prediction prototypes, preventing the model from learning in the wrong direction. Additionally, due to the complex backgrounds and dense distribution of objects in RSI, using point prompts may result in multiple objects being recognized as one. To solve this problem, we propose a negative prompt calibration method based on the non-overlapping nature of instance masks. In brief, we use the prompts of overlapping masks as corresponding negative signals, resulting in refined masks. Combining the above methods, we propose a novel Pointly-supervised Segment Anything Model named PointSAM. We conduct experiments on RSI datasets, including WHU, HRSID, and NWPU VHR-10, and the results show that our method significantly outperforms direct testing with SAM, SAM2, and other comparison methods. Furthermore, we introduce PointSAM as a point-to-box converter and achieve encouraging results, suggesting that this method can be extended to other point-supervised tasks. The code is available at https://github.com/Lans1ng/PointSAM.

cs.CV↗

arXiv2025

Augmented Contrastive Clustering with Uncertainty-Aware Prototyping for Time Series Test Time Adaptation

Test-time adaptation aims to adapt pre-trained deep neural networks using solely online unlabelled test data during inference. Although TTA has shown promise in visual applications, its potential in time series contexts remains largely unexplored. Existing TTA methods, originally designed for visual tasks, may not effectively handle the complex temporal dynamics of real-world time series data, resulting in suboptimal adaptation performance. To address this gap, we propose Augmented Contrastive Clustering with Uncertainty-aware Prototyping (ACCUP), a straightforward yet effective TTA method for time series data. Initially, our approach employs augmentation ensemble on the time series data to capture diverse temporal information and variations, incorporating uncertainty-aware prototypes to distill essential characteristics. Additionally, we introduce an entropy comparison scheme to selectively acquire more confident predictions, enhancing the reliability of pseudo labels. Furthermore, we utilize augmented contrastive clustering to enhance feature discriminability and mitigate error accumulation from noisy pseudo labels, promoting cohesive clustering within the same class while facilitating clear separation between different classes. Extensive experiments conducted on three real-world time series datasets and an additional visual dataset demonstrate the effectiveness and generalization potential of the proposed method, advancing the underexplored realm of TTA for time series data.

cs.LG↗

arXiv2024

Exploring Human-in-the-Loop Test-Time Adaptation by Synergizing Active Learning and Model Selection

Existing test-time adaptation (TTA) approaches often adapt models with the unlabeled testing data stream. A recent attempt relaxed the assumption by introducing limited human annotation, referred to as Human-In-the-Loop Test-Time Adaptation (HILTTA) in this study. The focus of existing HILTTA studies lies in selecting the most informative samples to label, a.k.a. active learning. In this work, we are motivated by a pitfall of TTA, i.e. sensitivity to hyper-parameters, and propose to approach HILTTA by synergizing active learning and model selection. Specifically, we first select samples for human annotation (active learning) and then use the labeled data to select optimal hyper-parameters (model selection). To prevent the model selection process from overfitting to local distributions, multiple regularization techniques are employed to complement the validation objective. A sample selection strategy is further tailored by considering the balance between active learning and model selection purposes. We demonstrate on 5 TTA datasets that the proposed HILTTA approach is compatible with off-the-shelf TTA methods and such combinations substantially outperform the state-of-the-art HILTTA methods. Importantly, our proposed method can always prevent choosing the worst hyper-parameters on all off-the-shelf TTA methods. The source code is available at https://github.com/Yushu-Li/HILTTA.

cs.CV↗

arXiv2024

CLIP-Guided Source-Free Object Detection in Aerial Images

Domain adaptation is crucial in aerial imagery, as the visual representation of these images can significantly vary based on factors such as geographic location, time, and weather conditions. Additionally, high-resolution aerial images often require substantial storage space and may not be readily accessible to the public. To address these challenges, we propose a novel Source-Free Object Detection (SFOD) method. Specifically, our approach begins with a self-training framework, which significantly enhances the performance of baseline methods. To alleviate the noisy labels in self-training, we utilize Contrastive Language-Image Pre-training (CLIP) to guide the generation of pseudo-labels, termed CLIP-guided Aggregation (CGA). By leveraging CLIP's zero-shot classification capability, we aggregate its scores with the original predicted bounding boxes, enabling us to obtain refined scores for the pseudo-labels. To validate the effectiveness of our method, we constructed two new datasets from different domains based on the DIOR dataset, named DIOR-C and DIOR-Cloudy. Experimental results demonstrate that our method outperforms other comparative algorithms. The code is available at https://github.com/Lans1ng/SFOD-RS.

cs.CV↗

arXiv2024

Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation

The success of large language models has inspired the computer vision community to explore image segmentation foundation model that is able to zero/few-shot generalize through prompt engineering. Segment-Anything(SAM), among others, is the state-of-the-art image segmentation foundation model demonstrating strong zero/few-shot generalization. Despite the success, recent studies reveal the weakness of SAM under strong distribution shift. In particular, SAM performs awkwardly on corrupted natural images, camouflaged images, medical images, etc. Motivated by the observations, we aim to develop a self-training based strategy to adapt SAM to target distribution. Given the unique challenges of large source dataset, high computation cost and incorrect pseudo label, we propose a weakly supervised self-training architecture with anchor regularization and low-rank finetuning to improve the robustness and computation efficiency of adaptation. We validate the effectiveness on 5 types of downstream segmentation tasks including natural clean/corrupted images, medical images, camouflaged images and robotic images. Our proposed method is task-agnostic in nature and outperforms pre-trained SAM and state-of-the-art domain adaptation methods on almost all downstream tasks with the same testing prompt inputs.

cs.CV↗