Wenbo Ding

Chengyang Ying, You Qiaoben, Xinning Zhou et al. · tsinghua

Embodied agents in vision navigation coupled with deep neural networks have attracted increasing attention. However, deep neural networks have been shown vulnerable to malicious adversarial noises, which may potentially cause catastrophic failures in Embodied Vision Navigation. Among different adversarial noises, universal adversarial perturbations (UAP), i.e., a constant image-agnostic perturbation applied on every input frame of the agent, play a critical role in Embodied Vision Navigation since they are computation-efficient and application-practical during the attack. However, existing UAP methods ignore the system dynamics of Embodied Vision Navigation and might be sub-optimal. In order to extend UAP to the sequential decision setting, we formulate the disturbed environment under the universal noise $δ$, as a $δ$-disturbed Markov Decision Process ($δ$-MDP). Based on the formulation, we analyze the properties of $δ$-MDP and propose two novel Consistent Attack methods, named Reward UAP and Trajectory UAP, for attacking Embodied agents, which consider the dynamic of the MDP and calculate universal noises by estimating the disturbed distribution and the disturbed Q function. For various victim models, our Consistent Attack can cause a significant drop in their performance in the PointGoal task in Habitat with different datasets and different scenes. Extensive experimental results indicate that there exist serious potential risks for applying Embodied Vision Navigation methods to the real world.

Yuzhu Mao, Zihao Zhao, Meilin Yang et al.

Federated learning (FL) enables edge devices to collaboratively learn a model in a distributed fashion. Many existing researches have focused on improving communication efficiency of high-dimensional models and addressing bias caused by local updates. However, most of FL algorithms are either based on reliable communications or assume fixed and known unreliability characteristics. In practice, networks could suffer from dynamic channel conditions and non-deterministic disruptions, with time-varying and unknown characteristics. To this end, in this paper we propose a sparsity enabled FL framework with both communication efficiency and bias reduction, termed as SAFARI. It makes novel use of a similarity among client models to rectify and compensate for bias that is resulted from unreliable communications. More precisely, sparse learning is implemented on local clients to mitigate communication overhead, while to cope with unreliable communications, a similarity-based compensation method is proposed to provide surrogates for missing model updates. We analyze SAFARI under bounded dissimilarity and with respect to sparse models. It is demonstrated that SAFARI under unreliable communications is guaranteed to converge at the same rate as the standard FedAvg with perfect communications. Implementations and evaluations on CIFAR-10 dataset validate the effectiveness of SAFARI by showing that it can achieve the same convergence speed and accuracy as FedAvg with perfect communications, with up to 80% of the model weights being pruned and a high percentage of client updates missing in each round.

Zihao Zhao, Yuzhu Mao, Zhenpeng Shi et al.

The widespread adoption of Federated Learning (FL), a privacy-preserving distributed learning methodology, has been impeded by the challenge of high communication overheads, typically arising from the transmission of large-scale models. Existing adaptive quantization methods, designed to mitigate these overheads, operate under the impractical assumption of uniform device participation in every training round. Additionally, these methods are limited in their adaptability due to the necessity of manual quantization level selection and often overlook biases inherent in local devices' data, thereby affecting the robustness of the global model. In response, this paper introduces AQUILA (adaptive quantization in device selection strategy), a novel adaptive framework devised to effectively handle these issues, enhancing the efficiency and robustness of FL. AQUILA integrates a sophisticated device selection method that prioritizes the quality and usefulness of device updates. Utilizing the exact global model stored by devices, it enables a more precise device selection criterion, reduces model deviation, and limits the need for hyperparameter adjustments. Furthermore, AQUILA presents an innovative quantization criterion, optimized to improve communication efficiency while assuring model convergence. Our experiments demonstrate that AQUILA significantly decreases communication costs compared to existing methods, while maintaining comparable model performance across diverse non-homogeneous FL settings, such as Non-IID data and heterogeneous model architectures.

Ruize Zhang, Zelai Xu, Chengdong Ma et al. · tsinghua

Self-play, a learning paradigm where agents iteratively refine their policies by interacting with historical or concurrent versions of themselves or other evolving agents, has shown remarkable success in solving complex non-cooperative multi-agent tasks. Despite its growing prominence in multi-agent reinforcement learning (MARL), such as Go, poker, and video games, a comprehensive and structured understanding of self-play remains lacking. This survey fills this gap by offering a comprehensive roadmap to the diverse landscape of self-play methods. We begin by introducing the necessary preliminaries, including the MARL framework and basic game theory concepts. Then, it provides a unified framework and classifies existing self-play algorithms within this framework. Moreover, the paper bridges the gap between the algorithms and their practical implications by illustrating the role of self-play in different non-cooperative scenarios. Finally, the survey highlights open challenges and future research directions in self-play.

Zijian Zhu, Yichi Zhang, Hai Chen et al.

3D object detection is an essential perception task in autonomous driving to understand the environments. The Bird's-Eye-View (BEV) representations have significantly improved the performance of 3D detectors with camera inputs on popular benchmarks. However, there still lacks a systematic understanding of the robustness of these vision-dependent BEV models, which is closely related to the safety of autonomous driving systems. In this paper, we evaluate the natural and adversarial robustness of various representative models under extensive settings, to fully understand their behaviors influenced by explicit BEV features compared with those without BEV. In addition to the classic settings, we propose a 3D consistent patch attack by applying adversarial patches in the 3D space to guarantee the spatiotemporal consistency, which is more realistic for the scenario of autonomous driving. With substantial experiments, we draw several findings: 1) BEV models tend to be more stable than previous methods under different natural conditions and common corruptions due to the expressive spatial representations; 2) BEV models are more vulnerable to adversarial noises, mainly caused by the redundant BEV features; 3) Camera-LiDAR fusion models have superior performance under different settings with multi-modal inputs, but BEV fusion model is still vulnerable to adversarial noises of both point cloud and image. These findings alert the safety issue in the applications of BEV detectors and could facilitate the development of more robust models.

Lingfeng Zhang, Xiaoshuai Hao, Yingbo Tang et al.

Navigation and manipulation are fundamental capabilities of embodied intelligence, enabling robots to interpret natural language commands and interact physically with their surroundings. However, current Vision-Language-Action (VLA) models remain constrained by task-specific architectures, specializing in either navigation or manipulation, which hinders the development of general-purpose robotic agents. To bridge this gap, we introduce OneVLA, a unified architecture that integrates these distinct tasks into a single, cohesive framework. Specifically, we design a unified action head capable of generating both navigation and manipulation actions without requiring task-specific variants. Furthermore, we propose a multi stage progressive training strategy-incorporating curated data construction and Chain-of-Thought (CoT) fine-tuning that facilitates strong positive transfer and mutual reinforcement between the two domains. Extensive experiments in both simulated and real-world environments demonstrate that OneVLA achieves state-of-the-art performance, significantly outperforming both specialized single-task and existing cross-task models. By unifying these core capabilities, OneVLA paves the way for truly general-purpose robotic systems. The model and source code will be publicly released.

Zihao Zhao, Yang Liu, Wenbo Ding et al.

Existing research has either adapted the Probably Approximately Correct (PAC) Bayesian framework for federated learning (FL) or used information-theoretic PAC-Bayesian bounds while introducing their theorems, but few considering the non-IID challenges in FL. Our work presents the first non-vacuous federated PAC-Bayesian bound tailored for non-IID local data. This bound assumes unique prior knowledge for each client and variable aggregation weights. We also introduce an objective function and an innovative Gibbs-based algorithm for the optimization of the derived bound. The results are validated on real-world datasets.

Hanhan Zhou, Tian Lan, Guru Venkataramani et al.

Cross-device Federated Learning (FL) faces significant challenges where low-end clients that could potentially make unique contributions are excluded from training large models due to their resource bottlenecks. Recent research efforts have focused on model-heterogeneous FL, by extracting reduced-size models from the global model and applying them to local clients accordingly. Despite the empirical success, general theoretical guarantees of convergence on this method remain an open question. This paper presents a unifying framework for heterogeneous FL algorithms with online model extraction and provides a general convergence analysis for the first time. In particular, we prove that under certain sufficient conditions and for both IID and non-IID data, these algorithms converge to a stationary point of standard FL for general smooth cost functions. Moreover, we introduce the concept of minimum coverage index, together with model reduction noise, which will determine the convergence of heterogeneous federated learning, and therefore we advocate for a holistic approach that considers both factors to enhance the efficiency of heterogeneous federated learning.

Shoujie Li, Haixin Yu, Wenbo Ding et al.

The accurate detection and grasping of transparent objects are challenging but of significance to robots. Here, a visual-tactile fusion framework for transparent object grasping under complex backgrounds and variant light conditions is proposed, including the grasping position detection, tactile calibration, and visual-tactile fusion based classification. First, a multi-scene synthetic grasping dataset generation method with a Gaussian distribution based data annotation is proposed. Besides, a novel grasping network named TGCNN is proposed for grasping position detection, showing good results in both synthetic and real scenes. In tactile calibration, inspired by human grasping, a fully convolutional network based tactile feature extraction method and a central location based adaptive grasping strategy are designed, improving the success rate by 36.7% compared to direct grasping. Furthermore, a visual-tactile fusion method is proposed for transparent objects classification, which improves the classification accuracy by 34%. The proposed framework synergizes the advantages of vision and touch, and greatly improves the grasping efficiency of transparent objects.

Yuzhu Mao, Siqi Ping, Zihao Zhao et al.

Large pre-trained models, such as large language models (LLMs), present significant resource challenges for fine-tuning due to their extensive parameter sizes, especially for applications in mobile systems. To address this, Low-Rank Adaptation (LoRA) has been developed to reduce resource consumption while maintaining satisfactory fine-tuning results. Despite its effectiveness, the original LoRA method faces challenges of suboptimal performance and overfitting. This paper investigates the intrinsic dimension of the matrix updates approximated by the LoRA method and reveals the performance benefits of increasing this intrinsic dimension. By employing regularization and a gradient masking method that encourages higher intrinsic dimension, the proposed method, termed Regularized and Masked LoRA (RM-LoRA), achieves superior generalization performance with the same or lower trainable parameter budget compared to the original LoRA and its latest variants across various open-source vision and language datasets.

Zhengru Fang, Yanan Ma, Yu Guo et al.

When a chest X-ray shows consolidation but the question asks which finding is present, a medical vision-language model may answer "No consolidation." This is more than an incorrect choice: it is a polarity reversal that emits a clinical statement contradicting the image. We study this failure as negated-option attraction, where a model is drawn to a negated answer option even when it conflicts with both the visual evidence and the question. We introduce CXR-ContraBench (Chest X-Ray Contradiction Benchmark), a diagnostic benchmark spanning internal ReXVQA slices and external OpenI and CheXpert protocols. The benchmark centers on present-finding questions, where selecting "No X" despite visible X creates the main clinical risk, and uses absent-finding questions as secondary tests of whether models copy negated wording. Across CheXpert protocols, the failure is substantial and persistent. On a strict direct presence probe, MedGemma and Qwen2.5-VL reach only 31.49% and 30.21% accuracy, respectively; on a matched 135,754-record CheXpert training-split protocol, both models select negated options on over 62% of presence questions. Chain-of-thought prompting reduces some presence-side reversals but does not eliminate them and can amplify absence-side contradictions. Finally, QCCV-Neg (Question-Conditioned Consistency Verifier for Negation) deterministically repairs the measured polarity-confused subset without retraining, raising MedGemma and Qwen2.5-VL to 96.60% and 95.32% accuracy on the direct presence probe. These results show that standard accuracy can hide a clinically meaningful inference-time polarity failure. Source code and benchmark construction scripts are available at https://github.com/fangzr/cxr-contrabench-code.

Tongcheng Fang, Hanling Zhang, Ruiqi Xie et al. · tsinghua

Diffusion Transformers have recently demonstrated remarkable performance in video generation. However, the long input sequences result in high computational latency due to the quadratic complexity of full attention. Various sparse attention mechanisms have been proposed. Training-free sparse attention is constrained by limited sparsity and thus offers modest acceleration, whereas training-based methods can reach much higher sparsity but demand substantial data and computation for training. In this work, we propose SALAD, introducing a lightweight linear attention branch in parallel with the sparse attention. By incorporating an input-dependent gating mechanism to finely balance the two branches, our method attains 90% sparsity and 1.72x inference speedup, while maintaining generation quality comparable to the full attention baseline. Moreover, our finetuning process is highly efficient, requiring only 2,000 video samples and 1,600 training steps with a batch size of 8.

Hengxi Zhang, Zhendong Shi, Yuanquan Hu et al.

Quantitative markets are characterized by swift dynamics and abundant uncertainties, making the pursuit of profit-driven stock trading actions inherently challenging. Within this context, reinforcement learning (RL), which operates on a reward-centric mechanism for optimal control, has surfaced as a potentially effective solution to the intricate financial decision-making conundrums presented. This paper delves into the fusion of two established financial trading strategies, namely the constant proportion portfolio insurance (CPPI) and the time-invariant portfolio protection (TIPP), with the multi-agent deep deterministic policy gradient (MADDPG) framework. As a result, we introduce two novel multi-agent RL (MARL) methods, CPPI-MADDPG and TIPP-MADDPG, tailored for probing strategic trading within quantitative markets. To validate these innovations, we implemented them on a diverse selection of 100 real-market shares. Our empirical findings reveal that the CPPI-MADDPG and TIPP-MADDPG strategies consistently outpace their traditional counterparts, affirming their efficacy in the realm of quantitative trading.

Zihao Zhao, Mengen Luo, Wenbo Ding

Distributed machine learning has been widely used in recent years to tackle the large and complex dataset problem. Therewith, the security of distributed learning has also drawn increasing attentions from both academia and industry. In this context, federated learning (FL) was developed as a "secure" distributed learning by maintaining private training data locally and only public model gradients are communicated between. However, to date, a variety of gradient leakage attacks have been proposed for this procedure and prove that it is insecure. For instance, a common drawback of these attacks is shared: they require too much auxiliary information such as model weights, optimizers, and some hyperparameters (e.g., learning rate), which are difficult to obtain in real situations. Moreover, many existing algorithms avoid transmitting model gradients in FL and turn to sending model weights, such as FedAvg, but few people consider its security breach. In this paper, we present two novel frameworks to demonstrate that transmitting model weights is also likely to leak private local data of clients, i.e., (DLM and DLM+), under the FL scenario. In addition, a number of experiments are performed to illustrate the effect and generality of our attack frameworks. At the end of this paper, we also introduce two defenses to the proposed attacks and evaluate their protection effects. Comprehensively, the proposed attack and defense schemes can be applied to the general distributed learning scenario as well, just with some appropriate customization.

Hongxiang Cai, Zeyuan Zhang, Zhenyu Zhou et al.

Integrating LiDAR and Camera information into Bird's-Eye-View (BEV) has become an essential topic for 3D object detection in autonomous driving. Existing methods mostly adopt an independent dual-branch framework to generate LiDAR and camera BEV, then perform an adaptive modality fusion. Since point clouds provide more accurate localization and geometry information, they could serve as a reliable spatial prior to acquiring relevant semantic information from the images. Therefore, we design a LiDAR-Guided View Transformer (LGVT) to effectively obtain the camera representation in BEV space and thus benefit the whole dual-branch fusion system. LGVT takes camera BEV as the primitive semantic query, repeatedly leveraging the spatial cue of LiDAR BEV for extracting image features across multiple camera views. Moreover, we extend our framework into the temporal domain with our proposed Temporal Deformable Alignment (TDA) module, which aims to aggregate BEV features from multiple historical frames. Including these two modules, our framework dubbed BEVFusion4D achieves state-of-the-art results in 3D object detection, with 72.0% mAP and 73.5% NDS on the nuScenes validation set, and 73.3% mAP and 74.7% NDS on nuScenes test set, respectively.

Jianyong Ai, Wenbo Ding, Jiuhua Zhao et al.

Compared to 2D lanes, real 3D lane data is difficult to collect accurately. In this paper, we propose a novel method for training 3D lanes with only 2D lane labels, called weakly supervised 3D lane detection WS-3D-Lane. By assumptions of constant lane width and equal height on adjacent lanes, we indirectly supervise 3D lane heights in the training. To overcome the problem of the dynamic change of the camera pitch during data collection, a camera pitch self-calibration method is proposed. In anchor representation, we propose a double-layer anchor with a improved non-maximum suppression (NMS) method, which enables the anchor-based method to predict two lane lines that are close. Experiments are conducted on the base of 3D-LaneNet under two supervision methods. Under weakly supervised setting, our WS-3D-Lane outperforms previous 3D-LaneNet: F-score rises to 92.3% on Apollo 3D synthetic dataset, and F1 rises to 74.5% on ONCE-3DLanes. Meanwhile, WS-3D-Lane in purely supervised setting makes more increments and outperforms state-of-the-art. To the best of our knowledge, WS-3D-Lane is the first try of 3D lane detection under weakly supervised setting.

Pengcheng Hao, Huaze Tang, Ercan Engin Kuruoglu et al.

Bayesian deep learning (BDL) provides a principled framework for reliable uncertainty quantification by combining deep neural networks with Bayesian inference. A central challenge in BDL lies in the design of informative prior distributions that scale effectively to high-dimensional data. Recent functional variational inference (VI) approaches address this issue by imposing priors directly in function space; however, most existing methods rely on Gaussian process (GP) priors, whose expressiveness and generalisation capabilities become limited in high-dimensional regimes. In this work, we propose VLM-FS-EB, a novel function-space empirical Bayes regularisation framework, leveraging large vision-language models (VLMs) to generates semantically meaningful context points. These synthetic samples are then used VLMs for embeddings to construct expressive functional priors. Furthermore, the proposed method is evaluated against various baselines, and experimental results demonstrate that our method consistently improves predictive performance and yields more reliable uncertainty estimates, particularly in out-of-distribution (OOD) detection tasks and data-scarce regimes.

Jian Xu, Meiling Yang, Wenbo Ding et al.

The label distribution skew induced data heterogeniety has been shown to be a significant obstacle that limits the model performance in federated learning, which is particularly developed for collaborative model training over decentralized data sources while preserving user privacy. This challenge could be more serious when the participating clients are in unstable circumstances and dropout frequently. Previous work and our empirical observations demonstrate that the classifier head for classification task is more sensitive to label skew and the unstable performance of FedAvg mainly lies in the imbalanced training samples across different classes. The biased classifier head will also impact the learning of feature representations. Therefore, maintaining a balanced classifier head is of significant importance for building a better global model. To this end, we propose a simple yet effective framework by introducing a prior-calibrated softmax function for computing the cross-entropy loss and a prototype-based feature augmentation scheme to re-balance the local training, which are lightweight for edge devices and can facilitate the global model aggregation. The improved model performance over existing baselines in the presence of non-IID data and client dropout is demonstrated by conducting extensive experiments on benchmark classification tasks.

Shoujie Li, Changqing Guo, Jianle Xu et al.

Robotic manipulators for aerospace applications require a delicate balance between lightweight construction and fault-tolerant operation to satisfy strict weight limitations and ensure reliability in remote, hazardous environments. This paper presents Time-Division Multiplexing Actuation (TDMA), a practical approach for tendon-driven robots that significantly reduces actuator count while preserving high torque output and intrinsic fault tolerance. The key hardware employs a vertically-stacked rotational selection structure that integrates self-rotating TDM motors for rapid configuration, electromagnetic clutches enabling sub-0.1 second engagement, a worm gear reducer for enhanced load capacity and self-locking capability, and a dual-encoder system for precise, long-term positioning. Leveraging TDMA, the proposed MuxArm achieves a self-weight of 2.17 kg, supports an actuator driving capacity of 10 kg, and maintains end-effector accuracy up to 1% of its length, even under partial servo failure. Additionally, an actuation space trajectory planning algorithm is developed, enabling fault-tolerant control and reducing tendon load by up to 50% compared to conventional methods. Comprehensive experiments demonstrate MuxArm's robust performance in diverse settings, including free-space, cluttered, and confined environments.

Hengxi Zhang, Huaze Tang, Wenbo Ding et al.

The Space-Air-Ground Integrated Network (SAGIN), integrating heterogeneous devices including low earth orbit (LEO) satellites, unmanned aerial vehicles (UAVs), and ground users (GUs), holds significant promise for advancing smart city applications. However, resource management of the SAGIN is a challenge requiring urgent study in that inappropriate resource management will cause poor data transmission, and hence affect the services in smart cities. In this paper, we develop a comprehensive SAGIN system that encompasses five distinct communication links and propose an efficient cooperative multi-type multi-agent deep reinforcement learning (CMT-MARL) method to address the resource management issue. The experimental results highlight the efficacy of the proposed CMT-MARL, as evidenced by key performance indicators such as the overall transmission rate and transmission success rate. These results underscore the potential value and feasibility of future implementation of the SAGIN.

Ran Yu, Haixin Yu, Shoujie Li et al.

Transparent objects are common in daily life, while their optical properties pose challenges for RGB-D cameras to capture accurate depth information. This issue is further amplified when these objects are hand-held, as hand occlusions further complicate depth estimation. For assistant robots, however, accurately perceiving hand-held transparent objects is critical to effective human-robot interaction. This paper presents a Hand-Aware Depth Restoration (HADR) method based on creating an implicit neural representation function from a single RGB-D image. The proposed method utilizes hand posture as an important guidance to leverage semantic and geometric information of hand-object interaction. To train and evaluate the proposed method, we create a high-fidelity synthetic dataset named TransHand-14K with a real-to-sim data generation scheme. Experiments show that our method has better performance and generalization ability compared with existing methods. We further develop a real-world human-to-robot handover system based on HADR, demonstrating its potential in human-robot interaction applications.

Zihao Zhao, Zhenpeng Shi, Yang Liu et al.

Federated Learning (FL) is often impeded by communication overhead issues. Prompt tuning, as a potential solution, has been introduced to only adjust a few trainable parameters rather than the whole model. However, current single-modality prompt tuning approaches fail to comprehensively portray local clients' data. To overcome this limitation, we present Twin Prompt Federated learning (TPFL), a pioneering solution that integrates both visual and textual modalities, ensuring a more holistic representation of local clients' data characteristics. Furthermore, in order to tackle the data heterogeneity issues, we introduce the Augmented TPFL (ATPFL) employing the contrastive learning to TPFL, which not only enhances the global knowledge acquisition of client models but also fosters the development of robust, compact models. The effectiveness of TPFL and ATPFL is substantiated by our extensive evaluations, consistently showing superior performance compared to all baselines.

Jinkun Liu, Haohan Chi, Lingfeng Zhang et al.

Long-horizon robotic manipulation requires plans that are both logically coherent and geometrically grounded. Existing Vision-Language-Action policies usually hide planning in latent states or expose only one modality: text-only chain-of-thought encodes causal order but misses spatial constraints, while visual prediction provides geometric cues but often remains local and semantically underconstrained. We introduce Interleaved Vision--Language Reasoning (IVLR), a policy framework built around \trace{}, an explicit intermediate representation that alternates textual subgoals with visual keyframes over the full task horizon. At test time, a single native multimodal transformer self-generates this global semantic-geometric trace from the initial observation and instruction, caches it, and conditions a closed-loop action decoder on the trace, original instruction, and current observation. Because standard robot datasets lack such traces, we construct pseudo-supervision by temporally segmenting demonstrations and captioning each stage with a vision-language model. Across simulated benchmarks for long-horizon manipulation and visual distribution shift, \method{} reaches 95.5\% average success on LIBERO, including 92.4\% on LIBERO-Long, and 59.4\% overall success on SimplerEnv-WidowX. Ablations show that both modalities are necessary: without traces, LIBERO-Long success drops to 37.7\%; text-only and vision-only traces reach 62.0\% and 68.4\%, while the full interleaved trace reaches 92.4\%. Stress tests with execution perturbations and masked trace content show moderate degradation, suggesting that the trace can tolerate local corruption and moderate execution drift, but remains limited under stale or incorrect global plans.

Junhao Gong, Shoujie Li, Kit-Wa Sou et al.

Conventional suction cups lack sensing capabilities for contact-aware manipulation in unstructured environments. This paper presents FlexiCup, a multimodal suction cup with wireless electronics that integrate dual-zone vision-tactile sensing. The central zone dynamically switches between vision and tactile modalities via illumination control, while the peripheral zone provides continuous spatial awareness. The modular mechanical design supports both vacuum (sustained-contact adhesion) and Bernoulli (contactless lifting) actuation while maintaining the identical dual-zone sensing architecture, demonstrating sensing-actuation decoupling where sensing and actuation principles are orthogonally separable. We validate hardware versatility through dual control paradigms. Modular perception-driven grasping achieves comparable success rates across vacuum (90.0%) and Bernoulli (86.7%) modes using identical sensing and control pipelines, validating the sensing architecture's effectiveness across fundamentally different pneumatic principles. Diffusion-based end-to-end learning achieves 73.3% and 66.7% success on contact-aware manipulation tasks, with ablation studies confirming 13% improvements from multi-head attention coordinating dual-zone observations. Hardware designs, firmware, and experimental videos are available at the companion website: https://flexicup.junhaogong.top.

Xingting Li, Yifan Xie, Han Liu et al.

Robotic contact-rich and fine-grained manipulation remains a significant challenge due to complex interaction dynamics and the competing requirements of multi-timescale control. While current visual imitation learning methods excel at long-horizon planning, they often fail to perceive critical interaction cues like friction variations or incipient slip, and struggle to balance global task coherence with local reactive feedback. To address these challenges, we propose M2-ResiPolicy, a novel Master-Micro residual control architecture that synergizes high-level action guidance with low-level correction. The framework consists of a Master-Guidance Policy (MGP) operating at 10 Hz, which generates temporally consistent action chunks via a diffusion-based backbone and employs a tactile-intensity-driven adaptive fusion mechanism to dynamically modulate perceptual weights between vision and touch. Simultaneously, a high-frequency (60 Hz) Micro-Residual Corrector (MRC) utilizes a lightweight GRU to provide real-time action compensation based on TCP wrench feedback. This policy is further integrated with a force-mixed PBIC execution layer, effectively regulating contact forces to ensure interaction safety. Experiments across several demanding tasks including fragile object grasping and precision insertion, demonstrate that M2-ResiPolicy significantly outperforms standard Diffusion Policy (DP) and state-of-the-art Reactive Diffusion Policy (RDP), achieving a 93\% damage-free success rate in chip grasping and superior force regulation stability.

Hongsheng Li, Lingfeng Zhang, Zexian Yang et al.

Robust 3D object detection in adverse weather is highly challenging due to the varying reliability of different sensors. While existing LiDAR-4D radar fusion methods improve robustness, they predominantly rely on fixed or weakly adaptive pipelines, failing to dy-namically adjust modality preferences as environmental conditions change. To bridge this gap, we reformulate multi-modal perception as a weather-conditioned branch routing problem. Instead of computing a single fused output, our framework explicitly maintains three parallel 3D feature streams: a pure LiDAR branch, a pure 4D radar branch, and a condition-gated fusion branch. Guided by a condition token extracted from visual and semantic prompts, a lightweight router dynamically predicts sample-specific weights to softly aggregate these representations. Furthermore, to prevent branch collapse, we introduce a weather-supervised learning strategy with auxiliary classification and diversity regularization to enforce distinct, condition-dependent routing behaviors. Extensive experiments on the K-Radar benchmark demonstrate that our method achieves state-of-the-art performance. Furthermore, it provides explicit and highly interpretable insights into modality preferences, transparently revealing how adaptive routing robustly shifts reliance between LiDAR and 4D radar across diverse adverse-weather scenarios. The source code with be released.

Tengbo Yu, Guanxing Lu, Zaijia Yang et al.

Multi-task robotic bimanual manipulation is becoming increasingly popular as it enables sophisticated tasks that require diverse dual-arm collaboration patterns. Compared to unimanual manipulation, bimanual tasks pose challenges to understanding the multi-body spatiotemporal dynamics. An existing method ManiGaussian pioneers encoding the spatiotemporal dynamics into the visual representation via Gaussian world model for single-arm settings, which ignores the interaction of multiple embodiments for dual-arm systems with significant performance drop. In this paper, we propose ManiGaussian++, an extension of ManiGaussian framework that improves multi-task bimanual manipulation by digesting multi-body scene dynamics through a hierarchical Gaussian world model. To be specific, we first generate task-oriented Gaussian Splatting from intermediate visual features, which aims to differentiate acting and stabilizing arms for multi-body spatiotemporal dynamics modeling. We then build a hierarchical Gaussian world model with the leader-follower architecture, where the multi-body spatiotemporal dynamics is mined for intermediate visual representation via future scene prediction. The leader predicts Gaussian Splatting deformation caused by motions of the stabilizing arm, through which the follower generates the physical consequences resulted from the movement of the acting arm. As a result, our method significantly outperforms the current state-of-the-art bimanual manipulation techniques by an improvement of 20.2% in 10 simulated tasks, and achieves 60% success rate on average in 9 challenging real-world tasks. Our code is available at https://github.com/April-Yz/ManiGaussian_Bimanual.

Jingjing Fan, Yushan Liu, Shoujie Li et al.

General vision-language models increasingly support unified spatiotemporal reasoning over long video streams, yet deploying such capabilities on robots remains constrained by the prohibitive latency of processing long-horizon histories and generating high-dimensional future predictions. To bridge this gap, we present FUTURE-VLA, a unified architecture that reformulates long-horizon control and future forecasting as a monolithic sequence-generation task. Adopting a dual-sided efficiency paradigm, FUTURE-VLA leverages a temporally adaptive compression strategy to maximize spatiotemporal information density, enabling the ingestion of extensive multi-view histories while maintaining constant inference latency. Simultaneously, it performs latent-space autoregression to align actionable dynamics with reviewable visual look-aheads in a single forward pass. These real-time predictive capabilities further enable a prediction-guided Human-In-the-Loop mechanism via interactive execution gating, allowing operators to dynamically validate behaviors based on interpretable future previews. Extensive evaluations demonstrate that FUTURE-VLA establishes new state-of-the-art performance, attaining success rates of 99.2% on LIBERO, 75.4% on RoboTwin, and 78.0% on a real-world Piper platform, all with a $16\times$ extended spatiotemporal window while maintaining the inference latency of a single-frame baseline.

Yinghao Shuai, Ran Yu, Yuantao Chen et al.

Current robotic systems can understand the categories and poses of objects well. But understanding physical properties like mass, friction, and hardness, in the wild, remains challenging. We propose a new method that reconstructs 3D objects using the Gaussian splatting representation and predicts various physical properties in a zero-shot manner. We propose two techniques during the reconstruction phase: a geometry-aware regularization loss function to improve the shape quality and a region-aware feature contrastive loss function to promote region affinity. Two other new techniques are designed during inference: a feature-based property propagation module and a volume integration module tailored for the Gaussian representation. Our framework is named as zero-shot physical understanding with Gaussian splatting, or PUGS. PUGS achieves new state-of-the-art results on the standard benchmark of ABO-500 mass prediction. We provide extensive quantitative ablations and qualitative visualization to demonstrate the mechanism of our designs. We show the proposed methodology can help address challenging real-world grasping tasks. Our codes, data, and models are available at https://github.com/EverNorif/PUGS

Zelai Xu, Zhexuan Xu, Ruize Zhang et al.

Recent advancements in Large Language Models (LLMs) have largely focused on depth scaling, where a single agent solves long-horizon problems with multi-turn reasoning and tool use. However, as tasks grow broader, the key bottleneck shifts from individual competence to organizational capability. In this work, we explore a complementary dimension of width scaling with multi-agent systems to address broad information seeking. Existing multi-agent systems often rely on hand-crafted workflows and turn-taking interactions that fail to parallelize work effectively. To bridge this gap, we propose WideSeek-R1, a lead-agent-subagent framework trained via multi-agent reinforcement learning (MARL) to synergize scalable orchestration and parallel execution. By utilizing a shared LLM with isolated contexts and specialized tools, WideSeek-R1 jointly optimizes the lead agent and parallel subagents on a curated dataset of 20k broad information-seeking tasks. Extensive experiments show that WideSeek-R1-4B achieves an item F1 score of 40.0% on the WideSearch benchmark, which is comparable to the performance of single-agent DeepSeek-R1-671B. Furthermore, WideSeek-R1-4B exhibits consistent performance gains as the number of parallel subagents increases, highlighting the effectiveness of width scaling.

Jianyi Zhou, Ziteng Gao, Feiyang Hong et al.

Egocentric human video data, which captures rich human-environment interactions and can be collected at scale, has become a key driver of embodied intelligence research. However, existing egocentric datasets typically lack tactile sensing, a critical modality that provides direct cues about contact, force, and pressure in human-object interaction. Without such signals, models struggle to learn physically grounded representations of real-world interaction dynamics. While tactile sensors provide these cues, deploying high-quality tactile hardware at scale remains expensive and cumbersome. This raises a central question: can tactile feedback be inferred directly from visual observations, enabling scalable tactile supervision for egocentric video data and supporting physically grounded embodied learning? To enable research in this direction, we introduce EgoTouch, a large-scale multi-view egocentric dataset with dense tactile supervision for bimanual hand-object interaction. EgoTouch comprises 208 manipulation tasks spanning 1,891 episodes in diverse indoor and outdoor environments, with synchronized multi-view RGB (head-mounted egocentric and dual wrist-mounted cameras), bimanual 3D hand pose, and continuous pressure maps from wearable tactile sensors. Building on EgoTouch, we introduce TouchAnything, a baseline multi-view vision-to-touch prediction framework that uses the egocentric view as the primary input and flexibly leverages available wrist-mounted views at inference time. Experiments show that incorporating wrist-mounted views generally improves tactile prediction over egocentric-only input, achieving up to 5.0% relative improvement in Contact IoU and 6.1% relative improvement in Volumetric IoU. We will publicly release the dataset, code, and benchmark.

Jiahao Lai, Jiaqi Li, Jian Xu et al.

Federated Learning (FL) offers a decentralized approach to model training, where data remains local and only model parameters are shared between the clients and the central server. Traditional methods, such as Federated Averaging (FedAvg), linearly aggregate these parameters which are usually trained on heterogeneous data distributions, potentially overlooking the complex, high-dimensional nature of the parameter space. This can result in degraded performance of the aggregated model. While personalized FL approaches can mitigate the heterogeneous data issue to some extent, the limitation of linear aggregation remains unresolved. To alleviate this issue, we investigate the generative approach of diffusion model and propose a novel generative parameter aggregation framework for personalized FL, \texttt{pFedGPA}. In this framework, we deploy a diffusion model on the server to integrate the diverse parameter distributions and propose a parameter inversion method to efficiently generate a set of personalized parameters for each client. This inversion method transforms the uploaded parameters into a latent code, which is then aggregated through denoising sampling to produce the final personalized parameters. By encoding the dependence of a client's model parameters on the specific data distribution using the high-capacity diffusion model, \texttt{pFedGPA} can effectively decouple the complexity of the overall distribution of all clients' model parameters from the complexity of each individual client's parameter distribution. Our experimental results consistently demonstrate the superior performance of the proposed method across multiple datasets, surpassing baseline approaches.

Xiaoshuai Hao, Lei Zhou, Zhijian Huang et al.

We open-source MiMo-Embodied, the first cross-embodied foundation model to successfully integrate and achieve state-of-the-art performance in both Autonomous Driving and Embodied AI. MiMo-Embodied sets new records across 17 embodied AI benchmarks in Task Planning, Affordance Prediction and Spatial Understanding, while also excelling in 12 autonomous driving benchmarks across Environmental Perception, Status Prediction, and Driving Planning. Across these tasks, MiMo-Embodied significantly outperforms existing open-source, closed-source, and specialized baselines. Our results indicate that through multi-stage learning, curated data construction, and CoT/RL fine-tuning, these two domains exhibit strong positive transfer and mutually reinforce one another. We provide a detailed analysis of our model design and training methodologies to facilitate further research. Code and models are available at https://github.com/XiaomiMiMo/MiMo-Embodied.

Lingfeng Zhang, Yuchen Zhang, Hongsheng Li et al.

Vision-Language Models (VLMs), leveraging their powerful visual perception and reasoning capabilities, have been widely applied in Unmanned Aerial Vehicle (UAV) tasks. However, the spatial intelligence capabilities of existing VLMs in UAV scenarios remain largely unexplored, raising concerns about their effectiveness in navigating and interpreting dynamic environments. To bridge this gap, we introduce SpatialSky-Bench, a comprehensive benchmark specifically designed to evaluate the spatial intelligence capabilities of VLMs in UAV navigation. Our benchmark comprises two categories-Environmental Perception and Scene Understanding-divided into 13 subcategories, including bounding boxes, color, distance, height, and landing safety analysis, among others. Extensive evaluations of various mainstream open-source and closed-source VLMs reveal unsatisfactory performance in complex UAV navigation scenarios, highlighting significant gaps in their spatial capabilities. To address this challenge, we developed the SpatialSky-Dataset, a comprehensive dataset containing 1M samples with diverse annotations across various scenarios. Leveraging this dataset, we introduce Sky-VLM, a specialized VLM designed for UAV spatial reasoning across multiple granularities and contexts. Extensive experimental results demonstrate that Sky-VLM achieves state-of-the-art performance across all benchmark tasks, paving the way for the development of VLMs suitable for UAV scenarios. The source code is available at https://github.com/linglingxiansen/SpatialSKy.

Chao Wang, Tao Yang, Hongtao Tian et al.

Critic-free methods like GRPO reduce memory demands by estimating advantages from multiple rollouts but tend to converge slowly, as critical learning signals are diluted by an abundance of uninformative samples and tokens. To tackle this challenge, we propose the \textbf{Dynamic Dual-Level Down-Sampling (D$^3$S)} framework that prioritizes the most informative samples and tokens across groups to improve the efficient of policy optimization. D$^3$S operates along two levels: (1) the sample-level, which selects a subset of rollouts to maximize advantage variance ($\text{Var}(A)$). We theoretically proven that this selection is positively correlated with the upper bound of the policy gradient norms, yielding higher policy gradients. (2) the token-level, which prioritizes tokens with a high product of advantage magnitude and policy entropy ($|A_{i,t}|\times H_{i,t}$), focusing updates on tokens where the policy is both uncertain and impactful. Moreover, to prevent overfitting to high-signal data, D$^3$S employs a dynamic down-sampling schedule inspired by curriculum learning. This schedule starts with aggressive down-sampling to accelerate early learning and gradually relaxes to promote robust generalization. Extensive experiments on Qwen2.5 and Llama3.1 demonstrate that integrating D$^3$S into advanced RL algorithms achieves state-of-the-art performance and generalization while requiring \textit{fewer} samples and tokens across diverse reasoning benchmarks. Our code is added in the supplementary materials and will be made publicly available.

Ziqi Yuan, Liyuan Wang, Wenbo Ding et al.

In real-world applications, an object detector often encounters object instances from new classes and needs to accommodate them effectively. Previous work formulated this critical problem as incremental object detection (IOD), which assumes the object instances of new classes to be fully annotated in incremental data. However, as supervisory signals are usually rare and expensive, the supervised IOD may not be practical for implementation. In this work, we consider a more realistic setting named semi-supervised IOD (SSIOD), where the object detector needs to learn new classes incrementally from a few labelled data and massive unlabelled data without catastrophic forgetting of old classes. A commonly-used strategy for supervised IOD is to encourage the current model (as a student) to mimic the behavior of the old model (as a teacher), but it generally fails in SSIOD because a dominant number of object instances from old and new classes are coexisting and unlabelled, with the teacher only recognizing a fraction of them. Observing that learning only the classes of interest tends to preclude detection of other classes, we propose to bridge the coexistence of unlabelled classes by constructing two teacher models respectively for old and new classes, and using the concatenation of their predictions to instruct the student. This approach is referred to as DualTeacher, which can serve as a strong baseline for SSIOD with limited resource overhead and no extra hyperparameters. We build various benchmarks for SSIOD and perform extensive experiments to demonstrate the superiority of our approach (e.g., the performance lead is up to 18.28 AP on MS-COCO). Our code is available at \url{https://github.com/chuxiuhong/DualTeacher}.

Yushan Liu, Peibo Sun, Shoujie Li et al.

World Action Models (WAMs) enhance Vision-Language-Action policies by jointly predicting scene evolution and robot actions, but existing methods usually represent the predicted world as holistic images, video tokens, or global latents. These representations are difficult for an action decoder to address when an instruction refers to a particular object, especially under scene shifts where object identity is entangled with context. We propose OA-WAM, an Object-Addressable World Action Model for robust robot manipulation. OA-WAM decomposes each frame into N+1 slot states, with one robot slot and N object slots. Each slot contains a persistent address vector and a time-varying content vector, and is fused with text, image, proprioception, and past-action tokens in a block-causal sequence. A world head predicts next-frame slot states, while a flow-matching action head decodes a 16-step continuous action chunk in the same forward pass. Addressability is enforced by routing cross-slot attention through address-only keys and resetting the address slice at every transformer layer, separating which object to act on from what that object currently is without adding extra tokens. OA-WAM matches strong VLA and WAM baselines on LIBERO (97.8%) and SimplerEnv (79.3%), reaches state-of-the-art performance on the most relevant LIBERO-Plus geometric axes, and remains competitive on the seven-axis aggregate. A causal slot-intervention test yields a swap-binding cosine of 0.87, versus at most 0.09 for holistic baselines. These results suggest that addressable object states provide an effective interface for robust world-action modeling under scene perturbations.

Lingfeng Zhang, Xiaoshuai Hao, Xizhou Bu et al.

Assisting humans in open-world outdoor environments requires robots to translate high-level natural-language intentions into safe, long-horizon, and socially compliant navigation behavior. Existing map-based methods rely on costly pre-built HD maps, while learning-based policies are mostly limited to indoor and short-horizon settings. To bridge this gap, we propose Walk with Me, a map-free framework for long-horizon social navigation from high-level human instructions. Walk with Me leverages GPS context and lightweight candidate points-of-interest from a public map API for semantic destination grounding and waypoint proposal. A High-Level Vision-Language Model grounds abstract instructions into concrete destinations and plans coarse waypoint sequences. During execution, an observation-aware routing mechanism determines whether the Low-Level Vision-Language-Action policy can handle the current situation or whether explicit safety reasoning from the High-Level VLM is needed. Routine segments are executed by the Low-Level VLA, while complex situations such as crowded crossings trigger high-level reasoning and stop-and-wait behavior when unsafe. By combining semantic intent grounding, map-free long-horizon planning, safety-aware reasoning, and low-level action generation, Walk with Me enables practical outdoor social navigation for human-centric assistance.

Zhengru Fang, Yu Guo, Fei Liu et al.

Visual reinforcement learning aims to empower an agent to learn policies from visual observations, yet it remains vulnerable to dynamic visual perturbations, such as unpredictable shifts in corruption types. To systematically study this, we introduce the Visual Degraded Control Suite (VDCS), a benchmark extending DeepMind Control Suite with Markov-switching degradations to simulate non-stationary real-world perturbations. Experiments on VDCS reveal severe performance degradation in existing methods. We theoretically prove via information-theoretic analysis that this failure stems from reconstruction-based objectives inevitably entangling perturbation artifacts into latent representations. To mitigate this negative impact, we propose Agent-Centric Observations with Mixture-of-Experts (ACO-MoE) to robustify visual RL against perturbations. The proposed framework leverages unique agent-centric restoration experts, achieving restoration from corruptions and task-relevant foreground extraction, thereby decoupling perception from perturbation before being processed by the RL agent. Extensive experiments on VDCS show our ACO-MoE outperforms strong baselines, recovering 95.3% of clean performance under challenging Markov-switching corruptions. Moreover, it achieves SOTA results on DMControl Generalization with random-color and video-background perturbations, demonstrating a high level of robustness.

Yifan Xie, YuAn Wang, Guangyu Chen et al.

Human videos contain rich manipulation priors, but using them for robot learning remains difficult because raw observations entangle scene understanding, human motion, and embodiment-specific action. We introduce MoT-HRA, a hierarchical vision-language-action framework that learns human-intention priors from large-scale human demonstrations. We first curate HA-2.2M, a 2.2M-episode action-language dataset reconstructed from heterogeneous human videos through hand-centric filtering, spatial reconstruction, temporal segmentation, and language alignment. On top of this dataset, MoT-HRA factorizes manipulation into three coupled experts: a vision-language expert predicts an embodiment-agnostic 3D trajectory, an intention expert models MANO-style hand motion as a latent human-motion prior, and a fine expert maps the intention-aware representation to robot action chunks. A shared-attention trunk and read-only key-value transfer allow downstream control to use human priors while limiting interference with upstream representations. Experiments on hand motion generation, simulated manipulation, and real-world robot tasks show that MoT-HRA improves motion plausibility and robust control under distribution shift.

Keyan Guo, Ayush Utkarsh, Wenbo Ding et al.

Online user generated content games (UGCGs) are increasingly popular among children and adolescents for social interaction and more creative online entertainment. However, they pose a heightened risk of exposure to explicit content, raising growing concerns for the online safety of children and adolescents. Despite these concerns, few studies have addressed the issue of illicit image-based promotions of unsafe UGCGs on social media, which can inadvertently attract young users. This challenge arises from the difficulty of obtaining comprehensive training data for UGCG images and the unique nature of these images, which differ from traditional unsafe content. In this work, we take the first step towards studying the threat of illicit promotions of unsafe UGCGs. We collect a real-world dataset comprising 2,924 images that display diverse sexually explicit and violent content used to promote UGCGs by their game creators. Our in-depth studies reveal a new understanding of this problem and the urgent need for automatically flagging illicit UGCG promotions. We additionally create a cutting-edge system, UGCG-Guard, designed to aid social media platforms in effectively identifying images used for illicit UGCG promotions. This system leverages recently introduced large vision-language models (VLMs) and employs a novel conditional prompting strategy for zero-shot domain adaptation, along with chain-of-thought (CoT) reasoning for contextual identification. UGCG-Guard achieves outstanding results, with an accuracy rate of 94% in detecting these images used for the illicit promotion of such games in real-world scenarios.

Siqi Ping, Yuzhu Mao, Yang Liu et al.

Although large-scale pre-trained models hold great potential for adapting to downstream tasks through fine-tuning, the performance of such fine-tuned models is often limited by the difficulty of collecting sufficient high-quality, task-specific data. Federated Learning (FL) offers a promising solution by enabling fine-tuning across large-scale clients with a variety of task data, but it is bottlenecked by significant communication overhead due to the pre-trained models' extensive size. This paper addresses the high communication cost for fine-tuning large pre-trained models within FL frameworks through low-rank fine-tuning. Specifically, we train a low-rank adapter for each individual task on the client side, followed by server-side clustering for similar group of adapters to achieve task-specific aggregation. Extensive experiments on various language and vision tasks, such as GLUE and CIFAR-10/100, reveal the evolution of task-specific adapters throughout the FL training process and verify the effectiveness of the proposed low-rank task-specific adapter clustering (TAC) method.

Yixian Zhang, Huaze Tang, Chao Wang et al.

Reinforcement learning (RL) policies represented in Reproducing Kernel Hilbert Spaces (RKHS) offer powerful representational capabilities. While second-order optimization methods like Newton's method demonstrate faster convergence than first-order approaches, current RKHS-based policy optimization remains constrained to first-order techniques. This limitation stems primarily from the intractability of explicitly computing and inverting the infinite-dimensional Hessian operator in RKHS. We introduce Policy Newton in RKHS, the first second-order optimization framework specifically designed for RL policies represented in RKHS. Our approach circumvents direct computation of the inverse Hessian operator by optimizing a cubic regularized auxiliary objective function. Crucially, we leverage the Representer Theorem to transform this infinite-dimensional optimization into an equivalent, computationally tractable finite-dimensional problem whose dimensionality scales with the trajectory data volume. We establish theoretical guarantees proving convergence to a local optimum with a local quadratic convergence rate. Empirical evaluations on a toy financial asset allocation problem validate these theoretical properties, while experiments on standard RL benchmarks demonstrate that Policy Newton in RKHS achieves superior convergence speed and higher episodic rewards compared to established first-order RKHS approaches and parametric second-order methods. Our work bridges a critical gap between non-parametric policy representations and second-order optimization methods in reinforcement learning.

Yifan Xie, Mingyang Li, Shoujie Li et al.

Tactile perception is essential for embodied agents to understand physical attributes of objects that cannot be determined through visual inspection alone. While existing approaches have made progress in visual and language modalities for physical understanding, they fail to effectively incorporate tactile information that provides crucial haptic feedback for real-world interaction. In this paper, we present VTV-LLM, the first multi-modal large language model for universal Visuo-Tactile Video (VTV) understanding that bridges the gap between tactile perception and natural language. To address the challenges of cross-sensor and cross-modal integration, we contribute VTV150K, a comprehensive dataset comprising 150,000 video frames from 100 diverse objects captured across three different tactile sensors (GelSight Mini, DIGIT, and Tac3D), annotated with four fundamental tactile attributes (hardness, protrusion, elasticity, and friction). We develop a novel three-stage training paradigm that includes VTV enhancement for robust visuo-tactile representation, VTV-text alignment for cross-modal correspondence, and text prompt finetuning for natural language generation. Our framework enables sophisticated tactile reasoning capabilities including feature assessment, comparative analysis, scenario-based decision making and so on. Experimental evaluations demonstrate that VTV-LLM achieves superior performance in tactile video understanding tasks, establishing a foundation for more intuitive human-machine interaction in tactile domains.

Zelai Xu, Ruize Zhang, Chao Yu et al. · tsinghua

Robot sports, characterized by well-defined objectives, explicit rules, and dynamic interactions, present ideal scenarios for demonstrating embodied intelligence. In this paper, we present VolleyBots, a novel robot sports testbed where multiple drones cooperate and compete in the sport of volleyball under physical dynamics. VolleyBots integrates three features within a unified platform: competitive and cooperative gameplay, turn-based interaction structure, and agile 3D maneuvering. These intertwined features yield a complex problem combining motion control and strategic play, with no available expert demonstrations. We provide a comprehensive suite of tasks ranging from single-drone drills to multi-drone cooperative and competitive tasks, accompanied by baseline evaluations of representative reinforcement learning (RL), multi-agent reinforcement learning (MARL) and game-theoretic algorithms. Simulation results show that on-policy RL methods outperform off-policy methods in single-agent tasks, but both approaches struggle in complex tasks that combine motion control and strategic play. We additionally design a hierarchical policy which achieves 69.5% win rate against the strongest baseline in the 3 vs 3 task, demonstrating its potential for tackling the complex interplay between low-level control and high-level strategy. To highlight VolleyBots' sim-to-real potential, we further demonstrate the zero-shot deployment of a policy trained entirely in simulation on real-world drones.

Erjia Xiao, Lingfeng Zhang, Yingbo Tang et al.

In this report, we describe the technical details of our submission to the IROS 2025 RoboSense Challenge Social Navigation Track. This track focuses on developing RGBD-based perception and navigation systems that enable autonomous agents to navigate safely, efficiently, and socially compliantly in dynamic human-populated indoor environments. The challenge requires agents to operate from an egocentric perspective using only onboard sensors including RGB-D observations and odometry, without access to global maps or privileged information, while maintaining social norm compliance such as safe distances and collision avoidance. Building upon the Falcon model, we introduce a Proactive Risk Perception Module to enhance social navigation performance. Our approach augments Falcon with collision risk understanding that learns to predict distance-based collision risk scores for surrounding humans, which enables the agent to develop more robust spatial awareness and proactive collision avoidance behaviors. The evaluation on the Social-HM3D benchmark demonstrates that our method improves the agent's ability to maintain personal space compliance while navigating toward goals in crowded indoor scenes with dynamic human agents, achieving 2nd place among 16 participating teams in the challenge.

Yixian Zhang, Shu'ang Yu, Tonghe Zhang et al.

Training expressive flow-based policies with off-policy reinforcement learning is notoriously unstable due to gradient pathologies in the multi-step action sampling process. We trace this instability to a fundamental connection: the flow rollout is algebraically equivalent to a residual recurrent computation, making it susceptible to the same vanishing and exploding gradients as RNNs. To address this, we reparameterize the velocity network using principles from modern sequential models, introducing two stable architectures: Flow-G, which incorporates a gated velocity, and Flow-T, which utilizes a decoded velocity. We then develop a practical SAC-based algorithm, enabled by a noise-augmented rollout, that facilitates direct end-to-end training of these policies. Our approach supports both from-scratch and offline-to-online learning and achieves state-of-the-art performance on continuous control and robotic manipulation benchmarks, eliminating the need for common workarounds like policy distillation or surrogate objectives.

Yixian Zhang, Huaze Tang, Changxu Wei et al.

The Soft Actor-Critic (SAC) algorithm, a state-of-the-art method in maximum entropy reinforcement learning, traditionally relies on minimizing reverse Kullback-Leibler (KL) divergence for policy updates. However, this approach leads to an intractable optimal projection policy, necessitating gradient-based approximations that can suffer from instability and poor sample efficiency. This paper investigates the alternative use of forward KL divergence within SAC. We demonstrate that for Gaussian policies, forward KL divergence yields an explicit optimal projection policy -- corresponding to the mean and variance of the target Boltzmann distribution's action marginals. Building on the distinct advantages of both KL directions, we propose Bidirectional SAC, an algorithm that first initializes the policy using the explicit forward KL projection and then refines it by optimizing the reverse KL divergence. Comprehensive experiments on continuous control benchmarks show that Bidirectional SAC significantly outperforms standard SAC and other baselines, achieving up to a $30\%$ increase in episodic rewards, alongside enhanced sample efficiency.

Ruize Zhang, Sirui Xiang, Zelai Xu et al. · tsinghua

In this paper, we tackle the problem of learning to play 3v3 multi-drone volleyball, a new embodied competitive task that requires both high-level strategic coordination and low-level agile control. The task is turn-based, multi-agent, and physically grounded, posing significant challenges due to its long-horizon dependencies, tight inter-agent coupling, and the underactuated dynamics of quadrotors. To address this, we propose Hierarchical Co-Self-Play (HCSP), a hierarchical reinforcement learning framework that separates centralized high-level strategic decision-making from decentralized low-level motion control. We design a three-stage population-based training pipeline to enable both strategy and skill to emerge from scratch without expert demonstrations: (I) training diverse low-level skills, (II) learning high-level strategy via self-play with fixed low-level skills, and (III) joint fine-tuning through co-self-play. Experiments show that HCSP achieves superior performance, outperforming non-hierarchical self-play and rule-based hierarchical baselines with an average 82.9% win rate and a 71.5% win rate against the two-stage variant. Moreover, co-self-play leads to emergent team behaviors such as role switching and coordinated formations, demonstrating the effectiveness of our hierarchical design and training scheme. The project page is at https://sites.google.com/view/hi-co-self-play.

Guanxing Lu, Zifeng Gao, Tianxing Chen et al.

Diffusion models have been verified to be effective in generating complex distributions from natural images to motion trajectories. Recent diffusion-based methods show impressive performance in 3D robotic manipulation tasks, whereas they suffer from severe runtime inefficiency due to multiple denoising steps, especially with high-dimensional observations. To this end, we propose a real-time robotic manipulation model named ManiCM that imposes the consistency constraint to the diffusion process, so that the model can generate robot actions in only one-step inference. Specifically, we formulate a consistent diffusion process in the robot action space conditioned on the point cloud input, where the original action is required to be directly denoised from any point along the ODE trajectory. To model this process, we design a consistency distillation technique to predict the action sample directly instead of predicting the noise within the vision community for fast convergence in the low-dimensional action manifold. We evaluate ManiCM on 31 robotic manipulation tasks from Adroit and Metaworld, and the results demonstrate that our approach accelerates the state-of-the-art method by 10 times in average inference speed while maintaining competitive average success rate.