Wenhao Tang

Jiayu Chen, Chao Yu, Guosheng Li et al. · tsinghua

Multi-UAV pursuit-evasion, where pursuers aim to capture evaders, poses a key challenge for UAV swarm intelligence. Multi-agent reinforcement learning (MARL) has demonstrated potential in modeling cooperative behaviors, but most RL-based approaches remain constrained to simplified simulations with limited dynamics or fixed scenarios. Previous attempts to deploy RL policy to real-world pursuit-evasion are largely restricted to two-dimensional scenarios, such as ground vehicles or UAVs at fixed altitudes. In this paper, we address multi-UAV pursuit-evasion by considering UAV dynamics and physical constraints. We introduce an evader prediction-enhanced network to tackle partial observability in cooperative strategy learning. Additionally, we propose an adaptive environment generator within MARL training, enabling higher exploration efficiency and better policy generalization across diverse scenarios. Simulations show our method significantly outperforms all baselines in challenging scenarios, generalizing to unseen scenarios with a 100% capture rate. Finally, we derive a feasible policy via a two-stage reward refinement and deploy the policy on real quadrotors in a zero-shot manner. To our knowledge, this is the first work to derive and deploy an RL-based policy using collective thrust and body rates control commands for multi-UAV pursuit-evasion in unknown environments. The open-source code and videos are available at https://sites.google.com/view/pursuit-evasion-rl.

Wenhao Tang, Sheng Huang, Xiaoxian Zhang et al.

The whole slide image (WSI) classification is often formulated as a multiple instance learning (MIL) problem. Since the positive tissue is only a small fraction of the gigapixel WSI, existing MIL methods intuitively focus on identifying salient instances via attention mechanisms. However, this leads to a bias towards easy-to-classify instances while neglecting hard-to-classify instances. Some literature has revealed that hard examples are beneficial for modeling a discriminative boundary accurately. By applying such an idea at the instance level, we elaborate a novel MIL framework with masked hard instance mining (MHIM-MIL), which uses a Siamese structure (Teacher-Student) with a consistency constraint to explore the potential hard instances. With several instance masking strategies based on attention scores, MHIM-MIL employs a momentum teacher to implicitly mine hard instances for training the student model, which can be any attention-based MIL model. This counter-intuitive strategy essentially enables the student to learn a better discriminating boundary. Moreover, the student is used to update the teacher with an exponential moving average (EMA), which in turn identifies new hard instances for subsequent training iterations and stabilizes the optimization. Experimental results on the CAMELYON-16 and TCGA Lung Cancer datasets demonstrate that MHIM-MIL outperforms other latest methods in terms of performance and training cost. The code is available at: https://github.com/DearCaat/MHIM-MIL.

Wenhao Tang, Junding Sun, Shuihua Wang et al.

In recent years, the rapid development of deep learning has led to a wide range of applications in the field of medical image classification. The variants of neural network models with ever-increasing performance share some commonalities: to try to mitigate overfitting, improve generalization, avoid gradient vanishing and exploding, etc. AlexNet first utilizes the dropout technique to mitigate overfitting and the ReLU activation function to avoid gradient vanishing. Therefore, we focus our discussion on AlexNet, which has contributed greatly to the development of CNNs in 2012. After reviewing over 40 papers, including journal papers and conference papers, we give a narrative on the technical details, advantages, and application areas of AlexNet.

Zhongying Deng, Cheng Tang, Ziyan Huang et al. · pku

Foundation models have demonstrated remarkable success across diverse domains and tasks, primarily due to the thrive of large-scale, diverse, and high-quality datasets. However, in the field of medical imaging, the curation and assembling of such medical datasets are highly challenging due to the reliance on clinical expertise and strict ethical and privacy constraints, resulting in a scarcity of large-scale unified medical datasets and hindering the development of powerful medical foundation models. In this work, we present the largest survey to date of medical image datasets, covering over 1,000 open-access datasets with a systematic catalog of their modalities, tasks, anatomies, annotations, limitations, and potential for integration. Our analysis exposes a landscape that is modest in scale, fragmented across narrowly scoped tasks, and unevenly distributed across organs and modalities, which in turn limits the utility of existing medical image datasets for developing versatile and robust medical foundation models. To turn fragmentation into scale, we propose a metadata-driven fusion paradigm (MDFP) that integrates public datasets with shared modalities or tasks, thereby transforming multiple small data silos into larger, more coherent resources. Building on MDFP, we release an interactive discovery portal that enables end-to-end, automated medical image dataset integration, and compile all surveyed datasets into a unified, structured table that clearly summarizes their key characteristics and provides reference links, offering the community an accessible and comprehensive repository. By charting the current terrain and offering a principled path to dataset consolidation, our survey provides a practical roadmap for scaling medical imaging corpora, supporting faster data discovery, more principled dataset creation, and more capable medical foundation models.

Heng Fang, Sheng Huang, Wenhao Tang et al.

Multiple Instance Learning (MIL) represents the predominant framework in Whole Slide Image (WSI) classification, covering aspects such as sub-typing, diagnosis, and beyond. Current MIL models predominantly rely on instance-level features derived from pretrained models such as ResNet. These models segment each WSI into independent patches and extract features from these local patches, leading to a significant loss of global spatial context and restricting the model's focus to merely local features. To address this issue, we propose a novel MIL framework, named SAM-MIL, that emphasizes spatial contextual awareness and explicitly incorporates spatial context by extracting comprehensive, image-level information. The Segment Anything Model (SAM) represents a pioneering visual segmentation foundational model that can capture segmentation features without the need for additional fine-tuning, rendering it an outstanding tool for extracting spatial context directly from raw WSIs. Our approach includes the design of group feature extraction based on spatial context and a SAM-Guided Group Masking strategy to mitigate class imbalance issues. We implement a dynamic mask ratio for different segmentation categories and supplement these with representative group features of categories. Moreover, SAM-MIL divides instances to generate additional pseudo-bags, thereby augmenting the training set, and introduces consistency of spatial context across pseudo-bags to further enhance the model's performance. Experimental results on the CAMELYON-16 and TCGA Lung Cancer datasets demonstrate that our proposed SAM-MIL model outperforms existing mainstream methods in WSIs classification. Our open-source implementation code is is available at https://github.com/FangHeng/SAM-MIL.

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.

Wenhao Tang, Sheng Huang, Xiaoxian Zhang et al.

Automatic pavement distress classification facilitates improving the efficiency of pavement maintenance and reducing the cost of labor and resources. A recently influential branch of this task divides the pavement image into patches and addresses these issues from the perspective of multi-instance learning. However, these methods neglect the correlation between patches and suffer from a low efficiency in the model optimization and inference. Meanwhile, Swin Transformer is able to address both of these issues with its unique strengths. Built upon Swin Transformer, we present a vision Transformer named \textbf{P}avement \textbf{I}mage \textbf{C}lassification \textbf{T}ransformer (\textbf{PicT}) for pavement distress classification. In order to better exploit the discriminative information of pavement images at the patch level, the \textit{Patch Labeling Teacher} is proposed to leverage a teacher model to dynamically generate pseudo labels of patches from image labels during each iteration, and guides the model to learn the discriminative features of patches. The broad classification head of Swin Transformer may dilute the discriminative features of distressed patches in the feature aggregation step due to the small distressed area ratio of the pavement image. To overcome this drawback, we present a \textit{Patch Refiner} to cluster patches into different groups and only select the highest distress-risk group to yield a slim head for the final image classification. We evaluate our method on CQU-BPDD. Extensive results show that \textbf{PicT} outperforms the second-best performed model by a large margin of $+2.4\%$ in P@R on detection task, $+3.9\%$ in $F1$ on recognition task, and 1.8x throughput, while enjoying 7x faster training speed using the same computing resources. Our codes and models have been released on \href{https://github.com/DearCaat/PicT}{https://github.com/DearCaat/PicT}.

Junzhi Ning, Jiashi Lin, Yingying Fang et al.

Multimodal large language models (MLLMs) have advanced clinical tasks for common conditions, but their performance on rare diseases remains largely untested. In rare-disease scenarios, clinicians often lack prior clinical knowledge, forcing them to rely strictly on case-level evidence for clinical judgments. Existing benchmarks predominantly evaluate common-condition, single-image settings, leaving multimodal and multi-image evidence integration under rare-disease data scarcity systematically unevaluated. We introduce MMRareBench, to our knowledge the first rare-disease benchmark jointly evaluating multimodal and multi-image clinical capability across four workflow-aligned tracks: diagnosis, treatment planning, cross-image evidence alignment, and examination suggestion. The benchmark comprises 1,756 question-answer pairs with 7,958 associated medical images curated from PMC case reports, with Orphanet-anchored ontology alignment, track-specific leakage control, evidence-grounded annotations, and a two-level evaluation protocol. A systematic evaluation of 23 MLLMs reveals fragmented capability profiles and universally low treatment-planning performance, with medical-domain models trailing general-purpose MLLMs substantially on multi-image tracks despite competitive diagnostic scores. These patterns are consistent with a capacity dilution effect: medical fine-tuning can narrow the diagnostic gap but may erode the compositional multi-image capability that rare-disease evidence integration demands.

Wenhao Tang, Fengtao Zhou, Sheng Huang et al.

Multiple instance learning (MIL) is the most widely used framework in computational pathology, encompassing sub-typing, diagnosis, prognosis, and more. However, the existing MIL paradigm typically requires an offline instance feature extractor, such as a pre-trained ResNet or a foundation model. This approach lacks the capability for feature fine-tuning within the specific downstream tasks, limiting its adaptability and performance. To address this issue, we propose a Re-embedded Regional Transformer (R$^2$T) for re-embedding the instance features online, which captures fine-grained local features and establishes connections across different regions. Unlike existing works that focus on pre-training powerful feature extractor or designing sophisticated instance aggregator, R$^2$T is tailored to re-embed instance features online. It serves as a portable module that can seamlessly integrate into mainstream MIL models. Extensive experimental results on common computational pathology tasks validate that: 1) feature re-embedding improves the performance of MIL models based on ResNet-50 features to the level of foundation model features, and further enhances the performance of foundation model features; 2) the R$^2$T can introduce more significant performance improvements to various MIL models; 3) R$^2$T-MIL, as an R$^2$T-enhanced AB-MIL, outperforms other latest methods by a large margin.The code is available at: https://github.com/DearCaat/RRT-MIL.

Jiayu Chen, Chao Yu, Yuqing Xie et al. · tsinghua

Executing precise and agile flight maneuvers is critical for quadrotors in various applications. Traditional quadrotor control approaches are limited by their reliance on flat trajectories or time-consuming optimization, which restricts their flexibility. Recently, RL-based policy has emerged as a promising alternative due to its ability to directly map observations to actions, reducing the need for detailed system knowledge and actuation constraints. However, a significant challenge remains in bridging the sim-to-real gap, where RL-based policies often experience instability when deployed in real world. In this paper, we investigate key factors for learning robust RL-based control policies that are capable of zero-shot deployment in real-world quadrotors. We identify five critical factors and we develop a PPO-based training framework named SimpleFlight, which integrates these five techniques. We validate the efficacy of SimpleFlight on Crazyflie quadrotor, demonstrating that it achieves more than a 50% reduction in trajectory tracking error compared to state-of-the-art RL baselines. The policy derived by SimpleFlight consistently excels across both smooth polynominal trajectories and challenging infeasible zigzag trajectories on small thrust-to-weight quadrotors. In contrast, baseline methods struggle with high-speed or infeasible trajectories. To support further research and reproducibility, we integrate SimpleFlight into a GPU-based simulator Omnidrones and provide open-source access to the code and model checkpoints. We hope SimpleFlight will offer valuable insights for advancing RL-based quadrotor control. For more details, visit our project website at https://sites.google.com/view/simpleflight/.

Yuxuan Li, Yicheng Zhang, Wenhao Tang et al.

Modern computer vision is converging on a closed loop in which perception, reasoning and generation mutually reinforce each other. However, this loop remains incomplete: the top-down influence of high-level reasoning on the foundational learning of low-level perceptual features is not yet underexplored. This paper addresses this gap by proposing a new paradigm for pretraining foundation models in downstream domains. We introduce Visual insTruction Pretraining (ViTP), a novel approach that directly leverages reasoning to enhance perception. ViTP embeds a Vision Transformer (ViT) backbone within a Vision-Language Model and pretrains it end-to-end using a rich corpus of visual instruction data curated from target downstream domains. ViTP is powered by our proposed Visual Robustness Learning (VRL), which compels the ViT to learn robust and domain-relevant features from a sparse set of visual tokens. Extensive experiments on 16 challenging remote sensing and medical imaging benchmarks demonstrate that ViTP establishes new state-of-the-art performance across a diverse range of downstream tasks. The code is available at https://github.com/zcablii/ViTP.

Wenhao Tang, Rong Qin, Heng Fang et al.

Pre-trained encoders for offline feature extraction followed by multiple instance learning (MIL) aggregators have become the dominant paradigm in computational pathology (CPath), benefiting cancer diagnosis and prognosis. However, performance limitations arise from the absence of encoder fine-tuning for downstream tasks and disjoint optimization with MIL. While slide-level supervised end-to-end (E2E) learning is an intuitive solution to this issue, it faces challenges such as high computational demands and suboptimal results. These limitations motivate us to revisit E2E learning. We argue that prior work neglects inherent E2E optimization challenges, leading to performance disparities compared to traditional two-stage methods. In this paper, we pioneer the elucidation of optimization challenge caused by sparse-attention MIL and propose a novel MIL called ABMILX. It mitigates this problem through global correlation-based attention refinement and multi-head mechanisms. With the efficient multi-scale random patch sampling strategy, an E2E trained ResNet with ABMILX surpasses SOTA foundation models under the two-stage paradigm across multiple challenging benchmarks, while remaining computationally efficient (<10 RTX3090 hours). We show the potential of E2E learning in CPath and calls for greater research focus in this area. The code is https://github.com/DearCaat/E2E-WSI-ABMILX.

Tao He, Sheng Huang, Wenhao Tang et al.

Scene text detection is a challenging computer vision task due to the high variation in text shapes and ratios. In this work, we propose a scene text detector named Deformable Kernel Expansion (DKE), which incorporates the merits of both segmentation and contour-based detectors. DKE employs a segmentation module to segment the shrunken text region as the text kernel, then expands the text kernel contour to obtain text boundary by regressing the vertex-wise offsets. Generating the text kernel by segmentation enables DKE to inherit the arbitrary-shaped text region modeling capability of segmentation-based detectors. Regressing the kernel contour with some sampled vertices enables DKE to avoid the complicated pixel-level post-processing and better learn contour deformation as the contour-based detectors. Moreover, we propose an Optimal Bipartite Graph Matching Loss (OBGML) that measures the matching error between the predicted contour and the ground truth, which efficiently minimizes the global contour matching distance. Extensive experiments on CTW1500, Total-Text, MSRA-TD500, and ICDAR2015 demonstrate that DKE achieves a good tradeoff between accuracy and efficiency in scene text detection.

Yuqing Xie, Jiayu Chen, Wenhao Tang et al.

Designing effective reward functions remains a central challenge in reinforcement learning, especially in multi-objective environments. In this work, we propose Multi-Objective Reward Shaping with Exploration (MORSE), a general framework that automatically combines multiple human-designed heuristic rewards into a unified reward function. MORSE formulates the shaping process as a bi-level optimization problem: the inner loop trains a policy to maximize the current shaped reward, while the outer loop updates the reward function to optimize task performance. To encourage exploration in the reward space and avoid suboptimal local minima, MORSE introduces stochasticity into the shaping process, injecting noise guided by task performance and the prediction error of a fixed, randomly initialized neural network. Experimental results in MuJoCo and Isaac Sim environments show that MORSE effectively balances multiple objectives across various robotic tasks, achieving task performance comparable to those obtained with manually tuned reward functions.

Wenhao Tang, Sheng Huang, Heng Fang et al.

Digitizing pathological images into gigapixel Whole Slide Images (WSIs) has opened new avenues for Computational Pathology (CPath). As positive tissue comprises only a small fraction of gigapixel WSIs, existing Multiple Instance Learning (MIL) methods typically focus on identifying salient instances via attention mechanisms. However, this leads to a bias towards easy-to-classify instances while neglecting challenging ones. Recent studies have shown that hard examples are crucial for accurately modeling discriminative boundaries. Applying such an idea at the instance level, we elaborate a novel MIL framework with masked hard instance mining (MHIM-MIL), which utilizes a Siamese structure with a consistency constraint to explore the hard instances. Using a class-aware instance probability, MHIM-MIL employs a momentum teacher to mask salient instances and implicitly mine hard instances for training the student model. To obtain diverse, non-redundant hard instances, we adopt large-scale random masking while utilizing a global recycle network to mitigate the risk of losing key features. Furthermore, the student updates the teacher using an exponential moving average, which identifies new hard instances for subsequent training iterations and stabilizes optimization. Experimental results on cancer diagnosis, subtyping, survival analysis tasks, and 12 benchmarks demonstrate that MHIM-MIL outperforms the latest methods in both performance and efficiency. The code is available at: https://github.com/DearCaat/MHIM-MIL.

Wenhao Tang, Heng Fang, Ge Wu et al.

Computational pathology (CPath) digitizes pathology slides into whole slide images (WSIs), enabling analysis for critical healthcare tasks such as cancer diagnosis and prognosis. However, WSIs possess extremely long sequence lengths (up to 200K), significant length variations (from 200 to 200K), and limited supervision. These extreme variations in sequence length lead to high data heterogeneity and redundancy. Conventional methods often compromise on training efficiency and optimization to preserve such heterogeneity under limited supervision. To comprehensively address these challenges, we propose a pack-based MIL framework. It packs multiple sampled, variable-length feature sequences into fixed-length ones, enabling batched training while preserving data heterogeneity. Moreover, we introduce a residual branch that composes discarded features from multiple slides into a hyperslide which is trained with tailored labels. It offers multi-slide supervision while mitigating feature loss from sampling. Meanwhile, an attention-driven downsampler is introduced to compress features in both branches to reduce redundancy. By alleviating these challenges, our approach achieves an accuracy improvement of up to 8% while using only 12% of the training time in the PANDA(UNI). Extensive experiments demonstrate that focusing data challenges in CPath holds significant potential in the era of foundation models. The code is https://github.com/FangHeng/PackMIL

Sheng Huang, Wenhao Tang, Guixin Huang et al.

Automatic image-based pavement distress detection and recognition are vital for pavement maintenance and management. However, existing deep learning-based methods largely omit the specific characteristics of pavement images, such as high image resolution and low distress area ratio, and are not end-to-end trainable. In this paper, we present a series of simple yet effective end-to-end deep learning approaches named Weakly Supervised Patch Label Inference Networks (WSPLIN) for efficiently addressing these tasks under various application settings. WSPLIN transforms the fully supervised pavement image classification problem into a weakly supervised pavement patch classification problem for solutions. Specifically, WSPLIN first divides the pavement image under different scales into patches with different collection strategies and then employs a Patch Label Inference Network (PLIN) to infer the labels of these patches to fully exploit the resolution and scale information. Notably, we design a patch label sparsity constraint based on the prior knowledge of distress distribution and leverage the Comprehensive Decision Network (CDN) to guide the training of PLIN in a weakly supervised way. Therefore, the patch labels produced by PLIN provide interpretable intermediate information, such as the rough location and the type of distress. We evaluate our method on a large-scale bituminous pavement distress dataset named CQU-BPDD and the augmented Crack500 (Crack500-PDD) dataset, which is a newly constructed pavement distress detection dataset augmented from the Crack500. Extensive results demonstrate the superiority of our method over baselines in both performance and efficiency. The source codes of WSPLIN are released on https://github.com/DearCaat/wsplin.

Wenhao Tang, Sheng Huang, Qiming Zhao et al.

We present a novel deep learning framework named the Iteratively Optimized Patch Label Inference Network (IOPLIN) for automatically detecting various pavement distresses that are not solely limited to specific ones, such as cracks and potholes. IOPLIN can be iteratively trained with only the image label via the Expectation-Maximization Inspired Patch Label Distillation (EMIPLD) strategy, and accomplish this task well by inferring the labels of patches from the pavement images. IOPLIN enjoys many desirable properties over the state-of-the-art single branch CNN models such as GoogLeNet and EfficientNet. It is able to handle images in different resolutions, and sufficiently utilize image information particularly for the high-resolution ones, since IOPLIN extracts the visual features from unrevised image patches instead of the resized entire image. Moreover, it can roughly localize the pavement distress without using any prior localization information in the training phase. In order to better evaluate the effectiveness of our method in practice, we construct a large-scale Bituminous Pavement Disease Detection dataset named CQU-BPDD consisting of 60,059 high-resolution pavement images, which are acquired from different areas at different times. Extensive results on this dataset demonstrate the superiority of IOPLIN over the state-of-the-art image classification approaches in automatic pavement distress detection. The source codes of IOPLIN are released on \url{https://github.com/DearCaat/ioplin}, and the CQU-BPDD dataset is able to be accessed on \url{https://dearcaat.github.io/CQU-BPDD/}.

Ming Hu, Chenglong Ma, Wei Li et al. · pku

Scientific Large Language Models (Sci-LLMs) are transforming how knowledge is represented, integrated, and applied in scientific research, yet their progress is shaped by the complex nature of scientific data. This survey presents a comprehensive, data-centric synthesis that reframes the development of Sci-LLMs as a co-evolution between models and their underlying data substrate. We formulate a unified taxonomy of scientific data and a hierarchical model of scientific knowledge, emphasizing the multimodal, cross-scale, and domain-specific challenges that differentiate scientific corpora from general natural language processing datasets. We systematically review recent Sci-LLMs, from general-purpose foundations to specialized models across diverse scientific disciplines, alongside an extensive analysis of over 270 pre-/post-training datasets, showing why Sci-LLMs pose distinct demands -- heterogeneous, multi-scale, uncertainty-laden corpora that require representations preserving domain invariance and enabling cross-modal reasoning. On evaluation, we examine over 190 benchmark datasets and trace a shift from static exams toward process- and discovery-oriented assessments with advanced evaluation protocols. These data-centric analyses highlight persistent issues in scientific data development and discuss emerging solutions involving semi-automated annotation pipelines and expert validation. Finally, we outline a paradigm shift toward closed-loop systems where autonomous agents based on Sci-LLMs actively experiment, validate, and contribute to a living, evolving knowledge base. Collectively, this work provides a roadmap for building trustworthy, continually evolving artificial intelligence (AI) systems that function as a true partner in accelerating scientific discovery.

Zongyuan Chen, Yan Xia, Jiayuan Liu et al. · tsinghua

Soft robots exhibit inherent compliance and safety, which makes them particularly suitable for applications requiring direct physical interaction with humans, such as surgical procedures. However, their nonlinear and hysteretic behavior, resulting from the properties of soft materials, presents substantial challenges for accurate modeling and control. In this study, we present a soft robotic system designed for surgical applications and propose a hysteresis-aware whole-body neural network model that accurately captures and predicts the soft robot's whole-body motion, including its hysteretic behavior. Building upon the high-precision dynamic model, we construct a highly parallel simulation environment for soft robot control and apply an on-policy reinforcement learning algorithm to efficiently train whole-body motion control strategies. Based on the trained control policy, we developed a soft robotic system for surgical applications and validated it through phantom-based laser ablation experiments in a physical environment. The results demonstrate that the hysteresis-aware modeling reduces the Mean Squared Error (MSE) by 84.95 percent compared to traditional modeling methods. The deployed control algorithm achieved a trajectory tracking error ranging from 0.126 to 0.250 mm on the real soft robot, highlighting its precision in real-world conditions. The proposed method showed strong performance in phantom-based surgical experiments and demonstrates its potential for complex scenarios, including future real-world clinical applications.

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.

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.

Le Qiu, Zelai Xu, Qixin Tan et al. · tsinghua

Assessing the safety of autonomous driving policy is of great importance, and reinforcement learning (RL) has emerged as a powerful method for discovering critical vulnerabilities in driving policies. However, existing RL-based approaches often struggle to identify vulnerabilities that are both effective-meaning the autonomous vehicle is genuinely responsible for the accidents-and diverse-meaning they span various failure types. To address these challenges, we propose AED, a framework that uses large language models (LLMs) to automatically discover effective and diverse vulnerabilities in autonomous driving policies. We first utilize an LLM to automatically design reward functions for RL training. Then we let the LLM consider a diverse set of accident types and train adversarial policies for different accident types in parallel. Finally, we use preference-based learning to filter ineffective accidents and enhance the effectiveness of each vulnerability. Experiments across multiple simulated traffic scenarios and tested policies show that AED uncovers a broader range of vulnerabilities and achieves higher attack success rates compared with expert-designed rewards, thereby reducing the need for manual reward engineering and improving the diversity and effectiveness of vulnerability discovery.