These 239 IROS 2025 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every IROS 2025 paper, free trial on arXivSub.
360Recon: An Accurate Reconstruction Method based on Depth Fusion from 360 Images
3D Hierarchical Panoptic Segmentation in Real Orchard Environments Across Different Sensors
Matteo Sodano (University of Bonn), C. Stachniss (University of Bonn)
CodeSegmentationPoint CloudAgriculture Related
🎯 What it does: Proposed a hierarchical panoramic segmentation method for 3D data in apple orchards, which can simultaneously perform semantic segmentation, stem and fruit instance segmentation, as well as tree (stem + fruit) instance segmentation; the method can identify individual plants, fruits, and stems, and calculate the number of fruits per tree.
4D-ROLLS: 4D Radar Occupancy Learning via LiDAR Supervision
Ruihan Liu, Yunjiang Lou
CodeAutonomous DrivingPoint Cloud
🎯 What it does: Designed and implemented a 4D radar occupancy estimation method based on LiDAR point cloud supervision, proposing pseudo LiDAR label generation and multi-stage supervision, leveraging LiDAR occupancy maps for model alignment and fine-tuning.
A Coarse-to-Fine Approach to Multi-Modality 3D Occupancy Grounding
Zhan Shi, Jianke Zhu
CodeDepth EstimationAutonomous DrivingVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a 3D occupancy ground benchmark and an end-to-end multi-modal model called GroundingOcc for precisely locating and predicting occupied voxels in outdoor scenes based on natural language descriptions.
🎯 What it does: Built and released a multimodal Dex Hand Dataset integrating visual, point cloud, and kinematic data, aligning human and robotic hand poses in a shared canonical space using neural rendering and kinematic optimization to achieve fine-grained hand imitation learning.
A Spatiotemporal Downwash Modeling for Agile Close-Proximity Multirotor Flight
Pavel Kharitenko, Yang Wang
CodeRobotic IntelligenceTime SeriesPhysics Related
🎯 What it does: Studied the downwash interaction of a quadrotor in the high-speed range of 0.5-4.0 m/s through high-precision CFD simulations, and proposed a geometric deep neural network prediction model that simultaneously considers both absolute and relative velocities.
A System for Multi-View Mapping of Dynamic Scenes Using Time-Synchronized UAVs
Aniket Gupta, Hanumant Singh
CodeRobotic IntelligenceSimultaneous Localization and MappingImageBenchmark
🎯 What it does: Developed a multi-perspective synchronous acquisition system based on drones, and provided a corresponding unrestricted outdoor scene synchronous multi-perspective image dataset.
ACP-MVS: Efficient Multi-View Stereo with Attention-based Context Perception
Hao Jia, Xin Yang
CodeDepth EstimationImage
🎯 What it does: Proposed ACP-MVS, an efficient multi-view stereo network that improves pixel correspondence accuracy and reduces noise by constructing a context-aware cost volume.
🎯 What it does: Propose an adaptive pipeline suitable for large-scale outdoor traffic scenarios, constructing a high-precision 3D Surfel model and real-time synthesizing realistic novel view images.
Adaptive Sliding Window Optimization for Multi-Modal LiDAR Inertial Odometry and Mapping
G. Han, Yu Hu
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed an uncertainty-based adaptive sliding window (ASW) strategy and implemented a multi-modal LiDAR-inertial odometry and mapping framework that integrates mechanical and solid-state LiDAR.
🎯 What it does: Propose a supervised surround view depth estimation framework based on Transformer, named AVT-SSDepth, which can jointly predict depth maps from multiple cameras;
AirSwarm: Enabling Cost-Effective Multi-UAV Research with COTS drones
Xiaowei Li, Lihua Xie
CodeRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Developed the AirSwarm platform, enabling multi-drone collaboration using low-cost commercial drones for research and educational purposes.
AKF-LIO: LiDAR-Inertial Odometry with Gaussian Map by Adaptive Kalman Filter
Xupeng Xie, Boyu Zhou
CodeAutonomous DrivingSimultaneous Localization and MappingMultimodalityPoint CloudBenchmark
🎯 What it does: Propose an adaptive Kalman filter framework that dynamically estimates time-varying noise covariance of LiDAR and IMU measurements, uses Gaussian maps to represent environmental planarity and spatial noise, and achieves accurate plane normal estimation through pseudo-merging via correlation registration, enhancing localization robustness in LiDAR failure or dynamic scenes.
🎯 What it does: Proposed a defensive network ADNet for reliable segmentation in maritime environments under various adversarial attacks, thereby enhancing the perception reliability of unmanned surface vessels.
Antagonistic Physical-Virtual Framework for the Development of Soft Actuators
Diogo Fonseca, P. Neto
CodeRobotic Intelligence
🎯 What it does: A framework for developing and integrating soft actuator models that includes high-fidelity digital twins and mechanical integration platforms is proposed, enabling testing and validation in both real and virtual environments.
AnyTSR: Any-Scale Thermal Super-Resolution for UAV
Mengyuan Li, Liangliang Yao
CodeSuper ResolutionImage
🎯 What it does: Proposed a UAV-specific arbitrary-scale thermal imaging super-resolution method called AnyTSR, which achieves image reconstruction at arbitrary scales using a single model.
ATARS: An Aerial Traffic Atomic Activity Recognition and Temporal Segmentation Dataset
Zihao Chen, Yan-Tsung Peng
CodeRecognitionSegmentationVideoBenchmark
🎯 What it does: Proposed and constructed the ATARS dataset, defined the multi-label temporal atomic activity recognition task, and conducted experimental evaluation on the performance of existing SOTA models in atomic activity recognition and temporal segmentation.
Autonomous Subtask Generation for Indoor Search and Rescue Mission via Large-Language-Model and Behavior-Tree Integration
Junfeng Shi, Hui Zhang
CodeRobotic IntelligenceLarge Language Model
🎯 What it does: AutoExpand proposes a high-level framework that tightly couples a large language model (LLM) with behavior trees to automatically generate responsive, context-aware subtasks for indoor search and rescue missions;
Bag-of-Word-Groups (BoWG): A Robust and Efficient Loop Closure Detection Method Under Perceptual Aliasing
Xiang Fei, Lu Li
CodeComputational EfficiencySimultaneous Localization and MappingImage
🎯 What it does: Proposed and implemented a loop closure detection method called Bag-of-Word-Groups (BoWG), aiming to improve precision-recall rate, robustness, and computational efficiency in perceptually confusing environments.
BaTCAVe: Trustworthy Explanations for Robot Behaviors
Som Sagar, Ransalu Senanayake
CodeExplainability and InterpretabilityRobotic Intelligence
🎯 What it does: Propose a trustworthy explanation technique based on human-interpretable high-level concepts for explaining neural networks in robot decision-making.
Bench4Merge: A Comprehensive Benchmark for Merging in Realistic Dense Traffic with Micro-Interactive Vehicles
Zhengming Wang, Yilun Chen
CodeAutonomous DrivingLarge Language ModelBenchmark
🎯 What it does: Proposed a closed-loop evaluation benchmark for assessing the motion planning capability of merging in high-density traffic, using micro-interaction vehicles trained on a large-scale dataset to enhance scenario diversity, and reconstructing the evaluation mechanism through large language models (LLM).
BEV-LIO(LC): BEV Image Assisted LiDAR-Inertial Odometry with Loop Closure
Haoxin Cai, Jianqi Liu
CodeAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposes a BEV-LIO(LC) framework that combines bird's-eye view (BEV) image representations with LiDAR-Inertial Odometry, achieving loop closure through BEV image features.
🎯 What it does: Proposed the BEVDiffLoc framework, treating LiDAR localization as a problem of conditional pose generation, and achieving end-to-end localization through BEV perspective, data augmentation, maximum feature aggregation module, Vision Transformer, and diffusion model;
Bio-Inspired Hybrid Map: Spatial Implicit Local Frames and Topological Map for Mobile Cobot Navigation
T. Dang, M. Huber
CodeRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Propose a hybrid map construction and navigation method based on human perceptual approaches, utilizing spatial implicit local frames and topological maps for mobile collaborative robot navigation.
BookBot: A Robotic Manipulation Benchmark for Voice-Driven Book Recognition and Grasping in Cluttered Environments
Huaqiang Wang, Shen Wang
CodeRobotic IntelligenceLarge Language ModelImageTextBenchmark
🎯 What it does: Constructed the THU-Book dataset and developed the Voice-driven BookBot robotic system for book recognition, localization, and grasping.
BoRe-Depth: Self-Supervised Monocular Depth Estimation with Boundary Refinement for Embedded Systems
Chang Liu, Xu Zhang
CodeDepth EstimationImage
🎯 What it does: Proposed the BoRe-Depth lightweight monocular depth estimation model, achieving efficient depth image estimation and boundary refinement on embedded systems.
Bridging Text and Vision: A Multi-View Text-Vision Registration Approach for Cross-Modal Place Recognition
Tianyi Shang, Weijun Hu
CodeRecognitionRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposed a multi-view text-visual registration method called Text4VPR for cross-modal place recognition, achieving the first-ever matching between text descriptions and image databases using only text.
🎯 What it does: Enhance keypoint detection and description through the interaction between global Transformer information and keypoints with descriptors.
🎯 What it does: Proposed a real-time error detection framework based on contrastive learning, classifying calibration states into calibrated or misaligned through binary classification.
🎯 What it does: Proposed a LiDAR-camera online calibration network called CalibMutiL based on multi-level visual feature fusion, capable of end-to-end alignment of point clouds and RGB images
Category-Level 6D Object Pose Estimation in Agricultural Settings Using a Lattice-Deformation Framework and Diffusion-Augmented Synthetic Data
Marios Glytsos, Petros Maragos
CodeData SynthesisPose EstimationDiffusion modelImageAgriculture Related
🎯 What it does: Proposed the PLANTPose framework, which can perform category-level 6D pose estimation using only RGB images, and adapts to the shapes of different instances by predicting deformation parameters;
Category-level Meta-learned NeRF Priors for Efficient Object Mapping
Saad Ejaz, J. L. Sánchez-López
CodePose EstimationOptimizationMeta LearningNeural Radiance Field
🎯 What it does: Propose PRENOM, a prior-based efficient neural object mapper that combines category-level priors with object-level NeRF to achieve efficient reconstruction and canonical pose estimation.
🎯 What it does: Propose Causal-Planner, which uses attention adversarial graph learning to decouple causal and confounding factors in the scene interaction graph, and introduces the Long Short-Term Memory Gating Module (LSTEM) to enhance the capture of causal relationships in dynamic environments.
🎯 What it does: This paper proposes the CLAIM method, which utilizes a powerful monocular depth model to achieve data alignment between the camera and LiDAR.
🎯 What it does: Proposed the CoCMT framework, which utilizes object queries to select and transmit key features, and constructed the Efficient Query Transformer (EQFormer) to efficiently fuse multi-agent object queries, combined with deep supervision to enhance overall performance.
ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments
Rohit Dwivedula, Chris Rossbach
CodeOptimizationRobotic Intelligence
🎯 What it does: Propose ConfigBot, which dynamically reconfigures robot applications to meet predefined performance specifications through runtime analysis and automatic configuration tuning, and experimentally validates its stability and resource optimization effects on multiple real robots.
Convex Hull-based Algebraic Constraint for Visual Quadric SLAM
Xiaolong Yu, T. Feng
CodeSegmentationPose EstimationOptimizationSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose and apply convex hull-based algebraic constraints to object reconstruction, front-end pose estimation, and back-end bundle adjustment for quadratic curves.
ConvoyLLM: Dynamic Multi-Lane Convoy Control Using LLMs
Liping Lu, Pan Zhou
CodeAutonomous DrivingOptimizationTransformerLarge Language Model
🎯 What it does: Proposed a dynamic multi-lane platoon control method based on large language models (LLMs), utilizing knowledge-driven real-time adaptive decision-making to achieve tasks such as obstacle avoidance, platoon joining/leaving, and formation switching, while ensuring the stability and flexibility of the overall platoon structure through an interleaved formation control strategy.
🎯 What it does: Propose a self-supervised learning framework, CooPre, which leverages unlabeled 3D V2X data to aggregate multi-agent perception information and enhances collaborative perception performance through a point cloud reconstruction proxy task.
CRESSim–MPM: A Material Point Method Library for Surgical Soft Body Simulation with Cutting and Suturing
Yafei Ou, Mahdi Tavakoli
CodePhysics Related
🎯 What it does: Designed and implemented CRESSim-MPM, a GPU-accelerated MPM library for surgical soft tissue physics simulation, supporting cutting and suturing.
🎯 What it does: Propose a cross-layer fusion paradigm to achieve bidirectional information flow between object lists and raw visual features in 3D object detection and tracking, integrated into an end-to-end Transformer framework.
🎯 What it does: Proposed the CM-SSM framework based on cross-modal state space modeling for real-time RGB-thermal pixel semantic segmentation in outdoor scenes.
🎯 What it does: Proposed a cross-modal visual-lidar localization method based on BEV feature distillation, achieving end-to-end cross-modal localization between surrounding images and point clouds for the first time.
CrowdQuery: Density-Guided Query Module for Enhanced 2D and 3D Detection in Crowded Scenes
Marius Dähling, J. M. Zöllner
CodeObject DetectionTransformerImagePoint Cloud
🎯 What it does: Proposes CrowdQuery, a density-guided query module designed to enhance 2D and 3D detectors, improving detection performance in crowded scenarios.
🎯 What it does: Built a reconstruction and synthesis framework for V2X driving environments, utilizing decomposed Gaussian splatting to achieve scene reconstruction and flexible editing, and rendering images from both vehicle and infrastructure perspectives for large-scale data augmentation.
CSVO: Complementary-Pathway Spatial-Enhanced Visual Odometry for Extreme Environments with Brain-Inspired Vision Sensors
Yihan Lin, Rong Zhao
CodePose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingImageVideoMultimodality
🎯 What it does: Proposed and implemented a CSVO visual odometry method based on the brain-inspired visual sensor Tianmouc, enhancing pose estimation in extreme environments by fusing information from the cognitive channel (COP) and action channel (AOP).
🎯 What it does: Designed and implemented a low-light clothing grasping detection model called DarkSeg, utilizing a student-teacher model for feature alignment, learning illumination-invariant structural representations from an infrared teacher model, and proposing a depth-aware grasping strategy while constructing the DarkClothes dataset;
🎯 What it does: Collected driver gaze data using in-vehicle cameras and constructed the DashGaze large-scale synchronized perspective dataset along with the baseline model DashGazeNet.
Decentralized Uncertainty-Aware Multi-Agent Collision Avoidance With Model Predictive Path Integral*
S. Dergachev, Konstantin S. Yakovlev
CodeAutonomous DrivingOptimizationSafty and Privacy
🎯 What it does: Proposed a decentralized multi-agent collision avoidance method that integrates Model Predictive Path Integral (MPPI) with probability-adapted Optimal Reciprocal Collision Avoidance, and directly embeds probabilistic safety constraints into the MPPI sampling process through second-order cone programming (SOCP);
Delving into Mapping Uncertainty for Mapless Trajectory Prediction
Zongzheng Zhang, Hao Zhao
CodeAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Analyze the impact of map uncertainty in mapless trajectory prediction, and propose Proprioceptive Scenario Gating and Covariance Map Uncertainty methods to adaptively integrate map uncertainty into trajectory prediction models.
DHC-ME: A Decentralized Hybrid Cooperative Approach for Multi-Robot Autonomous Exploration
Wenhao Jia, Liang Li
CodeRobotic IntelligenceAgentic AISimultaneous Localization and Mapping
🎯 What it does: Proposed and implemented DHC-ME: a decentralized hybrid collaboration strategy for multi-robot range-aware exploration, enhancing team coordination and exploration efficiency.
Directed Spatial Consistency-Based Partial-to-Partial Point Cloud Registration with Deep Graph Matching
Jingwen Zhou, Zhe Min
CodePose EstimationGraph Neural NetworkPoint CloudBiomedical Data
🎯 What it does: Propose a partial-to-partial point cloud registration framework based on directional space consistency, which first extracts overlapping regions and obtains a hard matching matrix via graph matching, then generates translation-invariant edge vectors through sampling nodes, combines point-level and edge-level geometric constraints for dual optimization, and introduces a bidirectional registration mechanism to enhance registration stability.
DogLegs: Robust Proprioceptive State Estimation for Legged Robots Using Multiple Leg-Mounted IMUs
Yibin Wu, H. Kuhlmann
CodeRobotic Intelligence
🎯 What it does: Propose a state estimation system named DogLegs, which achieves robust body pose estimation for legged robots by fusing body IMU, joint encoders, and multi-leg IMUs.
Domain-Conditioned Scene Graphs for State-Grounded Task Planning
Jonas Herzog, Yue Wang
CodeRobotic IntelligenceVision Language ModelMultimodality
🎯 What it does: This paper proposes and implements a domain-conditioned scene graph-based state induction framework for perception state induction and subsequent state-based planning in robot task planning; meanwhile, it provides a lightweight vision-language implementation scheme that classifies domain-specific predicates based on domain-related object detection to generate scene graphs; through this framework, scene graphs can be directly mapped to symbolic states in planning languages (e.g., PDDL).
🎯 What it does: Propose the DPR-Splat framework, which utilizes a sparse-view 3D Gaussian projection model and improves the accuracy of camera poses and depth maps through pose and depth refinement modules, thereby achieving higher quality view synthesis.
DriveLMM-o1: A Step-by-Step Reasoning Dataset and Large Multimodal Model for Driving Scenario Understanding
Ayesha Ishaq, Salman H. Khan
CodeAutonomous DrivingLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed a step-by-step visual reasoning dataset called DriveLMM-o1 for autonomous driving scenarios, and fine-tuned a large multimodal model on this dataset to provide a step-by-step reasoning process and answers.
DroneKey: Drone 3D Pose Estimation in Image Sequences using Gated Key-representation and Pose-adaptive Learning
Seo-Bin Hwang, Yeong-Jun Cho
CodePose EstimationTransformerVideo
🎯 What it does: Propose the DroneKey framework, combining 2D keypoint detection with 3D pose estimation to address the challenge of drone keypoint detection.
DRP: A Decomposition-Reflection-Prediction Framework for Long-Horizon Robot Task Planning using Large Language Models
Zhaowen Zheng, Jing Wang
CodeRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes a robot task planning framework DRP based on large language models (LLMs), supporting environment knowledge injection to address feasibility issues in long-term tasks.
DynamicGSG: Dynamic 3D Gaussian Scene Graphs for Environment Adaptation
Luzhou Ge, Xuesong Li
CodeDomain AdaptationGraph Neural NetworkVision Language ModelGaussian Splatting
🎯 What it does: Propose the DynamicGSG system, which utilizes high-precision Gaussian light scattering technology to construct a dynamically updatable 3D Gaussian scene graph for environmental adaptation.
🎯 What it does: Proposed a small object detection framework for aerial images based on encoder-decoder architecture, diffusion models, and Swin Transformer, redefining the SOD task as a Noise-to-Box process;
🎯 What it does: Proposed a lightweight real-time underwater object detection model named EFCWM-Mamba-YOLO, and created the UOD-SZTU-2025 dataset containing 3,133 high-quality images.
🎯 What it does: Proposed a low-resolution Transformer tracker called LoReTrack, which enhances tracking performance at low resolutions through dual knowledge distillation.
🎯 What it does: Proposed an instance motion-aware network called IMPNet for predicting future 3D point cloud scenes through historical LiDAR scans, explicitly integrating motion and instance-level information to enhance prediction accuracy.
🎯 What it does: Propose a fast and effective multi-modal 3D object detector that combines the Instance-level Contrastive Distillation (ICD) framework and the Cross Linear Attention Fusion Module (CLFM).
EfficientEQA: An Efficient Approach to Open-Vocabulary Embodied Question Answering
Kai Cheng, Aniket Bera
CodeComputational EfficiencyRobotic IntelligenceVision Language ModelVision-Language-Action ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the EfficientEQA framework to address robot-assisted open-vocabulary embodied question answering tasks, integrating efficient exploration with free-text answer generation.
🎯 What it does: Proposed the EFFOcc framework, which includes an efficient fused OccNet and a multi-stage occupancy-guided distillation, for training lightweight occupancy networks under scenarios with extremely few labels.
🎯 What it does: Proposed ELPTNet, a LiDAR-based efficient 3D pedestrian tracking network for autonomous navigation social robots, achieving real-time accurate tracking.
EndoMUST: Monocular Depth Estimation for Robotic Endoscopy via End-to-end Multi-step Self-supervised Training
Liangjing Shao, Xinrong Chen
CodeDepth EstimationRobotic IntelligenceSupervised Fine-TuningOptical FlowImageBiomedical Data
🎯 What it does: Propose a multi-step efficient fine-tuning framework for end-to-end training in robotic endoscopy monocular depth estimation, where the training process is divided into three steps: optical flow registration, multi-scale image decomposition, and multi-transformation alignment.
🎯 What it does: Proposed ET-Former, which achieves 3D semantic scene completion using a monocular camera and generates semantic occupancy maps and uncertainty estimates.
ET-Plan-Bench: Embodied Task-level Planning Benchmark Towards Spatial-Temporal Cognition with Foundation Models
Lingfeng Zhang, Jianye Hao
CodeLarge Language ModelBenchmark
🎯 What it does: Proposed a new embodied task planning benchmark called ET-Plan-Bench to evaluate LLMs' understanding of spatial, temporal, and causal relationships
EventSync: Joint Recovery of Temporal Offsets and Relative Orientations for Wide-Baseline Event Cameras
Wanli Xing, Jia Pan
CodePose EstimationOptical Flow
🎯 What it does: Developed a software algorithm called EventSync for synchronizing multi-camera event streams and estimating the relative pose between cameras.
Experimental Open-Source Framework for Underwater Pick-and-Place Studies with Lightweight UVMS – An Extensive Quantitative Analysis
Nathalie Bauschmann, Daniel A. Duecker
CodeOptimizationRobotic IntelligenceBenchmark
🎯 What it does: This paper proposes a complete open-source software framework for fully automated grasping and releasing experiments with a lightweight underwater manipulator (UVMS), extending the previous task-priority control framework by incorporating high-level decision-making and grasping detection methods, and verifying them on BlueROV2 and Alpha5 Manipulator.
Exponentially Weighted Instance-Aware Repeat Factor Sampling for Long-Tailed Object Detection Model Training in Unmanned Aerial Vehicles Surveillance Scenarios
🎯 What it does: Developed an exponentially weighted instance-aware resampling method (E-IRFS) to address long-tailed class imbalance in UAV monitoring scenarios.
Fast Policy: Accelerating Visuomotor Policies without Re-training
Tong Wu, Zheng Wang
CodeComputational EfficiencyRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningDiffusion model
🎯 What it does: Propose Fast Policy (FP), which accelerates visual motion policies by reusing UNet encoder features in non-critical denoising steps and dynamically selecting key steps.
FGS-SLAM: Fourier-based Gaussian Splatting for Real-time SLAM with Sparse and Dense Map Fusion
Yansong Xu, Weijia Zhou
CodePose EstimationAutonomous DrivingGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Proposes an adaptive densification method based on Fourier frequency domain analysis to achieve fast convergence of Gaussian representations, and constructs independent and unified sparse and dense maps. The sparse map is used for efficient tracking, while the dense map is used for high-fidelity visual representation.
FlowMP: Learning Motion Fields for Robot Planning with Conditional Flow Matching
K. Nguyen, Minh Nhat Vu
CodeRobotic IntelligenceFlow-based Model
🎯 What it does: By extending the flow matching method to capture second-order trajectory dynamics, learning a motion field incorporating acceleration information, directly mapping a simple prior distribution to smooth, executable robot trajectories.
🎯 What it does: Propose a new feature alignment strategy that combines the Perspective-Driven Attention Fusion (PDAF) module with the Projection-Stable Patch-Guided Pose Optimizer (PSPG) for cross-perspective (ground-to-satellite) geolocation, addressing occlusion and distortion caused by perspective changes.
🎯 What it does: Developed an FPGA-based neural controller for high-speed control in robotic systems such as cartpole and F1TENTH racing cars, demonstrating its ability to achieve kilohertz-level control rates.
From Learning to Mastery: Achieving Safe and Efficient Real-World Autonomous Driving with Human-in-the-Loop Reinforcement Learning
Zeqiao Li, Z. Zuo
CodeAutonomous DrivingReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Proposed a reward-agnostic active human-in-the-loop learning method called H-DSAC to achieve safe and efficient real-world autonomous driving;
Gaze-Guided Task Decomposition for Imitation Learning in Robotic Manipulation
Ryo Takizawa, Y. Kuniyoshi
CodeRobotic Intelligence
🎯 What it does: Decompose object manipulation tasks demonstrated by humans into subtasks using gaze transitions, enabling skill reuse and combination in imitation learning.
GazeTarget360: Towards Gaze Target Estimation in 360-Degree for Robot Perception
Zhu Dai, Chen Li
CodeRobotic IntelligenceImage
🎯 What it does: Proposed a system called GazeTarget360 for estimating human gaze targets from panoramic scenes in images, integrating an eye contact detector, a pre-trained visual encoder, and a multi-scale fusion decoder.
🎯 What it does: Proposes a visual drone navigation framework GRaD-Nav that combines 3D Gaussian point mist technology with differentiable depth reinforcement learning, and verifies its training efficiency and zero-shot sim-to-real transfer capability in hardware experiments.
🎯 What it does: Proposed a multimodal localization network called GSPR based on 3D Gaussian mapping, which fuses multi-view RGB images with LiDAR point clouds into a unified spatiotemporal scene representation and extracts global descriptors through 3D graph convolution and Transformer.
🎯 What it does: Proposed a high-speed and high-accuracy LiDAR semantic segmentation network called HARP-NeXt, along with an efficient preprocessing method and a multi-scale range-point fusion backbone.
HFDNet: High-Frequency Divergence Attention Network for Underwater Segmentation
Hongbo Xie, Chunlei Wang
CodeSegmentationTransformerImage
🎯 What it does: Proposed a high-frequency difference attention network (HFDNet) based on Transformer, improving underwater image semantic segmentation through frequency domain analysis
CodeAutonomous DrivingComputational EfficiencySupervised Fine-TuningPrompt EngineeringVision Language ModelImageChain-of-Thought
🎯 What it does: Proposed a hierarchical question-answering (QA) method for autonomous driving scene understanding, efficiently capturing key visual elements by refining high-level and detailed sub-questions.
High-Precision and High-Efficiency Trajectory Tracking for Excavators Based on Closed-Loop Dynamics
Ziqing Zou, Yue Wang
CodeOptimizationRobotic IntelligenceWorld Model
🎯 What it does: Proposed and verified a trajectory tracking method called EfficientTrack, combining model-based learning with closed-loop dynamics to address the nonlinear dynamics of excavators and improve tracking accuracy and efficiency.