These 182 ICRA 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 ICRA 2025 paper, free trial on arXivSub.
$U^2$ Frame: A Unified and Unsupervised Learning Framework for LiDAR-Based Loop Closing
Yixin Zhang, Yulan Guo
CodePose EstimationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes U^2 Frame, a unified LiDAR loop closure framework that can simultaneously perform loop closure detection and relative pose estimation without any ground truth training data.
A Comprehensive LLM-powered Framework for Driving Intelligence Evaluation
Shanhe You, Jiangtao Gong
CodeAutonomous DrivingTransformerLarge Language ModelText
🎯 What it does: Proposed and validated an LLM-based intelligent driving behavior assessment framework, constructed a natural language assessment dataset, and verified its effectiveness through CARLA simulation and human evaluation.
🎯 What it does: Proposes a local trajectory planning and speed prediction method based on VPMPCC, learning optimal parameters through Bayesian optimization and a specially designed racing objective function OFR.
🎯 What it does: Proposes a modular indoor social navigation framework that combines classical global planning with deep reinforcement learning (DRL) local planning, enabling rapid goal-reaching while avoiding collisions with people in crowded and confined environments.
A Light-Weight Framework for Open-Set Object Detection with Decoupled Feature Alignment in Joint Space
Yonghao He, Song Liu
CodeObject DetectionConvolutional Neural NetworkTransformerVision Language ModelMultimodality
🎯 What it does: Proposed a lightweight open-set object detection framework called DOSOD, combining YOLO-World with a visual language model and an MLP adapter to align cross-modal features in a joint space, enhancing real-time performance.
A Skeleton-Based Topological Planner for Exploration in Complex Unknown Environments
Haochen Niu, Peilin Liu
CodeOptimizationRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Propose an autonomous exploration framework based on skeleton topological graphs, utilizing global topological information to improve exploration efficiency and reduce computational costs
🎯 What it does: Proposes a collaborative learning framework for shape estimation and shape-aware whole-body control strategies for tendon-driven continuous robots;
Active Semantic Mapping with Mobile Manipulator in Horticultural Environments
José Cuarán, Girish Chowdhary
CodeRobotic IntelligenceSimultaneous Localization and MappingImagePoint CloudAgriculture Related
🎯 What it does: Proposes an efficient and scalable active semantic mapping method for horticultural environments, utilizing a mobile robot gripper and RGB-D camera. The method employs probabilistic semantic maps to detect semantic targets, generate candidate viewpoints, and calculate information gain, while introducing an efficient ray casting strategy and a novel information utility function.
An Active Perception Game for Robust Information Gathering
Siming He, Pratik Chaudhari
CodeOptimizationRobotic Intelligence
🎯 What it does: Proposes a game-theoretic active sensing framework that estimates the difference between predicted information gain and actual information gain, achieving sublinear regret through online estimation to reduce the suboptimality of active sensing systems.
🎯 What it does: Demonstrate integrating the annealed Winner-Takes-All (aWTA) loss into state-of-the-art motion prediction models to minimize the number of hypotheses and improve performance.
Asymptotically-Optimal Multi-Query Path Planning for a Polygonal Robot
Duo Zhang, Jingjin Yu
CodeOptimizationRobotic Intelligence
🎯 What it does: For multi-query 2D environment polyhedral global robot path planning, the rotational stacked visibility graph (RVG) algorithm is proposed, achieving fast computation of near-optimal paths while supporting simultaneous translation and rotation.
Asynchronous Multi-Object Tracking with an Event Camera
Angus Apps, Robert Mahony
CodeObject TrackingOptical FlowVideo
🎯 What it does: Proposed the AEMOT algorithm, which achieves the detection and tracking of multiple objects by asynchronously processing raw events using an event camera.
Autonomous Navigation in Ice-Covered Waters with Learned Predictions on Ship-Ice Interactions
Ninghan Zhong, Stephen L. Smith
CodeAutonomous DrivingRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Proposed a deep learning-based ice movement prediction model integrated into a graph search planner, achieving real-time autonomous navigation in icy areas, with verification showing a significant reduction in collisions with ice floes.
CAFE-AD: Cross-Scenario Adaptive Feature Enhancement for Trajectory Planning in Autonomous Driving
Junrui Zhang, Yanyong Zhang
CodeAutonomous DrivingPoint Cloud
🎯 What it does: Propose the CAFE-AD method, which enhances feature representation in trajectory planning through adaptive feature pruning and cross-scenario feature interpolation, addressing causal confusion and long-tail scenario distribution issues caused by open training.
CAO-RONet: A Robust 4D Radar Odometry with Exploring More Information from Low-Quality Points
Zhiheng Li, Zheng Fang
CodePose EstimationAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a learning-based 4D mmWave radar odometry framework, including local completion, context-aware hierarchical association, and window-based optimizer, to achieve robust ego-motion estimation under low-quality radar points.
CDMFusion: RGB-T Image Fusion Based on Conditional Diffusion Models via Few Denoising Steps in Open Environments
Luojie Yang, Yufeng Yue
CodeRestorationDiffusion modelMultimodality
🎯 What it does: Proposes CDMFusion, a three-branch conditional diffusion model for RGB-T image fusion, which adaptively enhances multimodal features through a dynamic gating module and accelerates the generation process using a skip patrol mechanism.
CGTrack: Cascade Gating Network with Hierarchical Feature Aggregation for UAV Tracking
Weihong Li, Libo Zhang
CodeObject Tracking
🎯 What it does: Proposes CGTrack, a UAV tracker based on Hierarchical Feature Cascade (HFC) and Lightweight Gated Center Head (LGCH), which expands network capacity through a coarse-to-fine two-tier framework.
Chameleon: Fast-Slow Neuro-Symbolic Lane Topology Extraction
Zongzheng Zhang, Hao Zhao
CodeAutonomous DrivingComputational EfficiencyVision Language ModelImageChain-of-Thought
🎯 What it does: Proposed a fast-slow neuro-symbolic lane topology extraction algorithm called Chameleon, which alternates between using a fast system for direct inference of detection instances and a slow system that leverages VLM chain-of-thought to handle edge cases
🎯 What it does: Proposed the Chemistry3D robot interaction toolkit, enabling robots to perform chemical experiments and real-time visualization of temperature, color, and pH changes.
CloudTrack: Scalable UAV Tracking with Cloud Semantics
Yannik Blei, Wolfram Burgard
CodeObject TrackingVision Language ModelImageText
🎯 What it does: Proposes a scalable drone-based object tracking method leveraging cloud semantics, enabling open-vocabulary tracking based on verbal descriptions (e.g., shirt color).
CoDynTrust: Robust Asynchronous Collaborative Perception via Dynamic Feature Trust Modulus
Yunjiang Xu, Jianping Wang
CodeObject DetectionAutonomous Driving
🎯 What it does: Proposes the CoDynTrust framework, leveraging dynamic feature trust modulus (DFTM) and a multi-scale fusion module to achieve robust fusion for asynchronous collaborative perception.
🎯 What it does: Built a cognitive robot operating system called CognitiveOS that can run on multiple robot platforms, utilizing a multimodal large model multi-agent system, modular and configurable, supporting complex tasks.
COHERENT: Collaboration of Heterogeneous Multi-Robot System with Large Language Models
Kehui Liu, Xuelong Li
CodeRobotic IntelligenceLarge Language ModelTextBenchmark
🎯 What it does: Proposes the COHERENT framework, which utilizes large language models (LLMs) for task planning in heterogeneous multi-robot collaboration, and designs a Proposal-Execution-Feedback-Adjustment (PEFA) loop mechanism.
Context Graph-Based Visual-Language Place Recognition
Soojin Woo, Seong-Woo Kim
CodeRetrievalRepresentation LearningConvolutional Neural NetworkVision Language ModelMultimodality
🎯 What it does: Proposed a visual-language place recognition method based on a zero-shot language-driven semantic segmentation model, leveraging pixel-level embeddings to construct semantic image descriptors;
Dashing for the Golden Snitch: Multi-Drone Time-Optimal Motion Planning with Multi-Agent Reinforcement Learning
Xian Wang, Shuo Li
CodeOptimizationReinforcement Learning
🎯 What it does: Proposed a decentralized multi-agent reinforcement learning strategy network for time-optimal flight in multi-quadcopters, incorporating a soft collision avoidance mechanism.
🎯 What it does: Proposed the Deep Height Decoupling (DHD) framework, which predicts height maps and uses Mask Guided Height Sampling (MGHS) to separate image features by height ranges, followed by a Synergistic Feature Aggregation (SFA) module to enhance feature representation, thereby achieving more accurate vision-based 3D occupancy prediction.
🎯 What it does: Proposes the DENSER framework based on 3D Gaussian splatting for dynamic urban environment reconstruction, significantly enhancing the appearance and shape modeling of foreground dynamic objects.
Distributed Invariant Kalman Filter for Object-Level Multi-Robot Pose SLAM
Haoying Li, Junfeng Wu
CodeRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Propose a distributed invariant Kalman filter based on covariance intersection (CI) for multi-robot pose estimation, and adopt an object-level measurement model to reduce communication burden.
🎯 What it does: Propose two cross-modal knowledge distillation methods (UDAKD and FSKD), leveraging synchronized camera and LiDAR data to achieve 3D point cloud semantic segmentation without 3D annotations.
🎯 What it does: Propose the Doppler Former (DPF) module to efficiently extract velocity information from raw radar data, and propose a Fully Complex Convolutional Network (FCCN) backbone network that is more suitable for raw data; integrate DPF into FCCN to enhance the performance of downstream radar perception tasks.
DP-Habitat: Bridging the Gap Between Simulation and Reality for Visual Navigation in Dynamic Pedestrian Environments
Liang Qin, Houqiang Li
CodeAutonomous DrivingBenchmark
🎯 What it does: Developed the DP-Habitat dynamic pedestrian simulator on the Habitat platform, and proposed the Adaptive Object Navigation with Dynamic Mapping (AON-DM) baseline method specifically designed for dynamic pedestrian environments.
Dual-AEB: Synergizing Rule-Based and Multimodal Large Language Models for Effective Emergency Braking
Wei Zhang, Yilun Chen
CodeAutonomous DrivingTransformerLarge Language ModelMultimodality
🎯 What it does: Proposed the Dual-AEB system, which integrates advanced multimodal large language models (MLLM) with traditional rule-based fast AEB to enhance adaptability in open scenarios.
Dynamic End Effector Trajectory Tracking for Small-Scale Underwater Vehicle-Manipulator Systems (UVMS): Modeling, Control, and Experimental Validation
Niklas Trekel, R. Seifried
CodeRobotic Intelligence
🎯 What it does: Implement dynamic end-effector trajectory tracking on a small underwater vehicle-manipulator system (UVMS), and complete modeling, control, and experimental validation using a task-priority control approach.
🎯 What it does: Propose a data and model-agnostic, efficient 7-DoF grasping method that generates grasping configurations for arbitrary target objects using single-view point cloud data, and rapidly adjusts the grasping configurations through object detection and multi-region point cloud distribution perception, enabling real-time precise grasping by robots in dense cluttered environments; meanwhile, design a grasping framework to reduce the time consumed during grasping and improve the grasping efficiency for specified target objects.
Ego-$A^{\mathbf{3}}$: Adaptive Fusion-Based Disentangled Transformer for Egocentric Action Anticipation
Minhyuk Kim, S. Yoo
CodeRecognitionTransformerVideo
🎯 What it does: Proposed the Ego-A³ model, which improves action prediction from the wearable camera perspective using an adaptive fusion and separation Transformer architecture;
🎯 What it does: Propose a training objective that simultaneously minimizes prediction loss and mutual information under adversarial perturbations, and gradually introduces adversarial objectives through a curriculum learning guide to enhance the robustness of 3D vision.
CodeAutonomous DrivingSimultaneous Localization and MappingWorld ModelPoint Cloud
🎯 What it does: We propose the ELite framework, which achieves seamless alignment of multi-session LiDAR data, dynamic object removal, and end-to-end map updating.
ETSM: Automating Dissection Trajectory Suggestion and Confidence Map-Based Safety Margin Prediction for Robot-Assisted Endoscopic Submucosal Dissection
Mengya Xu, Hongliang Ren
CodeRobotic IntelligenceVideoBiomedical Data
🎯 What it does: Constructed the ETSM dataset and proposed RCMNet, which automatically suggests resection trajectories and predicts safety margins based on confidence maps.
🎯 What it does: This paper proposes a visual localization method based on an event camera, which utilizes existing LiDAR maps for pose refinement. The process includes first projecting LiDAR point clouds into 2D to obtain a depth map with a rough initial pose, then aligning events with the 2D depth map using an optical flow estimation network, followed by estimating the camera pose via a PnP solver. Additionally, the authors introduce a novel frame-based event representation to enhance structural clarity and design an auxiliary variable prediction module as a regularization term to mitigate the impact of biases in real poses on network convergence.
Expert-Enhanced Masked Point Modeling for Point Cloud Self-Supervised Learning
YuJun Liu, Shu-Tao Xia
CodeRepresentation LearningMixture of ExpertsPoint Cloud
🎯 What it does: Introduce an expert-enhanced masked point modeling method in self-supervised learning for point clouds, achieving routing and analysis of different semantics by inserting a Sparse Mixture of Experts (SMoE) layer after each backbone block.
🎯 What it does: Propose FedDet, a method for detecting data poisoning attacks in federated skeleton action recognition, and develop a prototype-based attack detector.
🎯 What it does: This paper proposes the 3F-OVD task, extending supervised fine-grained object detection to open-vocabulary scenarios, creating the NEU-171K fine-grained dataset, benchmarking existing state-of-the-art object detectors on this dataset, and proposing a simple and effective post-processing technique.
FogROS2-PLR: Probabilistic Latency-Reliability for Cloud Robotics
Kai-Peng Chen, Kenneth Y. Goldberg
CodeOptimizationRobotic Intelligence
🎯 What it does: This paper proposes the FogROS2-PLR framework, which utilizes multi-path network interfaces and cloud server redundancy to enhance the latency reliability of cloud robotics systems.
🎯 What it does: Construct a dynamic driving scene NeRF using only continuous RGB images without pose input, employing point-level dynamic-static decoupling and optical flow constraints.
CodePose EstimationAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a gravity-enhanced trajectory estimation method called GaRLIO, which integrates radar, LiDAR, and inertial measurement unit (IMU) data. The method leverages direct velocity information from radar to improve gravity estimation accuracy and reduces vertical drift by eliminating dynamic objects using radar data.
GenCo: A Dual VLM Generate-Correct Framework for Adaptive Peg-in-Hole Robotics
Zhengxue Zhou, A. I. Cooper
CodeRobotic IntelligenceSupervised Fine-TuningMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Implemented a Generate-Correct framework based on dual Vision-Language Models (VLMs) for adaptive blind peg-in-hole robotic tasks, integrating a motion generator and a motion expert to generate and correct actions during execution.
Gradient-Based Trajectory Optimization with Parallelized Differentiable Traffic Simulation
Sanghyun Son, Ming C. Lin
CodeAutonomous DrivingOptimizationWorld ModelVideoTime Series
🎯 What it does: Proposed a parallel differentiable traffic simulator based on the Intelligent Driver Model (IDM), utilizing the simulator to achieve trajectory denoising, dense trajectory reconstruction, and future trajectory prediction, and optimizing IDM parameters through gradient methods.
Ground-Optimized 4D Radar-Inertial Odometry Via Continuous Velocity Integration Using Gaussian Process
Wooseong Yang, Ayoung Kim
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint CloudTime Series
🎯 What it does: Proposes a ground-optimized noise filtering and continuous velocity pre-integration method based on radar to achieve more accurate radar-inertial odometry
HARP: Human-Assisted Regrouping With Permutation Invariant Critic for Multi-Agent Reinforcement Learning
Huawen Hu, Shu Zhang
CodeReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Proposes the HARP framework, integrating multi-agent automatic reorganization and deployment with human assistance, enabling non-experts to provide effective guidance with minimal intervention; during the training phase, agents dynamically adjust groupings, and during deployment, they proactively seek human assistance, utilizing permutation-invariant group evaluators to assess and refine human-proposed groupings.
🎯 What it does: Developed HeLiOS, a deep network for heterogeneous LiDAR, used for LiDAR location recognition, leveraging small local windows and spherical transformers along with clustering assignment based on optimal transport to generate robust global descriptors, and improving performance through overlapping foundation data mining and guided triplet loss.
🎯 What it does: Propose a lightweight, plug-and-play HGSLoc framework that combines 3D reconstruction with heuristic optimization to improve the accuracy of camera pose estimation.
Hier-SLAM: Scaling-Up Semantics in SLAM with a Hierarchically Categorical Gaussian Splatting
Boying Li, Hamid Rezatofighi
CodeLarge Language ModelGaussian SplattingSimultaneous Localization and Mapping
🎯 What it does: Propose Hier-SLAM, a semantic SLAM method that utilizes hierarchical category 3D Gaussian splatting to achieve accurate global semantic mapping, scalability, and explicit semantic label prediction.
Illumination Adaptation for SAM to Achieve Accurate Segmentation of Images Taken in Low-Light Scenes
Hongmin Mu, Zhengcai Cao
CodeSegmentationDomain AdaptationTransformerImage
🎯 What it does: An adaptive method for the Segment Anything Model (SAM) is proposed to address low-light scenarios, incorporating self-training, low-light feature enhancement head, and domain shift compensation loss to achieve more accurate image segmentation.
🎯 What it does: Designed and implemented the IMOST framework, leveraging incremental dynamic memory and online self-supervised annotation to achieve continuous traversability learning.
Implicit Articulated Robot Morphology Modeling with Configuration Space Neural Signed Distance Functions
Yiting Chen, A. Billard
CodeComputational EfficiencyRobotic Intelligence
🎯 What it does: Proposes an implicit robot morphology modeling method based on the configuration space signature distance function (Robot Neural Distance Function, RNDF), utilizing forward kinematics to achieve precise encoding and optimize the computational efficiency and accuracy of distance queries.
🎯 What it does: This paper proposes a 3D object detection method for a single source domain, aiming to enhance the model's generalization capability in target domains with different sensor configurations and scene distributions.
Improving Indoor Localization Accuracy by Using an Efficient Implicit Neural Map Representation
Haofei Kuang, C. Stachniss
CodePose EstimationRobotic IntelligenceSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Propose an implicit neural map representation to capture position and orientation geometric features from 2D LiDAR scans, and combine a lightweight neural network with a traditional Monte Carlo localization framework to design an efficient observation model, achieving real-time robot pose estimation.
Indoor Localization of UAVs Using Only Few Measurements by Output-Sensitive Preimage Intersection
Michael M. Bilevich, Dan Halperin
CodeRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Propose a deterministic indoor UAV positioning method using only a small number of downward distance measurements and corresponding odometry, achieved through preimage intersection and spatial subdivision search.
🎯 What it does: Proposed InsCMPR, an instance-aware cross-modal pose recognition method, which generates descriptors through pixel-level and instance-level modality alignment as well as a dual-branch hybrid Mamba-Transformer network;
IRef-VLA: A Benchmark for Interactive Referential Grounding with Imperfect Language in 3D Scenes
Haochen Zhang, Wenshan Wang
CodeGraph Neural NetworkVision Language ModelPoint CloudGraphBenchmark
🎯 What it does: Constructed and released the IRef-VLA dataset for interactive referential localization tasks, containing natural language instructions and navigation goals in 3D scenes.
🎯 What it does: Proposed a anchor-based collaborative perception framework named ACCO for target detection in multi-agent systems; the framework includes anchor feature block (AFB), anchor confidence generator (ACG), and local-global fusion modules (LAAF and SACA)
JORD: A Benchmark Dataset for Off-Road LiDAR Place Recognition and SLAM
Wei Zhou, Gang Wang
CodeSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: Proposed and released the first benchmark dataset JORD specifically designed for airborne LiDAR SLAM, and conducted benchmark testing with multiple advanced methods.
KARMA: Augmenting Embodied AI Agents with Long-and-Short Term Memory Systems
Zixuan Wang, Yiming Gan
CodeRobotic IntelligenceLarge Language ModelPrompt EngineeringRetrieval-Augmented Generation
🎯 What it does: Proposed the KARMA memory system, integrating long-term and short-term memory modules to enhance the planning capabilities of embodied AI agents when performing complex household tasks.
🎯 What it does: Developed an open-source C++ library called KISS-Matcher for point cloud registration, integrating a new feature detector Faster-PFH and a graph theory pruning algorithm based on k-core, forming a complete and user-friendly registration pipeline.
CodeObject TrackingVision Language ModelVideoBenchmark
🎯 What it does: Proposed a language-oriented unified task framework for multi-object tracking, created the LaMOT benchmark dataset, and introduced a concise and effective tracker called LaMOTer.
🎯 What it does: This study addresses the problem of pedestrian 3D tracking in highly crowded urban environments by constructing an offline automatic annotation system and proposing a high-resolution, density-aware, and relation-aware representation learning method;
Learning Multiple Probabilistic Decisions from Latent World Model in Autonomous Driving
Lingyu Xiao, Jingdong Wang
CodeAutonomous DrivingWorld ModelBenchmark
🎯 What it does: Propose the LatentDriver framework, which uses an autoregressive world model to infer a mixed distribution of the next environmental state and possible vehicle actions through multiple probabilistic hypotheses, and generates deterministic control signals from it
🎯 What it does: Learning control for quadrotors using differentiable simulation, comparing with model-agnostic RL, and finding that differentiable simulation significantly outperforms traditional RL in sample efficiency and training time, capable of achieving recovery within seconds using vehicle states and accomplishing control within minutes relying solely on visual features.
LidarDM: Generative LiDAR Simulation in a Generated World
Vlas Zyrianov, Shenlong Wang
CodeData SynthesisAutonomous DrivingDiffusion modelWorld ModelPoint CloudTime Series
🎯 What it does: Proposes LidarDM, a generative model capable of generating realistic, layout-aware, physically feasible, and temporally consistent LiDAR videos and 4D point cloud sequences; the model can guide LiDAR generation based on driving scenarios.
LiDARDustX: A LiDAR Dataset for Dusty Unstructured Road Environments
Chenfeng Wei, Shenhong Wang
CodeAutonomous DrivingPoint CloudBenchmark
🎯 What it does: This paper proposes the LiDARDustX dataset, which collects 30,000 frames of LiDAR point clouds in dusty environments, equipped with 3D bounding box annotations and semantic segmentation from six types of sensors. Based on this dataset, a benchmark experiment for 3D detection and segmentation is established, further analyzing the impact of dust on perception accuracy.
LiLoc: Lifelong Localization Using Adaptive Submap Joining and Egocentric Factor Graph
Yixin Fang, Gim Hee Lee
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a lifelong LiDAR-based localization framework named LiLoc, which utilizes adaptive subgraph connections between central sessions and sub-sessions, coarse-to-fine pose initialization, and joint optimization with self-centered factor graphs, supporting switching between relocalization and incremental localization modes.
Logic-RAG: Augmenting Large Multimodal Models with Visual-Spatial Knowledge for Road Scene Understanding
Imran Kabir, S. Billah
CodeAutonomous DrivingLarge Language ModelVision Language ModelVideoRetrieval-Augmented Generation
🎯 What it does: Propose the Logic-RAG framework, applying visual-spatial knowledge retrieval-augmented generation (RAG) technology to large-scale multimodal models to enhance road scene understanding.
🎯 What it does: Proposed the LoGS visual localization pipeline, utilizing 3D Gaussian Splatting as the scene representation to achieve high-quality view synthesis and accomplish localization.
🎯 What it does: Study the performance of differentiable driving policies under distribution shift, propose the LoRD low-rank residual decoder and multi-task fine-tuning, and improve model performance in closed-loop evaluation.
Maintaining Strong $r$-Robustness in Reconfigurable Multi-Robot Networks Using Control Barrier Functions
Haejoon Lee, Dimitra Panagou
CodeRobotic Intelligence
🎯 What it does: Propose a control barrier function to ensure sufficient strong $r$-robustness in multi-robot networks during reconfiguration, enabling leader-follower consensus.
MambaGlue: Fast and Robust Local Feature Matching with Mamba
Kihwan Ryoo, Hyun Myung
CodeImage
🎯 What it does: Proposed a Mamba-based local feature matching method called MambaGlue, and designed two modules: MambaAttention mixer and depth confidence score regressor.
🎯 What it does: Propose MDC-Seg, which utilizes multi-directional convolution (MDConv) to perform parallel sparse feature encoding on bird's eye view (BEV) and range view (RV) planes, combined with attention mechanisms, effective multi-feature fusion (EMFF) module, and point voxel constraint (PVC) module, achieving effective receptive field expansion and accuracy improvement for 3D point cloud semantic segmentation.
ME-PATS: Mutually Enhancing Search-Based Planner and Learning-Based Agent for Tractor-Trailer Systems
Ke Fan, Zufeng Zhang
CodeAutonomous Driving
🎯 What it does: Proposed the ME-PATS framework, aiming to plan dynamically feasible paths for tractor-trailer systems through mutually enhancing search planners and learning agents;
Monocular Depth Estimation and Segmentation for Transparent Object with Iterative Semantic and Geometric Fusion
Jiangyuan Liu, Wei Zou
CodeSegmentationDepth EstimationImage
🎯 What it does: Proposed a monocular framework that simultaneously performs segmentation and depth estimation of transparent objects using a single RGB image.