These 156 IROS 2024 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 2024 paper, free trial on arXivSub.
A Framework for Reproducible Benchmarking and Performance Diagnosis of SLAM Systems
Nikola Radulov, Mikel Luján
CodeSimultaneous Localization and MappingMultimodalityBenchmark
🎯 What it does: Proposes SLAMFuse, an open-source benchmark framework for multi-modal SLAM algorithms, providing a consistent cross-platform environment, data fuzz testing, failure detection, and diagnostic tools;
A Generic Trajectory Planning Method for Constrained All-Wheel-Steering Robots
Ren Xin, Ming Liu
CodeOptimizationRobotic Intelligence
🎯 What it does: Proposes a general trajectory planning method for four-wheel steering (AWS) robots with fixed steering axes and limited steering angles for each wheel, called C-AWS.
A Hybrid Model and Learning-Based Force Estimation Framework for Surgical Robots
Hao Yang, Jie Ying Wu
CodeRobotic Intelligence
🎯 What it does: Proposed a hybrid framework combining model-based and learning methods for surgical robotic force estimation, specifically targeting the Patient Side Manipulators of the da Vinci research platform.
Accelerating Model Predictive Control for Legged Robots through Distributed Optimization
Lorenzo Amatucci, Claudio Semini
CodeOptimizationRobotic Intelligence
🎯 What it does: By decomposing robot dynamics into parallel subsystems and utilizing ADMM to achieve consistency between subsystems, the performance of model predictive control for legged robots is improved.
🎯 What it does: Proposes an adaptive stochastic nonlinear model predictive control method based on deep reinforcement learning for autonomous vehicle motion control
Agile and Safe Trajectory Planning for Quadruped Navigation with Motion Anisotropy Awareness
Wentao Zhang, Lijun Zhu
CodeOptimizationRobotic Intelligence
🎯 What it does: Proposes a navigation framework for quadruped robots that considers the anisotropy of their motion, encompassing kinodynamic trajectory generation, nonlinear trajectory optimization, and nonlinear model predictive control.
🎯 What it does: Propose a method for depth estimation under different lighting conditions during the day and night, utilizing multi-sensor fusion—projecting synchronous sparse laser point clouds onto the image plane to generate sparse depth maps, along with camera images as input;
🎯 What it does: Propose an asymmetric fair fusion model ASY-VRNet based on vision and 4D millimeter-wave radar for waterway panoramic driving perception;
Asynchronous Microphone Array Calibration using Hybrid TDOA Information
Chengjie Zhang, He Kong
CodeOptimizationSimultaneous Localization and MappingAudio
🎯 What it does: Propose a batch SLAM method using hybrid TDOA (TDOA-S and TDOA-M) combined with kinematic information for asynchronous microphone array calibration.
🎯 What it does: Propose an unsupervised LiDAR 3D scanning instance segmentation method that first generates initial instance masks through pseudo annotations and then refines the segmentation results using a self-training algorithm.
Belief-Aided Navigation using Bayesian Reinforcement Learning for Avoiding Humans in Blind Spots
Jinyeob Kim, Donghan Kim
CodeAutonomous DrivingReinforcement Learning
🎯 What it does: Proposed and verified a BNBRL+ algorithm based on Bayesian reinforcement learning for avoiding humans in blind spots and achieving socially aware navigation.
🎯 What it does: Propose a monocular camera-based BEV-enhanced visual localization framework that generates global descriptors combining visual and structural information;
Blending Distributed NeRFs with Tri-stage Robust Pose Optimization
Baijun Ye, Guyue Zhou
CodePose EstimationOptimizationNeural Radiance Field
🎯 What it does: Proposed a distributed NeRF system that employs a three-stage pose optimization (bundle adjust Mip-NeRF 360, Frame2Model optimization, and Model2Model optimization) to achieve precise image poses and NeRF transformations, and complete NeRF fusion;
CaFNet: A Confidence-Driven Framework for Radar Camera Depth Estimation
Huawei Sun, Robert Wille
CodeDepth EstimationImageMultimodalityPoint Cloud
🎯 What it does: Proposes a two-stage end-to-end trained confidence-driven radar-camera fusion network called CaFNet for dense depth estimation, fusing RGB images with sparse noisy radar point clouds.
🎯 What it does: Proposes a real-time UAV target tracking framework named CLAT (Convolutional Local Attention Tracker), combining a hierarchical convolutional local attention structure, a streamlined feature fusion network, and a redesigned upper controller to enhance tracking speed and control robustness.
Co-RaL: Complementary Radar-Leg Odometry with 4-DoF Optimization and Rolling Contact
Sangwoo Jung, Ayoung Kim
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingMultimodality
🎯 What it does: Proposes a cooperative radar-leg odometry algorithm integrating chip radar with leg-based systems to achieve robust and accurate localization
Communication-Constrained Multi-Robot Exploration with Intermittent Rendezvous
Alysson Ribeiro Da Silva, Ani Hsieh
CodeOptimizationRobotic Intelligence
🎯 What it does: Propose modeling the multi-robot exploration task as a Distributed Partially Observable Markov Decision Process (DEC-POMDP), and design a joint strategy following a rendezvous plan, enabling robots to occasionally share maps in unknown environments without being constrained by joint path optimization.
Conditional Generative Denoiser for Nighttime UAV Tracking
Yucheng Wang, Haobo Zuo
CodeObject TrackingTransformerVideo
🎯 What it does: Proposed a conditional generative denoiser (CG-Denoiser) for nighttime UAV target tracking, breaking the deterministic paradigm by conditionally generating and removing noise on the input.
Current-Based Impedance Control for Interacting with Mobile Manipulators
Jelmer de Wolde, Javier Alonso-Mora
CodeRobotic Intelligence
🎯 What it does: Proposed a current-based impedance control method to achieve compliance control on mobile robots without force/torque sensors, and verified its performance on the Kinova GEN3 Lite robotic arm.
🎯 What it does: Proposed a stepwise alignment paradigm named DaDiff to align low-resolution target features in nighttime UAV tracking to daytime features, and constructed the NUT-LR nighttime UAV tracking benchmark dataset.
🎯 What it does: Proposes a diffusion-based affordance prediction (DAP) process to address multi-modal object storage problems, first locating placeable regions and then precisely calculating relative poses.
DDS-SLAM: Dense Semantic Neural SLAM for Deformable Endoscopic Scenes
Jiwei Shan, Hesheng Wang
CodeSimultaneous Localization and MappingBiomedical Data
🎯 What it does: Propose DDS-SLAM for achieving accurate camera tracking, continuous dense reconstruction, and high-quality image rendering in deformable endoscopic scenarios.
DeepBHMR: Learning Bidirectional Hybrid Mixture Models for Generalized Rigid Point Set Registration
Zhe Min, M. Q. Meng
CodePose EstimationPoint CloudBiomedical Data
🎯 What it does: Propose a deep learning-based rigid registration method called DeepBHMR, which achieves precise registration by utilizing normals and a bidirectional hybrid model.
DeepMIF: Deep Monotonic Implicit Fields for Large-Scale LiDAR 3D Mapping
Kutay Yilmaz, A. Artemov
CodeAutonomous DrivingOptimizationNeural Radiance FieldSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: Propose a LiDAR large-scale 3D mapping method based on deep monotonic implicit fields, optimizing non-metric monotonic implicit fields to achieve dense 3D scenes.
DHP-Mapping: A Dense Panoptic Mapping System with Hierarchical World Representation and Label Optimization Techniques
Tianshuai Hu, Ming Liu
CodeSimultaneous Localization and MappingWorld ModelPoint Cloud
🎯 What it does: Proposes DHP-Mapping, a dense panoptic mapping system that employs hierarchical modeling using multiple TSDF submaps and panoptic labels, capable of simultaneously maintaining geometric and semantic information at both voxel-level and submap-level.
🎯 What it does: Built DMFuser, a Transformer-based end-to-end perception-control framework that fuses RGB-D camera data and generates vehicle throttle, steering, and braking commands using multi-task knowledge distillation.
Driving Animatronic Robot Facial Expression From Speech
Boren Li, Hangxin Liu
CodeGenerationRobotic IntelligenceAudio
🎯 What it does: Propose a voice-driven method for generating realistic animated robot facial expressions based on linear blend skinning (LBS), capable of producing expressions in real-time at high frame rates.
Dynamic SpectraFormer for Ultra-High-Definition Underwater Image Enhancement
Zhiqiang Hu, Masatoshi Ishikawa
CodeRestorationTransformerImage
🎯 What it does: Proposed and implemented Dynamic SpectraFormer, which enhances underwater images using a frequency-domain transformer, capable of simultaneously correcting high- and low-frequency mixed distortions.
Efficient Incremental Penetration Depth Estimation between Convex Geometries
Wei Gao
CodeOptimization
🎯 What it does: Proposes an optimization-based incremental penetration depth estimation algorithm capable of calculating the minimum penetration depth and its direction between convex geometries.
🎯 What it does: Proposed DPLNet, which adapts frozen pre-trained RGB models to multi-modal semantic segmentation through dual prompt learning, achieving efficient fusion with only a small number of trainable parameters.
🎯 What it does: Proposed Trajectory Conditional Flow Matching (T-CFM), unifying trajectory prediction and generation tasks, utilizing flow matching techniques to learn time-varying vector fields, achieving efficient and fast trajectory generation and prediction.
🎯 What it does: Proposed the PIP framework for efficiently generating pixel-level aligned data pairs of cross-temporal infrared-RGB images, and constructed the NUDT-PIP dataset;
Enhanced Language-guided Robot Navigation with Panoramic Semantic Depth Perception and Cross-modal Fusion
Liuyi Wang, Qi Chen
CodeRobotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: Proposed the SEAT model, utilizing panoramic multi-type visual encoder, region query pre-training task, and dual-scale cross-modal Transformer to achieve semantic-depth aware cross-modal navigation;
Enhancing Exploratory Capability of Visual Navigation Using Uncertainty of Implicit Scene Representation
Yichen Wang, Hesheng Wang
CodeAutonomous DrivingNeural Radiance FieldImage
🎯 What it does: Proposed an uncertainty-driven exploration navigation pipeline (NUE) based on NeRF, leveraging NeRF's uncertainty to enhance exploration behavior and integrating memory information to generate navigation actions.
Ensembling Prioritized Hybrid Policies for Multi-agent Pathfinding
Huijie Tang, Jinkyoo Park
CodeOptimizationReinforcement LearningMixture of Experts
🎯 What it does: Proposed and implemented the Ensembling Prioritized Hybrid Policies (EPH) method, which enhances performance in multi-agent path planning through a selective communication module, Q-learning training, integration of neural policies with single-intelligent expert guidance, Q-value prioritized conflict resolution, and robust ensemble methods.
🎯 What it does: Propose using an event camera for moving object segmentation, construct the first large-scale dataset DSEC-MOS, and design the EmoFormer network.
🎯 What it does: Proposes a point cloud semantic segmentation method that combines local features from 3D representations with range image-based representations, utilizing GPU-accelerated KDTree for fast construction, querying, and projection.
Exposing the Unseen: Exposure Time Emulation for Offline Benchmarking of Vision Algorithms
Olivier Gamache, Philippe Giguère
CodeAuto EncoderSimultaneous Localization and MappingImagePoint CloudBenchmark
🎯 What it does: Propose an exposure time simulator that generates images with arbitrary exposure times to enable offline benchmark testing of visual algorithms.
FlexLoc: Conditional Neural Networks for Zero-Shot Sensor Perspective Invariance in Object Localization with Distributed Multimodal Sensors
Jason Wu, Mani Srivastava
CodeObject DetectionMultimodality
🎯 What it does: Propose FlexLoc, which utilizes conditional neural networks to generate subsets of model weights at runtime based on node poses, achieving zero-shot sensor-invariant target localization;
🎯 What it does: Proposed an online 3D multi-object tracking framework based on camera-radar fusion, primarily relying on 3D object motion models while fully utilizing image appearance information and 2D detection results.
🎯 What it does: Proposed and implemented Hierarchical Action Chunking Transformer with Vector-quantization (HACT-Vq) for learning multimodal trajectories and fine-grained actions from multi-user demonstration data.
🎯 What it does: Proposed a leaderless hierarchical search-based cooperative motion planning (SCMP) method for unmanned ground vehicles with nonholonomic Ackermann models;
High-Fidelity SLAM Using Gaussian Splatting with Rendering-Guided Densification and Regularized Optimization
Shuo Sun, Martin Magnusson
CodePose EstimationDepth EstimationGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Proposed a dense RGB-D SLAM system based on 3D Gaussian Splatting, achieving accurate pose tracking and visually realistic reconstruction.
Intelligent Fish Detection System with Similarity-Aware Transformer
Shengchen Li, Zhiqiang Xu
CodeObject DetectionTransformerVideoBenchmark
🎯 What it does: Designed a lightweight, plug-and-play edge intelligent visual system and proposed FishViT for rapid fish identification in aquatic-terrestrial transition scenarios.
InverseMatrixVT3D: An Efficient Projection Matrix-Based Approach for 3D Occupancy Prediction
Zhenxing Ming, Stewart Worrall
CodeAutonomous DrivingImagePoint Cloud
🎯 What it does: Propose an InverseMatrixVT3D method that efficiently converts multi-view image features into 3D feature volumes using projection matrices for 3D semantic occupancy prediction.
🎯 What it does: Propose a deep reinforcement learning method called IR2 for information sharing in multi-robot exploration teams under sparse intermittent connectivity conditions;
JointLoc: A Real-time Visual Localization Framework for Planetary UAVs Based on Joint Relative and Absolute Pose Estimation
Xubo Luo, Leizheng Shu
CodePose EstimationSimultaneous Localization and MappingImage
🎯 What it does: Proposed JointLoc, a real-time visual localization framework that fuses absolute 2-DoF pose and relative 6-DoF pose estimation to determine the position of planetary drones.
Large Language Models Powered Context-aware Motion Prediction in Autonomous Driving
Xiaoji Zheng, Jiangtao Gong
CodeAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: Using large language models (LLM) combined with visualized traffic environment (TC-Map) and text prompts to encode traffic semantic context, and integrating this contextual information into motion prediction models to improve prediction accuracy.
LCP-Fusion: A Neural Implicit SLAM with Enhanced Local Constraints and Computable Prior
Jiahui Wang, Yufeng Yue
CodeRobotic IntelligenceSimultaneous Localization and MappingImage
🎯 What it does: Proposes LCP-Fusion, a neural implicit SLAM system that integrates sparse voxel octrees, feature grids, and SDF priors to achieve scalable and robust dense SLAM.
Learning from Spatio-temporal Correlation for Semi-Supervised LiDAR Semantic Segmentation
Seungho Lee, Hyunjung Shim
CodeSegmentationAutonomous DrivingPoint Cloud
🎯 What it does: Proposes a semi-supervised LiDAR semantic segmentation method leveraging spatiotemporal correlation, primarily generating high-quality pseudo-labels through spatial consistency of adjacent scans, and employing a dual-branch structure to alleviate label imbalance issues.
Learning Variable Compliance Control From a Few Demonstrations for Bimanual Robot with Haptic Feedback Teleoperation System
Tatsuya Kamijo, Masashi Hamaya
CodeRobotic IntelligenceTransformer
🎯 What it does: Demonstrating variable compliance control for rigid dual-arm robots using a VR hand controller-based haptic remote control and the Comp-ACT method, enabling improved flexibility and safety through learning from a small number of demonstrations, suitable for complex contact-rich operational tasks.
Leveraging Neural Radiance Field in Descriptor Synthesis for Keypoints Scene Coordinate Regression
Huy-Hoang Bui, Joo-Ho Lee
CodePose EstimationGraph Neural NetworkNeural Radiance Field
🎯 What it does: This paper proposes a pipeline that utilizes NeRF to generate keypoint descriptors in order to improve the localization accuracy of keypoint scene coordinate regression (KSCR).
LF-3PM: a LiDAR-based Framework for Perception-aware Planning with Perturbation-induced Metric
Kaixin Chai, Fei Gao
CodeAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed the LiDAR-based perception-aware planning framework LF-3PM, which improves the localization performance of autonomous robots in sparse environments by evaluating the impact of LiDAR observations on localization accuracy and stability, and accelerating motion planning using a static observation loss map
LiDAR-based 4D Occupancy Completion and Forecasting
Xinhao Liu, Chen Feng
CodeAutonomous DrivingPoint CloudBenchmark
🎯 What it does: Proposed the LiDAR perception task OCF, achieving a unified framework for sparse-to-dense, partial-to-complete, and 3D-to-4D prediction, and evaluated baselines based on the OCFBench dataset
🎯 What it does: Proposes LiOn-XA, an unsupervised domain adaptation method combining LiDAR-Only Cross-Modal learning with adversarial training, for 3D LiDAR point cloud semantic segmentation, aiming to bridge domain gaps caused by environmental and sensor setup changes.
🎯 What it does: By optimizing the extrinsic camera parameters in the 3D Gaussian Splatting framework, achieving fast 3D reconstruction and novel view synthesis using photometric residuals without requiring accurate camera pose initialization.
🎯 What it does: Proposes a method for loss distillation through gradient matching, searching for a weighted Chamfer distance (CD) loss that does not require parameter tuning, and introduces a two-layer optimization training framework based on this loss;
ManipVQA: Injecting Robotic Affordance and Physically Grounded Information into Multi-Modal Large Language Models
Siyuan Huang, Hao Dong
CodeRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: Developed the ManipVQA framework, which injects robotic grasping capabilities and physical knowledge into multimodal large language models through the VQA format, achieving tool detection, graspability recognition, and physical concept understanding.
🎯 What it does: Proposes a segmentation method for real and virtual image regions, utilizing synthetic images, domain-invariant information, motion entropy kernel, and epipolar geometric consistency to address segmentation under the influence of underwater dynamic disturbances;
🎯 What it does: Propose a sparse query framework MDHA, which constructs adaptive 3D proposals by combining hybrid anchors and depth prediction, and introduces Anchor Encoder and Circular Deformable Attention to improve efficiency.
🎯 What it does: Developed a high-fidelity hybrid reality sensor simulation framework for testing and evaluating the resilience of drones against false data injection attacks.
🎯 What it does: Propose the ModaLink framework, which uses the FoV conversion module and non-negative decomposition encoder to unify images and point clouds into position-discriminative features
Model Predictive Trees: Sample-Efficient Receding Horizon Planning with Reusable Tree Search
John Lathrop, Soon-Jo Chung
CodeAutonomous DrivingOptimization
🎯 What it does: Proposed a recursive vision tree search algorithm for Model Predictive Trees (MPT), leveraging the reuse of entire optimal subtrees to enhance planning efficiency and quality.
🎯 What it does: Proposes a general 3D plane detection and reconstruction framework called MonoPlane based on monocular geometric cues. It leverages pre-trained networks to extract depth and normals from single images, progressively fits planes using neighborhood-guided RANSAC, performs multi-plane joint optimization at the image level, and further extends to sparse view reconstruction.
Multi-Robot Active Graph Exploration with Reduced Pose-SLAM Uncertainty via Submodular Optimization
Ruofei Bai, Lihua Xie
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingGraph
🎯 What it does: Propose a two-phase strategy that first generates a fast coverage path and then reduces the uncertainty in multi-robot collaborative SLAM by adding information-rich loop closure actions, modeling loop closure selection as a non-monotonic submodular maximization problem.
🎯 What it does: Proposed a multi-modal evolutionary encoder (MEE) that integrates multiple modalities such as depth and sub-instructions, employing evolutionary pre-training to enhance environmental understanding and generalization in continuous vision-language navigation.
🎯 What it does: Reformulate single-agent and multi-agent trajectory planning problems as query problems for implicit neural trajectory representations, and propose Neural Trajectory Models (NTM) to generate near-optimal trajectories.
NRDF - Neural Region Descriptor Fields as Implicit ROI Representation for Robotic 3D Surface Processing
A. Pratheepkumar, Markus Vincze
CodeRobotic IntelligenceNeural Radiance Field
🎯 What it does: Proposed the Neural Region Descriptor Field (NRDF) to achieve unsupervised dense 3D surface region correspondence, enabling retrieval of arbitrary processing-related regions of interest (P-ROI) in new instances of known categories, and applying it for one-click P-ROI-level process knowledge transfer.
OBHMR: Robust Partial-to-full Generalized Point Set Registration with Overlap-guided Bidirectional Hybrid Mixture Model
Xinzhe Du, Zhe Min
CodePose EstimationOptimizationPoint CloudBiomedical Data
🎯 What it does: Proposed an overlap-guided bidirectional hybrid model point set registration method (OBHMR), achieving robust partial-to-global geometric registration;
🎯 What it does: Learning a system that can both classify the states of predicates in scenes and generate scene configurations that satisfy specified predicates
Optimal Robotic Assembly Sequence Planning (ORASP): A Sequential Decision-Making Approach
Kartik Nagpal, Negar Mehr
CodeRobotic IntelligenceReinforcement Learning
🎯 What it does: Propose a method that models robot assembly planning as a Markov decision process, and utilize dynamic programming, Graph Exploration Assembly Planner (GEAP), ORASP search, and deep reinforcement learning to generate optimal assembly sequences;
OVGNet: A Unified Visual-Linguistic Framework for Open-Vocabulary Robotic Grasping
Meng Li, Chenguang Yang
CodeRobotic IntelligenceVision Language ModelVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposed the OVGNet framework, integrating open-vocabulary learning to enable robots to grasp objects of new categories, and constructed a corresponding large-scale benchmark dataset.
🎯 What it does: Proposed an end-to-end parking network from RGB images to path planning, using imitation learning to learn and execute human driving trajectories
🎯 What it does: Propose a robot grasping agent (PGA) that achieves personalized grasping through a single human-robot interaction and a pseudo-label propagation algorithm utilizing unlabeled image data.
🎯 What it does: Proposed an instance-based transfer imitation learning method for personalized autonomous driving planning in complex urban environments.
🎯 What it does: Proposes a full-process solution based on HD map enhancement and trajectory synthesis, first generating synthetic driving data, then pre-training the trajectory prediction model using a masked autoencoder (MAE) to learn general representations.
Preventing Catastrophic Forgetting in Continuous Online Learning for Autonomous Driving
Rui Yang, Yassine Ruichek
CodeAutonomous DrivingPoint Cloud
🎯 What it does: Proposed an online learning framework called Long-Short-Term Online Learning (LSTOL), aimed at preventing catastrophic forgetting and achieving long-term unsupervised learning in autonomous driving
Progressive Query Refinement Framework for Bird's-Eye-View Semantic Segmentation from Surrounding Images
Dooseop Choi, KyoungWook Min
CodeSegmentationAutonomous DrivingImage
🎯 What it does: Propose an advanced query refinement framework to apply multi-resolution residual learning and view transformation encoder in bird's-eye view semantic segmentation.