These 140 ICRA 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 ICRA 2024 paper, free trial on arXivSub.
3D-OAE: Occlusion Auto-Encoders for Self-Supervised Learning on Point Clouds
🎯 What it does: Propose a self-supervised learning framework called 3D-OAE based on occlusion completion, which learns point cloud representations by randomly occluding local blocks of point clouds and completing them on the remaining points.
3QFP: Efficient neural implicit surface reconstruction using Tri-Quadtrees and Fourier feature Positional encoding
Shuo Sun, Martin Magnusson
CodeGenerationNeural Radiance FieldMesh
🎯 What it does: Proposes a neural implicit surface reconstruction method that utilizes a sparse tri-quadtree structure and Fourier feature position encoding.
A 3D Vector Field and Gaze Data Fusion Framework for Hand Motion Intention Prediction in Human-Robot Collaboration
Maleen Jayasuriya, Dikai Liu
CodeClassificationComputational EfficiencyRobotic IntelligenceMultimodalityTime Series
🎯 What it does: Propose a framework that combines hand trajectory and gaze data to achieve real-time prediction of hand motion intent with low computational resources
A Biomorphic Whisker Sensor for Aerial Tactile Applications
Chaoxiang Ye, S. Hamaza
CodeRobotic IntelligenceRecurrent Neural NetworkTime Series
🎯 What it does: Design, fabricate, and evaluate a lightweight follicle sensor based on MEMS barometers, and propose a recursive multi-output network (RMN) for 3D touch point localization.
A Point-to-distribution Degeneracy Detection Factor for LiDAR SLAM using Local Geometric Models
Sehua Ji, Haifei Zhu
CodeAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A LiDAR-SLAM degradation detection method based on a local geometric model using point-to-distribution matching is proposed, aiming to address the matching degradation caused by insufficient geometric features in environments such as corridors and tunnels.
🎯 What it does: Propose a soft tissue tracking method called Ada-Tracker, which uses optical flow for adjacent frame matching to obtain a rough ROI, and dynamically updates the tracking template based on estimated reliability through adaptive template matching.
Advancements in 3D Lane Detection Using LiDAR Point Clouds: From Data Collection to Model Development
Runkai Zhao, Weidong (Tom) Cai
CodeObject DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Propose the LiSV-3DLane large-scale 3D lane dataset and the LiLaDet 3D lane detection model, leveraging the geometric features and spatial attributes of LiDAR point clouds to achieve automatic annotation and detection;
🎯 What it does: Proposed the ADVGPS method to generate stealthy GPS adversarial signals, attacking multi-agent perception systems and significantly reducing target detection accuracy.
AgriSORT: A Simple Online Real-time Tracking-by-Detection framework for robotics in precision agriculture
Leonardo Saraceni, Thomas Alessandro Ciarfuglia
CodeObject TrackingVideoAgriculture Related
🎯 What it does: Propose AgriSORT, a simple, online, real-time multi-object tracking-by-detection framework based on motion information, specifically designed for precision agriculture robots.
An Extrinsic Calibration Method between LiDAR and GNSS/INS for Autonomous Driving
Jiahao Pi, Botian Shi
CodeAutonomous DrivingPoint Cloud
🎯 What it does: A three-stage extrinsic parameter calibration method between LiDAR and GNSS/INS was developed, enabling rapid and accurate calibration of the relative attitude between sensors.
An Onboard Framework for Staircases Modeling Based on Point Clouds
Chun Qing, Gan Ma
CodePose EstimationAutonomous DrivingPoint Cloud
🎯 What it does: Developed a vehicle-mounted framework for detecting navigable areas and physical property modeling of stairs based on point cloud data.
Augmenting Lane Perception and Topology Understanding with Standard Definition Navigation Maps
Katie Z Luo, Marco Pavone
CodeAutonomous DrivingTransformerGraph
🎯 What it does: Investigated the effectiveness of standard definition maps in real-time lane topology understanding, and proposed a framework to integrate standard definition maps into online map prediction.
🎯 What it does: Proposes the AYDIV framework, which integrates a three-stage alignment process to enhance long-range object detection performance under LiDAR and camera fusion.
Barrier Functions Inspired Reward Shaping for Reinforcement Learning
Nilaksh Nilaksh, Shishir N Y Kolathaya
CodeRobotic IntelligenceReinforcement Learning
🎯 What it does: Proposed and implemented a safety-oriented reward shaping framework based on barrier functions, evaluated in simulation environments such as CartPole, Ant, Humanoid, and on the Unitree Go1 quadruped robot.
Better Monocular 3D Detectors with LiDAR from the Past
Yurong You, Kilian Q. Weinberger
CodeObject DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposed the AsyncDepth framework, which provides asynchronous LiDAR features to monocular 3D detectors during inference using unlabeled historical LiDAR scans, thereby enhancing detection performance.
🎯 What it does: Proposes a multi-space aligned teacher-student framework (MATS), achieving domain adaptation for cross-domain bird's-eye view (BEV) 3D object detection through depth-aware teacher (DAT) and geometry space aligned student (GAS).
Block-Map-Based Localization in Large-Scale Environment
Yixiao Feng, Guyue Zhou
CodeAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: Propose a localization system based on Block Maps, which reduces computational load for large-scale maps through Block Map generation and switching strategies, and provides global localization using Branch-and-Bound Search and dynamic sliding window graph optimization.
🎯 What it does: Proposes the Feature Distribution-aware Aggregation (FDA) framework to address the distribution gap issue caused by independent private data in multi-agent perception systems, aiming to break data silos.
CoBT: Collaborative Programming of Behaviour Trees from One Demonstration for Robot Manipulation
Aayush Jain, M. Leva
CodeRobotic Intelligence
🎯 What it does: CoBT is a collaborative programming framework based on demonstrations that generates reactive and modular behavior trees to enable robotic manipulation.
CoLRIO: LiDAR-Ranging-Inertial Centralized State Estimation for Robotic Swarms
Shipeng Zhong, Ming Liu
CodeRobotic IntelligenceSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a centralized LiDAR-Ranging-Inertial state estimation system that enables robot swarms to collaborate in GPS-denied environments.
🎯 What it does: Proposes a complementary random masking strategy for RGB-thermal images and introduces a self-distillation loss between clean and masked input modalities;
Complementing Onboard Sensors with Satellite Maps: A New Perspective for HD Map Construction
Wenjie Gao, Nanning Zheng
CodeObject DetectionSegmentationAutonomous DrivingTransformerSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This paper proposes to utilize satellite maps to supplement vehicle-mounted sensors for high-definition (HD) map construction, and releases a corresponding supplementary dataset; meanwhile, a hierarchical fusion module is designed, which uses a mask generator and mask cross-attention to refine vehicle-mounted features at the feature level, and an alignment module to correct coordinate differences at the BEV level; this module can be seamlessly integrated into three existing HD map construction methods; experimental validation on the augmented nuScenes dataset shows significant performance improvements in HD map semantic segmentation and instance detection tasks.
Conditionally Combining Robot Skills using Large Language Models
K. Zentner, G. Sukhatme
CodeRobotic IntelligenceTransformerLarge Language ModelWorld ModelText
🎯 What it does: Proposed Language-World, an extension of the Meta-World, enabling large language models to manipulate in simulated robot environments through semi-structured natural language queries and script-based natural language descriptions, and introduced the Plan Conditioned Behavioral Cloning (PCBC) method to fine-tune high-level plans through end-to-end example demonstrations.
🎯 What it does: Proposed the CrackNex framework based on Retinex theory, which learns a unified illumination-invariant representation using reflectance information and addresses the issue of insufficient training data by combining few-shot segmentation.
🎯 What it does: Proposes an architecture named Darkness Clue-Prompted Tracking (DCPT) for nighttime drone tracking, which replaces the traditional enhancement-then-tracking workflow with darkness clue prompts.
🎯 What it does: Proposed the DeFlow network for scene flow estimation on large-scale point clouds, refining the conversion from voxel features to point features using GRU
🎯 What it does: Developed a direct imaging sonar odometry system called DISO for estimating the relative spatial transformation between two sonar image frames.
DMSA - Dense Multi Scan Adjustment for LiDAR Inertial Odometry and Global Optimization
D. Skuddis, Norbert Haala
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed the DMSA method to achieve dense joint registration of multi-frame point clouds, using full point cloud merging and gradually reducing scattering through a Gaussian distribution model based on uniform grid cells; combined with IMU measurements for sliding window continuous trajectory optimization and large-scale keyframe optimization.
🎯 What it does: Proposes a data-driven robot input vector exploration (DRIVE) protocol to collect input limits for unmanned ground vehicles (UGV) and gather training data for motion models, while introducing a slippage learning model that outperforms traditional acceleration-based learning methods.
DroneMOT: Drone-based Multi-Object Tracking Considering Detection Difficulties and Simultaneous Moving of Drones and Objects
Peng Wang, De-qin Li
CodeObject TrackingVideo
🎯 What it does: Proposed a multi-object tracking method for UAV scenarios called DroneMOT, including dual-domain integrated attention module, motion-driven association scheme, adaptive feature synchronization technology, and dual motion prediction methods, aiming to enhance detection and association performance in environments with small targets, blurriness, and occlusions.
DTPP: Differentiable Joint Conditional Prediction and Cost Evaluation for Tree Policy Planning in Autonomous Driving
Zhiyu Huang, Chen Lv
CodeAutonomous DrivingTransformer
🎯 What it does: This paper proposes a tree-structured policy planning approach and designs a differentiable joint training framework that integrates self-vehicle conditional motion prediction with cost assessment, directly enhancing the final planning performance.
DynaInsRemover: A Real-time Dynamic Instance-Aware Static 3D LiDAR Mapping Framework for Dynamic Environment
Huanfeng Zhao, Bo Zheng
CodeAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed a real-time dynamic instance-aware static 3D LiDAR mapping framework called DynaInsRemover, which efficiently removes dynamic objects while preserving more static map details by leveraging geometric differences between instances.
Fast and Robust Point Cloud Registration with Tree-based Transformer
Guangyan Chen (Beijing Institute of Technology), Yufeng Yue (Beijing Institute of Technology)
CodePose EstimationTransformerPoint Cloud
🎯 What it does: Proposed Tree-based Transformer (TrT) for point cloud registration, constructing a coarse-to-fine feature tree and using Tree-based Attention (TrA) to progressively focus on key points, achieving rich local and global feature extraction with linear complexity.
Fault Tolerant Neural Control Barrier Functions for Robotic Systems under Sensor Faults and Attacks
Hongchao Zhang, R. Poovendran
CodeRobotic Intelligence
🎯 What it does: Proposed and implemented a fault-tolerant neural control barrier function (FT-NCBF) applicable to environments with sensor faults and attacks, learned the function using a data-driven approach, and subsequently designed control inputs while formally proving its safety.
🎯 What it does: A keypoint detection and tracking method called FE-DeTr, which integrates image frames and event streams, was developed. It achieves stable and efficient keypoint detection through time response consistency supervision and employs a spatiotemporal nearest neighbor search strategy for robust tracking.
🎯 What it does: Propose a dynamic aggregation method to replace static aggregation (e.g., average or max pooling) during the fine-tuning of pre-trained point cloud Transformers
Follow the Footprints: Self-supervised Traversability Estimation for Off-road Vehicle Navigation based on Geometric and Visual Cues
Yurim Jeon, Seung-Woo Seo
CodeAutonomous DrivingRobotic IntelligenceImage
🎯 What it does: Propose a self-supervised off-road feasibility estimation method based on geometric and visual information, using a Guided Filtering Network (GFN) and Footprint Supervision Module (FSM) to predict the robot's traversable area.
GAM-Depth: Self-Supervised Indoor Depth Estimation Leveraging a Gradient-Aware Mask and Semantic Constraints
Anqi Cheng, Kezhi Mao
CodeDepth EstimationImage
🎯 What it does: Proposed the GAM-Depth model, which improves indoor self-supervised depth estimation using gradient-aware masks and semantic constraints to address depth inconsistency in textureless regions and depth differences at object boundaries.
GelRoller: A Rolling Vision-based Tactile Sensor for Large Surface Reconstruction Using Self-Supervised Photometric Stereo Method
Zhiyuan Zhang, Hua Yang
CodeDepth EstimationImagePoint Cloud
🎯 What it does: Designed a rolling cylindrical visual tactile sensor that achieves continuous and rapid perception over large surface areas through rolling, and proposed a self-supervised photometric stereo deep learning method. This method can obtain surface normals from a single frame image without prior calibration or stable illumination, followed by large-area surface reconstruction using normals and point cloud registration.
GrainGrasp: Dexterous Grasp Generation with Fine-grained Contact Guidance
Fuqiang Zhao, Qian Liu
CodeOptimizationRobotic IntelligencePoint Cloud
🎯 What it does: Proposed the GrainGrasp scheme, which uses a generative model to predict fine-grained contact maps for each finger on object point clouds, and generates precise, deterministic human-like grasping strategies based on an optimization algorithm with only point cloud input.
Ground-Fusion: A Low-cost Ground SLAM System Robust to Corner Cases
Jie Yin, Danping Zou
CodeAutonomous DrivingSimultaneous Localization and MappingMultimodality
🎯 What it does: Introduces a low-cost sensor fusion SLAM system called Ground-Fusion, suitable for ground vehicles, featuring efficient initialization, sensor anomaly detection and handling, real-time dense color mapping, and robust localization in diverse environments.
🎯 What it does: Proposed a human-robot interactive art portrait drawing system (HRICA), enabling humans and robots to alternate in drawing strokes on a canvas, and achieving collaborative creation through the robot's understanding of human intent.
Improving Radial Imbalances with Hybrid Voxelization and RadialMix for LiDAR 3D Semantic Segmentation
Jiale Li, Yong Ding
CodeSegmentationPoint Cloud
🎯 What it does: Propose the Hi-VoxelNet network, utilizing hybrid voxelization and RadialMix data augmentation to address the radial imbalance issue in LiDAR 3D semantic segmentation.
🎯 What it does: Developed a lightweight network based on key points, using an encoder-decoder framework to achieve single-view 3D human reconstruction.
Learning Agile Locomotion and Adaptive Behaviors via RL-augmented MPC
Lydia Y. Chen, Quan Nguyen
CodeRobotic IntelligenceReinforcement Learning
🎯 What it does: By combining reinforcement learning with model predictive control, the study investigates and implements adaptive balance and foot swing reflex in quadruped robots operating in vision-blind environments, thereby enhancing the robot's robustness and agility in complex terrains.
🎯 What it does: Propose the LHMap-loc pipeline to achieve accurate and efficient localization of a monocular camera in pre-built LiDAR point cloud maps
🎯 What it does: Propose Lightning NeRF, which uses an efficient hybrid scene representation and leverages LiDAR geometric priors to enhance the quality of novel view synthesis in autonomous driving environments, while reducing computational costs for training and rendering.
Lightweight Event-based Optical Flow Estimation via Iterative Deblurring
Yilun Wu, G. D. Croon
CodeComputational EfficiencyOptical FlowTime Series
🎯 What it does: Proposes a lightweight event-based optical flow estimation network, IDNet, which directly estimates optical flow using event trajectories without constructing correlation volumes.
LIKO: LiDAR, Inertial, and Kinematic Odometry for Bipedal Robots
Qingrui Zhao, Qiang Huang
CodeRobotic IntelligenceSimultaneous Localization and MappingPoint CloudTime Series
🎯 What it does: Propose a tightly coupled LiDAR-IMU-kinematic odometry (LIKO), achieving state estimation for biped robots through an iterative extended Kalman filter, while modeling and estimating foot contact positions.
🎯 What it does: Proposed a lightweight Transformer-based visual tracking model called LiteTrack, which is efficiently optimized for real-time robots and edge devices.
🎯 What it does: Propose a method to predict future driver actions by utilizing in-vehicle and external camera data, combined with manually extracted object and road-level features.
🎯 What it does: Propose a lightweight parameter-sharing network (LPS-Net) for point cloud scene recognition, which includes multi-scale bidirectional perception units and a parameter-shared NetVLAD aggregation module;
MEDL-U: Uncertainty-aware 3D Automatic Annotation based on Evidential Deep Learning
Helbert Paat, Tong Zhang
CodeAutonomous DrivingExplainability and InterpretabilityData-Centric LearningTransformerPoint Cloud
🎯 What it does: Proposed a 3D automatic annotation framework named MEDL-U based on Evidential Deep Learning (EDL), which can quantify uncertainty while generating pseudo-labels.
Merging Decision Transformers: Weight Averaging for Forming Multi-Task Policies
Daniel Lawson, A. H. Qureshi
CodeRobotic IntelligenceTransformer
🎯 What it does: Construct a multi-task model by averaging the parameter space of Decision Transformers trained on different MuJoCo locomotion tasks, without requiring centralized training.
🎯 What it does: Proposed a real-time dense depth estimation model that enhances monocular image depth prediction and resolves scale ambiguity by leveraging sparse depth priors generated from triangulated features.
🎯 What it does: Proposed MF-MOS, a motion-focused model for LiDAR moving object segmentation, which employs a dual-branch structure to separate spatial-temporal information.
🎯 What it does: We propose MonoOcc, an improved monocular semantic occupancy prediction framework that enhances performance using auxiliary semantic loss, image-conditioned cross-attention modules, and knowledge distillation.
MoPA: Multi-Modal Prior Aided Domain Adaptation for 3D Semantic Segmentation
Haozhi Cao, Lihua Xie
CodeSegmentationDomain AdaptationVision Language ModelPoint Cloud
🎯 What it does: Proposes a multi-modal prior-assisted domain adaptation method called MoPA to improve the recognition of rare objects in 3D semantic segmentation
🎯 What it does: A unified object attribute representation learning framework called MOSAIC based on multi-modal perception is studied, which aligns the knowledge of foundation models from visual, tactile, and auditory modalities, and experiments are conducted on robot interaction data to evaluate performance in object classification and grasping tasks.
Neural Informed RRT*: Learning-based Path Planning with Point Cloud State Representations under Admissible Ellipsoidal Constraints
Zhe Huang, K. Driggs-Campbell
CodeRobotic IntelligencePoint Cloud
🎯 What it does: Propose Neural Informed RRT*, combining the asymptotic optimality of RRT* with the acceleration advantages of rule-based information sampling. Utilize point cloud representation for free states, infer guided states near the optimal path through Neural Focus and PointNet++, and introduce Neural Connect to construct connectivity among guided states to improve planning efficiency.
🎯 What it does: A novel few-shot 3D point cloud scene semantic segmentation method is proposed through a meta-learning framework, utilizing multi-prototype graph construction, graph structure-based denoising, subgraph bagging schemes for semi-supervised transfer learning, and triplet contrastive loss to enhance prototype feature discriminability.
OCC-VO: Dense Mapping via 3D Occupancy-Based Visual Odometry for Autonomous Driving
Heng Li, Yanyong Zhang
CodeAutonomous DrivingTransformerSimultaneous Localization and MappingImage
🎯 What it does: Propose the OCC-VO framework, which uses deep learning to convert 2D camera images into 3D semantic occupancy grids, achieving visual odometry and global semantic map construction without requiring traditional pose and feature point estimation.
Offline Goal-Conditioned Reinforcement Learning for Safety-Critical Tasks with Recovery Policy
Chenyang Cao, Xueqian Wang
CodeReinforcement Learning
🎯 What it does: Proposes a method called Recovery-based Supervised Learning (RbSL) to address offline goal-conditional reinforcement learning tasks with safety constraints, aiming to achieve multi-objective safety-critical missions.
OmniColor: A Global Camera Pose Optimization Approach of LiDAR-360Camera Fusion for Colorizing Point Clouds
Bonan Liu, Pan Hui
CodePose EstimationAutonomous DrivingOptimizationSimultaneous Localization and MappingImageMultimodalityPoint Cloud
🎯 What it does: Propose OmniColor, which utilizes an independent 360-degree camera to jointly optimize camera poses through photometric consistency optimization, mapping images to LiDAR-based point clouds to achieve point cloud coloring.
🎯 What it does: Propose OmniLRS—a lighting-realistic lunar simulator based on the Nvidia robotics simulator, offering fast procedural environment generation, multi-robot capabilities, synthetic data pipelines for machine learning, and support for ROS1/ROS2 bindings.
CodeSimultaneous Localization and MappingBenchmark
🎯 What it does: Proposes conversion methods between three camera models, enabling the mutual conversion of data calibrated under different projection models, allowing the use of existing data without recalibration;
OptiState: State Estimation of Legged Robots using Gated Networks with Transformer-based Vision and Kalman Filtering
Alexander Schperberg, Dennis Hong
CodeOptimizationRobotic IntelligenceRecurrent Neural NetworkTransformerAuto EncoderSimultaneous Localization and MappingImage
🎯 What it does: A hybrid method is proposed, combining Kalman filtering, convex MPC optimization, and learning techniques. It utilizes joint encoders, IMU, and ground reaction force control outputs, and employs GRU gated networks and Vision Transformer autoencoders to perform semantic and height reasoning on depth images, achieving precise estimation of the quadruped robot's trunk state.
Outram: One-shot Global Localization via Triangulated Scene Graph and Global Outlier Pruning
Peng Yin, Lihua Xie
CodePose EstimationRetrievalAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Propose a hierarchical one-time LiDAR localization algorithm called Outram, which utilizes substructures of 3D scene graphs for local consistency correspondence search and global substructure-level outlier removal, combining feature retrieval and correspondence extraction to achieve consistency refinement from local to global, addressing substructure ambiguity issues.
🎯 What it does: Proposes a parameter-efficient Prompt tuning method called PPT to adapt large multimodal models for 3D point cloud understanding tasks.
PeLiCal: Targetless Extrinsic Calibration via Penetrating Lines for RGB-D Cameras with Limited Co-visibility
Jaeho Shin, Ayoung Kim
CodePose EstimationDepth EstimationImage
🎯 What it does: Proposes PeLiCal, a target-free, real-time, and robust RGB-D camera extrinsic calibration method based on long straight line features in the scene.
Physical Priors Augmented Event-Based 3D Reconstruction
Jiaxu Wang, Renjing Xu
CodeGenerationNeural Radiance FieldPhysics Related
🎯 What it does: Propose to enhance NeRF training by leveraging motion, geometry, and density priors, and introduce a density-guided patch sampling strategy that enables direct 3D scene reconstruction from event streams, while constructing the first large-scale event-based 3D reconstruction dataset.
🎯 What it does: This paper studies the robustness of RGB-D and point cloud visual control strategies under different visual conditions, and proposes the Point Cloud World Model (PCWM) and point cloud-based control strategies.
PointSSC: A Cooperative Vehicle-Infrastructure Point Cloud Benchmark for Semantic Scene Completion
Yuxiang Yan, Jian Pu
CodeSegmentationAutonomous DrivingTransformerVision Language ModelPoint CloudBenchmark
🎯 What it does: Proposed PointSSC, the first cooperative vehicle-infrastructure point cloud semantic scene completion benchmark, and developed an automatic annotation pipeline based on Semantic Segment Anything; simultaneously proposed a LiDAR-based model that combines Spatial-Aware Transformer for global and local feature extraction, and introduces a Completion and Segmentation Cooperative Module for joint completion and segmentation.
🎯 What it does: Proposed a new interactive object grasping task called Pragmatic-IOG, created the corresponding dataset IM-Dial, and designed a system named PROGrasp capable of interpreting user intent and completing target object recognition and grasping.
QuadricsNet: Learning Concise Representation for Geometric Primitives in Point Clouds
Ji Wu, Gui-Song Xia
CodeRepresentation LearningPoint Cloud
🎯 What it does: Proposed a new framework that represents 3D point cloud geometry primitives using quadratic surfaces and implemented an end-to-end learning model called QuadricsNet to parse quadratic surfaces.
RaSim: A Range-aware High-fidelity RGB-D Data Simulation Pipeline for Real-world Applications
Xingyu Liu, Xiangyang Ji
CodeData SynthesisImage
🎯 What it does: Developed RaSim, a range-aware high-fidelity RGB-D data simulation pipeline to generate realistic depth images and bridge the simulation-to-reality gap.
RaTrack: Moving Object Detection and Tracking with 4D Radar Point Cloud
Zhijun Pan, Chris Xiaoxuan Lu
CodeObject DetectionObject TrackingPoint Cloud
🎯 What it does: Proposes an algorithm called RaTrack for moving object detection and tracking using 4D radar point clouds, focusing on motion segmentation and clustering, and equipped with a motion estimation module;
🎯 What it does: Proposed a new contrastive divergence loss to address the non-IID problem of autonomous driving data in federated learning environments
🎯 What it does: Proposes the RenderOcc framework, which employs a NeRF-style 3D volume representation and volumetric rendering techniques to train a 3D occupancy prediction model using only 2D semantic and depth labels, and introduces an auxiliary ray method to address sparse viewpoint issues.
RGB-based Category-level Object Pose Estimation via Decoupled Metric Scale Recovery
Jiaxin Wei, Pan Ji
CodePose EstimationImage
🎯 What it does: Propose an RGB method that decouples 6D pose estimation and size estimation, leveraging a monocular pre-trained estimator to extract local geometric information and restoring the true scale through category-level scale statistics, finally using RANSAC-PnP for robust pose solving.
RoboLLM: Robotic Vision Tasks Grounded on Multimodal Large Language Models
Zijun Long, Gerardo Aragon Camarasa
CodeRecognitionObject DetectionSegmentationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelMultimodality
🎯 What it does: Propose the RoboLLM framework, which leverages the BEiT-3 multimodal large language model to uniformly address tasks such as object detection, segmentation, and recognition in robot vision.
Robust and Remote Center of Cyclic Motion Control for Redundant Robots with Partially Unknown Structure
Long Jin, Mei Liu
CodeRobotic Intelligence
🎯 What it does: Developed an acceleration-level cyclic motion remote center (ARC2M) control scheme, and proposed a computational method for estimating unknown end-effector parameters under noise influence.