IEEE/RSJ International Conference on Intelligent Robots and Systems · 239 papers
Hybrid Transformer-Mamba Model for 3D Semantic Segmentation
Xinyu Wang, Yingying Zhu
CodeSegmentationTransformerPoint Cloud
🎯 What it does: Propose the HybridTM architecture, a hybrid of Transformer and Mamba for 3D semantic segmentation, and introduce the Inner Layer Hybrid Strategy to fuse the two at a finer-grained level to simultaneously capture long-range dependencies and fine-grained features.
🎯 What it does: Developed the HyperGraph ROS system, unifying cross-process, cross-device, and cross-process computing into a computational hypergraph to enhance message passing and parallel execution efficiency.
iGaussian: Real-Time Camera Pose Estimation via Feed-Forward 3D Gaussian Splatting Inversion
Hao Wang, Haibin Yan
CodePose EstimationGaussian SplattingImage
🎯 What it does: Proposes iGaussian, a two-stage feedforward framework that realizes real-time camera pose estimation through direct 3D Gaussian inversion.
Image-Goal Navigation Using Refined Feature Guidance and Scene Graph Enhancement
Zhicheng Feng, Huimin Lu
CodeAutonomous DrivingKnowledge DistillationImage
🎯 What it does: Proposes an image goal navigation method called RFSG, which leverages fine-grained associations between goals, observations, and environments in limited image data while maintaining a lightweight and concise architecture.
🎯 What it does: Proposed a 3D multi-object tracking framework called IMM-MOT based on the Interacting Multiple Model (IMM) filter, introducing the Damping Window mechanism and the Distance-Based Score Enhancement module to improve trajectory lifecycle management and detection score processing.
ImpedanceGPT: VLM-driven Impedance Control of Swarm of Mini-drones for Intelligent Navigation in Dynamic Environment
Faryal Batool, D. Tsetserukou
CodeRobotic IntelligenceTransformerVision Language ModelRetrieval-Augmented Generation
🎯 What it does: Proposes ImpedanceGPT, a system based on vision-language models (VLM) and retrieval-augmented generation (RAG) frameworks, for achieving impedance control and adaptive navigation of small drone swarms in dynamic environments.
🎯 What it does: Propose and implement a learning scheme based on EER to leverage multi-view consistency and enhance the accuracy of a diffusion-based 2D hand trajectory prediction method.
Improved 3D Point-Line Mapping Regression for Camera Relocalization
B. Bui, Joo-Ho Lee
CodePose EstimationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a novel architecture that independently learns features for 3D points and lines, then combines them for camera relocalization.
Improved Calibration for Panoramic Annular Lens Systems with Angular Modulation
Ding Wang, Lingbao Kong
CodeImage
🎯 What it does: To address the calibration challenges of panoramic mirror systems, a new projection model incorporating angular modulation is proposed and validated on both synthetic and real datasets.
Intelligent LiDAR Navigation: Leveraging External Information and Semantic Maps with LLM as Copilot
Fujing Xie, Sören Schwertfeger
CodeAutonomous DrivingRobotic IntelligenceLarge Language ModelTextPoint Cloud
🎯 What it does: Propose combining osmAG (a semantic topological hierarchical map based on OpenStreetMap text) with a large language model (LLM) as a co-pilot for LiDAR navigation to achieve richer external information fusion;
Inverse-Free and Data-Driven Motion Tracking Control for Redundant Robot with Fuzzy Recurrent Neural Network
Min Yang, Hui Zhang
CodeRobotic IntelligenceRecurrent Neural Network
🎯 What it does: Proposed a data-driven fuzzy discrete recurrent neural network (D2-FDRNN) model to address the precise motion tracking control problem for redundant robots under unknown structural knowledge and noise interference.
J-ORA: A Framework and Multimodal Dataset for Japanese Object Identification, Reference, Action Prediction in Robot Perception
Jesse Atuhurra, Koichiro Yoshino
CodeRecognitionRobotic IntelligenceVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: Proposed the J-ORA multimodal dataset and evaluated visual language models (VLMs) on object recognition, coreference resolution, and next-action prediction tasks.
JAM: Keypoint-Guided Joint Prediction after Classification-Aware Marginal Proposal for Multi-Agent Interaction
Fangze Lin, Hong Zhang
CodeAutonomous DrivingPoint CloudBenchmark
🎯 What it does: Proposed a two-stage multi-agent interaction prediction framework named JAM, which first performs classification-aware edge prediction and then completes joint prediction under the guidance of key points to address the issue of low-quality generation of low-probability patterns.
🎯 What it does: Proposes a logits-based knowledge distillation framework, using a bird's eye view (BEV) projection model as the student, a non-projection model as the teacher, and separating moving and non-moving categories. Custom distillation strategies, dynamic upsampling, and network structure optimization are employed to improve accuracy while maintaining real-time inference speed.
Keypoint-Aware RAG for Robotic Manipulation: In-Context Constraint Learning via Large-Scale Retrieval
Jiuzhou Lin, Houde Liu
CodeRetrievalRobotic IntelligenceVision Language ModelRetrieval-Augmented Generation
🎯 What it does: Propose Keypoint-Aware Retrieval Augmented Generation (KARAG), achieving retrieval-augmented generation between vision-language models and robot datasets through keypoint constraints;
🎯 What it does: Proposed a reinforcement learning-based SE(3) manifold extrinsic calibration framework, treating calibration as a decision problem and directly optimizing extrinsic parameters to enhance odometry accuracy.
Language as Cost: Proactive Hazard Mapping using VLM for Robot Navigation
Mintaek Oh, Seong-Woo Kim
CodeRobotic IntelligenceLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposes a zero-shot language cost mapping framework based on VLM for predicting dynamic hazards and assigning risk costs to robot navigation, achieving active obstacle avoidance by integrating with geometric obstacle maps.
Large Language Model-Based Robot Task Planning from Voice Command Transcriptions
Afonso Certo, Pedro U. Lima
CodeRobotic IntelligenceLarge Language ModelText
🎯 What it does: Proposed a complete pipeline that leverages large language models (LLMs) to directly convert spoken instructions into coherent action plans, integrating environmental context into model inputs to generate more efficient and context-aware plans.
🎯 What it does: Using a Seq2Seq framework and a motion mapping model driven by deep neural networks, human speech and facial expressions are mapped to facial actions executable by robots, enabling robots to generate self-reactive facial expressions.
Learning Point Correspondences In Radar 3D Point Clouds For Radar-Inertial Odometry
J. Michalczyk, J. Steinbrener
CodePose EstimationTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Designed a Transformer-based learning framework for predicting robust point correspondences in noise-sparse SoC FMCW radar 3D point clouds.
Learning to Initialize Trajectory Optimization for Vision-Based Autonomous Flight in Unknown Environments
Yicheng Chen, Qingdong Li
CodeOptimizationRobotic IntelligenceImage
🎯 What it does: Proposes a neural network enhanced trajectory planner (NEO-Planner), which learns to predict spatial and temporal parameters from raw visual sensor observations, providing heuristic initial values for non-convex trajectory optimization to accelerate planning and support robust online replanning.
Learning-Based Passive Fault-Tolerant Control of a Quadrotor with Rotor Failure
Jiehao Chen, Yunjiang Lou
CodeRobotic IntelligenceReinforcement Learning
🎯 What it does: Proposes a learning-based passive fault-tolerant control method that can handle any single rotor failure without requiring rotor fault information or controller switching.
LGDD: Local-Global Synergistic Dual-Branch 3D Object Detection Using 4D Radar
Xiaokai Bai, Hui-Liang Shen
CodeObject DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposed a local-global collaborative dual-branch 3D object detection framework based on 4D radar, named LGDD, which includes a point-based branch (employing a voxel attention point feature extractor and cluster voting) and a pillar-based branch (utilizing query-based feature pre-fusion and proposal masks), and achieves alignment between local instances and global context through a semantic-geometry awareness fusion module.
🎯 What it does: Proposes a visual place recognition (VPR) framework with a shared lightweight keypoint extraction module, which jointly learns local keypoint extraction and VPR by utilizing self-attention and cross-attention mechanisms to fuse irregularly distributed key features.
Lightweight Temporal Transformer Decomposition for Federated Autonomous Driving
Tuong Khanh Long Do, Anh Nguyen
CodeAutonomous DrivingFederated LearningTransformerImageTime Series
🎯 What it does: Propose a lightweight time Transformer decomposition method, which splits large attention maps into smaller matrices to process continuous image frames and steering sequences, reduces model complexity, and achieves real-time autonomous driving prediction in federated learning environments.
LiMo-Calib: On-Site Fast LiDAR-Motor Calibration for Quadruped Robot-Based Panoramic 3D Sensing System
Jianping Li, Lihua Xie
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed and validated a field rapid LiDAR-servo calibration method LiMo-Calib without external targets, applicable to panoramic 3D perception systems on quadruped robot platforms.
LMMCoDrive: Cooperative Driving with Large Multimodal Models
Haichao Liu, Jun Ma
CodeAutonomous DrivingOptimizationTransformerLarge Language ModelImageMultimodality
🎯 What it does: Propose the LMMCoDrive framework, which integrates scheduling and motion planning by leveraging Large Multimodal Models (LMM) combined with bird's-eye view (BEV) representations, and enhances traffic efficiency and passenger experience in AMoD systems through decentralized safety-constrained optimization via ADMM.
🎯 What it does: The study uses low-cost BeadSight tactile sensors for pre-training and employs only visual input in downstream tasks to enhance operational performance.
🎯 What it does: Proposes LSW-Net, a self-supervised framework for learning environmental perception features from raw 2D LiDAR point clouds, incorporating an LS-Encoder with local convolutional perception and global attention, along with an interpretable weight extraction module.
🎯 What it does: Developed the lunar surface simulation system LESS and the LunarSeg dataset, and proposed the LuSeg two-stage segmentation network for positive and negative obstacle segmentation.
M2H: Multi-Task Learning with Efficient Window-Based Cross-Task Attention for Monocular Spatial Perception
U. Udugama, F. Nex
CodeSegmentationDepth EstimationTransformerImage
🎯 What it does: Proposed and implemented the Multi-Mono-Hydra (M2H) multi-task learning framework for semantic segmentation, depth estimation, edge detection, and surface normal estimation from monocular images.
MambaMap: Online Vectorized HD Map Construction using State Space Model
Ruizi Yang, Jianke Zhu
CodeAutonomous DrivingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose the MambaMap framework, achieving online vectorized high-definition map construction by fusing long-range temporal features in the state space.
ManeuverGPT Agentic Control for Safe Autonomous Stunt Maneuvers
Shawn Azdam, Aliasghar Arab
CodeAutonomous DrivingLarge Language ModelPrompt Engineering
🎯 What it does: Developed the ManeuverGPT framework, which utilizes LLM-driven agents to generate and execute high-dynamic stunt maneuvers (e.g., J-turn) in the CARLA simulation environment, and adjusts control parameters through an iterative prompting method to accomplish complex actions without retraining model weights.
ManiGaussian++: General Robotic Bimanual Manipulation with Hierarchical Gaussian World Model
Tengbo Yu, Ziwei Wang
CodeRobotic IntelligenceGaussian SplattingWorld Model
🎯 What it does: Proposes a multi-task dual-arm manipulation method called ManiGaussian++, which models multi-body spatiotemporal dynamics using a hierarchical Gaussian world model.
🎯 What it does: Propose a map-free obstacle avoidance method based on MPC and dual KD-Tree, directly generating safe actions using sparse path points and depth camera point clouds without requiring 3D map construction or trajectory tracking.
MARS-FTCP: Robust Fault-Tolerant Control and Agile Trajectory Planning for Modular Aerial Robot Systems
Rui Huang, Lin Zhao
CodeOptimizationRobotic Intelligence
🎯 What it does: Proposes a fault-tolerant control reallocation method and flexible trajectory planning method for the modular aerial robotic system (MARS), achieving fault tolerance and collision avoidance flight under multi-module and different configurations.
🎯 What it does: A multi-robot automatic robust extrinsic calibration method based on spherical targets is proposed, applicable to outdoor and extraterrestrial environments. It can accurately extract spherical and elliptical centers and solve the transformation matrix even when the target or sensors are damaged.
MaskSem: Semantic-Guided Masking for Learning 3D Hybrid High-Order Motion Representation
Wei Wei, Jianqin Yin
CodeRecognitionTransformerGraphSequential
🎯 What it does: Propose MaskSem, a semantic-guided occlusion method for learning 3D mixed high-order motion representations to enhance skeletal action recognition.
MCTrack: A Unified 3D Multi-Object Tracking Framework for Autonomous Driving
Xiyang Wang, Mu Yang
CodeObject TrackingAutonomous DrivingPoint Cloud
🎯 What it does: Proposed a unified 3D multi-object tracking framework called MCTrack, which achieves good performance on the KITTI, nuScenes, and Waymo datasets.
Mesh-Learner: Texturing Mesh with Spherical Harmonics
Yunfei Wan, Fu Zhang
CodeGenerationPoint CloudMesh
🎯 What it does: Propose a 3D reconstruction and rendering framework named Mesh-Learner, which can directly integrate with traditional rasterization pipelines, learns perspective-related radiance for each mesh, and achieves end-to-end training through spherical harmonic (SH) texture mapping.
MIAT: Maneuver-Intention-Aware Transformer for Spatio-Temporal Trajectory Prediction
Chandra Raskoti, Weizi Li
CodeAutonomous DrivingTransformerTime Series
🎯 What it does: Proposed and implemented the Maneuver-Intention-Aware Transformer (MIAT) architecture, combining maneuver-intention-aware control mechanisms with spatiotemporal interaction modeling to enhance vehicle trajectory prediction
🎯 What it does: A multi-modal keypoint learning framework, MK-Pose, is proposed to achieve category-level object pose estimation by integrating RGB images, point clouds, and category-level text descriptions.
MM-Geo: Multi-Scale and Multi-Positive UAV-View Geo-Localization
Pan Ai, Guoquan Huang
CodeRetrievalContrastive LearningImage
🎯 What it does: Propose an MM-Geo method that uses uniformly sized satellite tiles and real-time retrieval of matching tiles from drone images at different altitudes, achieving geolocation from the drone's perspective.
MobiExo: GPS-SLAM Fusion for Seamless Indoor-Outdoor Mobile Manipulation with Hand-Foot Coordination
Jianpeng Wang, F. Yu
CodeFederated LearningRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Developed the MobiExo system to achieve remote control for seamless indoor and outdoor mobile operations, integrating GPS and SLAM localization with hand-foot coordinated control.
Modeling The States of Liquid Phase Change Pouch Actuators by Reservoir Computing
Cedric Caremel, Tung D. Ta
CodeRobotic IntelligenceRecurrent Neural NetworkTime SeriesPhysics Related
🎯 What it does: Proposed and implemented a reservoir computing-based modeling method for the inflation state of liquid-phase transition foam actuators, and designed a soft gripper with dual foam actuators.
Mr. Virgil: Learning Multi-robot Visual-range Relative Localization
Si Wang, Yue Wang
CodeOptimizationRobotic IntelligenceGraph Neural NetworkSimultaneous Localization and MappingMultimodalityGraph
🎯 What it does: Proposes an end-to-end learning framework named Mr. Virgil, which uses graph neural networks for data association between UWB ranging and visual detection, and combines differentiable pose graph optimization (PGO) to achieve relative localization within the visual range of multi-robot systems.
Multi-Cali Anything: Dense Feature Multi-Frame Structure-from-Motion for Large-Scale Camera Array Calibration
Jinjiang You, W. Pu
CodeOptimizationImage
🎯 What it does: Propose a multi-frame structured light reconstruction (SfM) camera array calibration method based on dense features, which directly optimizes camera intrinsic parameters from scene data without requiring additional calibration pattern capture.
🎯 What it does: A factor graph-based multi-target association and localization method is proposed, integrating sensor measurements and control constraints from distributed UAVs to achieve joint association and localization of multiple targets. The effectiveness and robustness of the method are validated through simulations and real-world experiments.
Normalized Triangulation for Calibrated Dual-View 3D Human Pose Estimation
Zijian Zhang, Tianyi Ma
CodePose EstimationImage
🎯 What it does: Propose decomposing dual-camera 3D human pose estimation into 2D pose estimation and 2D→3D lifting, and introduce a novel technical solution for the latter.
🎯 What it does: Propose the MMTwin model, which utilizes multimodal inputs (2D RGB, 3D point clouds, historical hand trajectory, and text prompts) for 3D hand trajectory prediction.
🎯 What it does: Propose a one-shot deformable object localization method (OS-AGDO) that identifies unknown deformable objects in egocentric scenes with minimal samples.
Open-World Task Planning for Humanoid Bimanual Dexterous Manipulation via Vision-Language Models
Zixin Tang, Fei Chen
CodeRobotic IntelligenceVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a large-scale benchmark called OBiMan-Bench for evaluating open-world task planning in dual-arm dexterous manipulation, and designs a zero-shot planning framework called OBiMan-Planner based on vision-language models, which includes two modules: scenario normalization and task planning;
🎯 What it does: Proposes the OpenMIGS framework, achieving multi-granularity, information-preserving open-vocabulary 3D Gaussian rendering through constructing object-level Gaussian fields and lightweight implicit fields.
🎯 What it does: Proposes an OSM generation method utilizing low-altitude aircraft equipped with aerial sensors, with the core being a binary dual-stream road segmentation model based on LiDAR and camera data to achieve efficient path finding and generate complete OSM maps.
PL-VIWO: A Lightweight and Robust Point-Line Monocular Visual Inertial Wheel Odometry
Zhixin Zhang, Pawel Ladosz
CodePose EstimationRobotic IntelligenceSimultaneous Localization and MappingImage
🎯 What it does: Proposes a lightweight, robust point-line monocular visual inertial wheel odometry system (VIWO) for localization of ground robots in long-term complex outdoor navigation.
Pseudo Depth Meets Gaussian: A Feed-forward RGB SLAM Baseline
Linqing Zhao, Jiwen Lu
CodePose EstimationDepth EstimationGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideo
🎯 What it does: Propose an online 3D reconstruction method based on 3D Gaussian mapping, which integrates a feedforward recurrent prediction module to directly estimate camera pose from optical flow, and introduces a local graph rendering technique to enhance the robustness of pose prediction.
RDMM: Enhancing Household Robotics with On-Device Contextual Memory and Decision Making
Shady Nasrat, Seung-Joon Yi
CodeRobotic IntelligenceTransformerLarge Language ModelMultimodality
🎯 What it does: Proposes the RDMM framework, which enhances robot autonomy by leveraging large language models for domain-specific robot decision-making, integrating agent-specific knowledge representation, visual perception, and real-time speech recognition;
🎯 What it does: Proposed a Factorized Graph Sequence Encoder network for real-time recognition of human manipulation actions, and introduced the Hand Pooling operation to enhance the focus of graph layer embeddings.
Real-time Photorealistic Mapping for Situational Awareness in Robot Teleoperation
I. Page, P. Wieber
CodeComputational EfficiencyRobotic IntelligenceGaussian SplattingSimultaneous Localization and Mapping
🎯 What it does: Propose a modular, efficient GPU-based integrated solution that combines Gaussian Splatting SLAM with an existing online map-based remote operating system to enhance remote operation efficiency in unknown environments.
🎯 What it does: Propose a real-time spatiotemporal traversability assessment method, which utilizes sparse Gaussian processes to extract point cloud geometric features and construct a high-resolution local traversability map, then designs a spatiotemporal Bayesian Gaussian kernel to infer real-time traversability scores.
🎯 What it does: Propose a streamlined scene graph generation model to reduce the number of parameters in UAV applications, directly using subject-object query pairs to predict triples;
ResLPR: A LiDAR Data Restoration Network and Benchmark for Robust Place Recognition Against Weather Corruptions
Wenqing Kuang, Xieyuanli Chen
CodeRecognitionRestorationPoint CloudBenchmark
🎯 What it does: Propose a LiDAR data recovery network called ResLPRNet based on wavelet transform, and evaluate local recognition under adverse weather conditions using the ResLPR dataset
Resource-Efficient Affordance Grounding with Complementary Depth and Semantic Prompts
Yizhou Huang, Kailun Yang
CodeSegmentationPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes the BiT-Align framework, which utilizes depth images and text prompts to perform multimodal affordance map mapping on RGB images, and achieves functional region localization through the Bypass Prompt Module (BPM) and Text Feature Guidance (TFG) attention selection mechanism.
🎯 What it does: Propose a single-stage architecture that can simultaneously predict traffic elements, lane centerlines, and their topological relationships, improving accuracy and inference speed.
Revisiting 3D Curve to Surface Registration using Tangent and Normal Vectors for Computer-Assisted Orthopedic Surgery
Zhengyan Zhang, Zhe Min
CodeOptimizationMeshBiomedical Data
🎯 What it does: Proposed a bidirectional hybrid model registration method called BiHMM-DTN, which uses dual-constrained tangent vectors and normal vectors for curve-to-surface registration.
🎯 What it does: Developed a reinforcement learning method called GORA-RL based on group opinion risk assessment to enhance the safety of autonomous driving decision-making
🎯 What it does: Proposed a BEV-based 3D object detection network called ROA-BEV, which uses 2D region-oriented attention to make the backbone network focus more on feature learning in regions where targets are located.
Roadside GNSS Aided Multi-Sensor Integrated System for Vehicle Positioning in Urban Areas
Feng Huang, Li-Ta Hsu
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes the RSG-GLIO method, which leverages roadside GNSS and C-V2X collaboration, combined with onboard GNSS/LiDAR/IMU sensors to achieve reliable odometry and map building.
Robust and High-Fidelity 3D Gaussian Splatting: Fusing Pose Priors and Geometry Constraints for Texture-Deficient Outdoor Scenes
Meijun Guo, Bin Liang
CodeGenerationPose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Improved pose estimation in large-scale outdoor scenes by combining prior poses from LiDAR-IMU Odometry with COLMAP triangulation, and enhanced directional and shape consistency in scene representation by incorporating normal vector constraints and effective rank regularization into 3D Gaussian Splatting (3DGS), achieving higher-quality digital asset rendering.
Robust and Modular Multi-Limb Synchronization in Motion Stack for Space Robots with Trajectory Clamping via Hypersphere
Elian Neppel, Kazuya Yoshida
CodeRobotic Intelligence
🎯 What it does: Propose a robust and modular multi-arm synchronization method that achieves trajectory synchronization for multi-dimensional states in space robots through hyper-sphere constraints and adapts to system changes
🎯 What it does: Proposed and implemented the Joint Semantic Learning (JSL) framework, integrating the marine scene segmentation module with the object detection network during training, and removing the segmentation module during inference to achieve robust object detection without additional computational overhead.
🎯 What it does: Propose a state-based search framework that concatenates state-action pairs from incomplete demonstrations into richer training trajectories to enhance policy learning.
RRT*former: Environment-Aware Sampling-Based Motion Planning using Transformer
Mingyang Feng, Xiang Yin
CodeOptimizationRobotic IntelligenceTransformer
🎯 What it does: Proposed a sampling-based optimal path planning algorithm called RRT*former, which combines RRT* and Transformer for robot path planning in complex dynamic environments
Scalable Outdoors Autonomous Drone Flight with Visual-Inertial SLAM and Dense Submaps Built without LiDAR
Sebastián Barbas Laina, Stefan Leutenegger
CodeAutonomous DrivingSimultaneous Localization and MappingImage
🎯 What it does: Developed a miniature drone system entirely based on low-cost passive visual and inertial sensors for large-scale autonomous navigation in outdoor, unstructured, cluttered environments.
SEB-Naver: A SE(2)-based Local Navigation Framework for Car-like Robots on Uneven Terrain
Xiaoying Li (Zhejiang University), Fei Gao (Zhejiang University)
CodeOptimizationRobotic Intelligence
🎯 What it does: Designed and implemented an SE(2)-based local navigation framework called SEB-Naver for real-time terrain assessment and trajectory optimization of wheeled robots on uneven terrain.
Self-supervised 3D Reconstruction of Tibia and Fibula from Biplanar X-rays
Kai Pan, Shoudong Huang
CodeGenerationGraph Neural NetworkImageBiomedical Data
🎯 What it does: Reconstruct patient-specific 3D models of the tibia and fibula using two X-ray images from the coronal and sagittal planes combined with a generic template, integrating point-based deformation and deep learning techniques.
🎯 What it does: Proposed a self-supervised learning framework named SelfToF, which generates scale-aware and detail-rich depth maps by fusing high-resolution RGB images with low-resolution ToF depth maps, and further improved to SelfToF* to adapt to ToF signals with varying sparsity.
Self-supervised Monocular Depth Estimation for Dynamic Objects with Ground Propagation
Huan Li, Stefano Mattoccia
CodeDepth EstimationImage
🎯 What it does: Proposes a self-supervised monocular depth estimation method that leverages the relationship between ground contact points and the depth of dynamic objects, recalibrating the depth of dynamic objects by iteratively propagating ground features to moving targets within the perception layer, without adding extra networks or complex training.
Semantic-Geometric-Physical-Driven Robot Manipulation Skill Transfer via Skill Library and Tactile Representation
Mingchao Qi, Panfeng Huang
CodeRobotic IntelligenceTransformerLarge Language ModelMultimodalityChain-of-Thought
🎯 What it does: Construct a skill library based on a knowledge graph and propose a hierarchical skill transfer framework to achieve collaboration between task-level thinking and action-level precise execution;
SemGauss-SLAM: Dense Semantic Gaussian Splatting SLAM
Siting Zhu, Hesheng Wang
CodePose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposed the SemGauss-SLAM system, achieving dense semantic SLAM using 3D Gaussian representations, supporting precise semantic mapping, robust camera tracking, and high-quality rendering.
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingImageMultimodalityPoint Cloud
🎯 What it does: Implemented a semi-distributed cross-modal aerial-ground relative localization framework, enabling unmanned ground vehicles (UGV) and unmanned aerial vehicles (UAV) to independently complete SLAM and use deep learning to extract keypoints and global descriptors for relative positioning
SGLoc: Semantic Localization System for Camera Pose Estimation from 3D Gaussian Splatting Representation
Beining Xu, Hesheng Wang
CodePose EstimationGaussian SplattingImage
🎯 What it does: Propose the SGLoc system, which directly utilizes 3D Gaussian Splatting representation and semantic information to regress the camera's 6DoF pose from a query image without prior pose information.