These 120 ICRA 2023 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 2023 paper, free trial on arXivSub.
3D-DAT: 3D-Dataset Annotation Toolkit for Robotic Vision
🎯 What it does: Proposed a 3D dataset annotation tool called 3D-DAT based on pure RGB, utilizing Neural Radiance Fields (NeRF) to achieve scene-level automatic object alignment and annotation;
🎯 What it does: Propose the 3DSGrasp strategy, which uses a Transformer encoder-decoder network to predict missing geometry, achieving a complete 3D point cloud and thereby generating reliable grasping poses.
A generic diffusion-based approach for 3D human pose prediction in the wild
Saeed Saadatnejad, Alexandre Alahi
CodePose EstimationDiffusion modelSequential
🎯 What it does: Proposes a 3D human pose prediction method based on diffusion models, utilizing a denoising framework to handle noisy inputs and missing elements, applicable for long-term prediction in real-world scenarios;
A Sequential Quadratic Programming Approach to the Solution of Open-Loop Generalized Nash Equilibria
Edward L. Zhu, F. Borrelli
CodeOptimization
🎯 What it does: Proposed a numerical method for solving the local generalized Nash equilibrium of open general and dynamic games with nonlinear dynamics and constraints.
AANet: Aggregation and Alignment Network with Semi-hard Positive Sample Mining for Hierarchical Place Recognition
Feng Lu, Chun Yuan
CodeRecognitionContrastive LearningImage
🎯 What it does: Proposed a unified AANet network, combining a global feature aggregation module and a dynamic alignment of local features module to achieve hierarchical visual localization, and introduced a semi-hard positive sample mining strategy to enhance network robustness.
CodeAutonomous DrivingExplainability and InterpretabilityTransformerVision-Language-Action ModelVideo
🎯 What it does: Proposes ADAPT, an end-to-end Transformer-based architecture that jointly trains driving captioning and vehicle control prediction tasks through shared video representations, providing user-friendly natural language narratives and reasoning for each decision step.
Adaptive Risk-Tendency: Nano Drone Navigation in Cluttered Environments with Distributional Reinforcement Learning
Cheng Liu, G. D. Croon
CodeReinforcement Learning
🎯 What it does: Propose a distributed reinforcement learning framework to generate adaptive risk-averse strategies, studying navigation of nano quadrotor robots in unknown crowded environments.
🎯 What it does: Propose a tracking and fusion algorithm based on computer vision, specifically designed for a 3D-printed gimbal system flying in natural environments, utilizing the skyline and ground plane as references to achieve robust camera orientation control.
🎯 What it does: This paper studies the infrastructure LiDAR localization problem, proposing an efficient process to find optimal installation positions in real simulation environments, and building a real LiDAR simulation library capable of mimicking the characteristics of various mainstream LiDARs. It uses generated high-fidelity point cloud data and multiple detection models to assess the perception accuracy of different localization schemes, further analyzing the correlation between perception performance and point cloud density and uniformity within regions of interest.
🎯 What it does: Proposes AvoidBench, a high-fidelity benchmark suite for evaluating multirotor visual obstacle avoidance algorithms, integrating RotorS dynamics with Unity3D virtual environments to provide performance and environmental metrics.
🎯 What it does: Propose the BEVFusion framework, unifying multi-modal features into a shared bird's-eye view (BEV) to achieve multi-task multi-sensor fusion.
🎯 What it does: Improved the accuracy of point cloud registration by transferring knowledge from pre-trained multi-modal models to a new point cloud descriptor neural network, using only single-modal point cloud data during inference.
🎯 What it does: Propose CalibDepth, which uses depth maps as a unified representation for images and LiDAR point clouds, and introduces a monocular depth estimation subnetwork to assist online calibration; treat online calibration as a sequence prediction problem, and optimize results using global and local losses.
Calibration and Uncertainty Characterization for Ultra-Wideband Two-Way-Ranging Measurements
Mohamed Fouad Shalaby, J. L. Ny
CodeAutonomous DrivingSimultaneous Localization and MappingTabular
🎯 What it does: Proposed a UWB bidirectional ranging protocol and developed a scalable antenna delay calibration method, while modeling error and uncertainty as a function of received signal power, and applying it to positioning using an extended Kalman filter.
CAROM Air - Vehicle Localization and Traffic Scene Reconstruction from Aerial Videos
Duo Lu, Yezhou Yang
CodeObject TrackingVideoBenchmark
🎯 What it does: Automatically extract vehicle trajectory data from aerial videos captured by consumer-grade drones to reconstruct traffic scenes and enable precise reproduction.
Cerberus: Low-Drift Visual-Inertial-Leg Odometry For Agile Locomotion
Shuozhi Yang, Zachary Manchester
CodeRobotic IntelligenceSimultaneous Localization and MappingImage
🎯 What it does: Proposed an open-source visual-inertial-leg odometry (VILO) state estimator called Cerberus, which can accurately estimate position in real-time across various terrains using stereo cameras, IMU, joint encoders, and contact sensors.
🎯 What it does: Proposes a continuity-aware latent cross-frame information mining framework called ClimRT, aimed at enhancing the reliability of drone tracking.
🎯 What it does: Proposes the CrossDTR method, which includes a lightweight depth predictor and a cross-perspective depth-guided Transformer for 3D object detection.
🎯 What it does: Propose a multi-modal imitation learning method based on curriculum learning, which uses data point weighting and entropy rewards to specialize the model on representable sub-data, and covers all data through a Mixture of Experts (MoE).
🎯 What it does: Propose a 3D object detection method called D-Align based on multi-frame point cloud sequences, which generates powerful bird's-eye-view (BEV) features and completes detection tasks by aligning and aggregating features from target frames and support frames.
Deep Underwater Monocular Depth Estimation with Single-Beam Echosounder
Haowen Liu, Alberto Quattrini Li
CodeData SynthesisDepth EstimationImage
🎯 What it does: Propose a self-supervised monocular depth estimation method for underwater environments using low-cost single-beam echo sounders (SBES) and generate a synthetic dataset.
Demonstration-Guided Reinforcement Learning with Efficient Exploration for Task Automation of Surgical Robot
Tao Huang, Qingxu Dou
CodeRobotic IntelligenceReinforcement LearningBiomedical Data
🎯 What it does: Proposed a reinforcement learning algorithm called DEX based on expert demonstrations to improve exploration efficiency in the automation of surgical robot tasks.
🎯 What it does: Propose to fine-tune depth estimation using LiDAR or RGB videos in an unsupervised manner to enhance monocular 3D detection performance.
DFR-FastMOT: Detection Failure Resistant Tracker for Fast Multi-Object Tracking Based on Sensor Fusion
Mohamed Nagy, S. Javed
CodeObject TrackingMultimodalityPoint Cloud
🎯 What it does: Propose a lightweight multi-object tracking method DFR-FastMOT based on camera and LiDAR sensor fusion, utilizing algebraic target association and fusion to achieve long-term memory for handling occlusions.
DS-K3DOM: 3-D Dynamic Occupancy Mapping with Kernel Inference and Dempster-Shafer Evidential Theory
Ju Han, Han-Lim Choi
CodeSimultaneous Localization and Mapping
🎯 What it does: Propose a 3-D dynamic occupancy mapping algorithm DS-K3DOM, which performs sequential updates on measurement streams using Bayesian methods based on the theory of random finite sets, and realizes real-time computation through particle approximation in the Dempster-Shafer domain. Furthermore, dense mapping from sparse measurements is achieved by employing kernel-based reasoning and Dirichlet basic belief allocation.
Dynamic Control Barrier Function-based Model Predictive Control to Safety-Critical Obstacle-Avoidance of Mobile Robot
Zhu Jian, Bin Liang
CodeRobotic IntelligencePoint Cloud
🎯 What it does: A safety avoidance method for dynamic obstacles in LiDAR-based mobile robots, which uses point clouds to generate real-time local grid maps, clusters obstacles with DBSCAN and applies minimum bounding ellipse (MBE) closure, estimates/predicts obstacle motion trajectories using data association and Kalman filtering, parameterizes trajectories as a set of ellipses, and achieves safe dynamic obstacle avoidance by combining extended dynamic control barrier functions (D-CBF) with model predictive control (MPC).
🎯 What it does: Propose an edge-guided multi-domain RGB-to-TIR image translation model, and use the generated realistic TIR images to train TIR tasks such as optical flow estimation and object detection.
🎯 What it does: Developed an exit-perception object tracker EXOT, which uses a robot's hand-mounted camera to detect whether the target object disappears during operation, thereby deciding whether to continue the operation.
FDLNet: Boosting Real-time Semantic Segmentation by Image-size Convolution via Frequency Domain Learning
Qingqing Yan, Qi Chen
CodeSegmentationConvolutional Neural NetworkImage
🎯 What it does: Proposes a real-time semantic segmentation network based on frequency domain learning called FDLNet, and designs Image Size Convolution (IS-Conv), Global Structural Representation Path (GSRP), and Decomposed Stereoscopic Attention (FSA) modules.
FloorplanNet: Learning Topometric Floorplan Matching for Robot Localization
Delin Feng, Liangjun Zhang
CodeData SynthesisRobotic IntelligenceGraph Neural NetworkSimultaneous Localization and MappingPoint CloudGraph
🎯 What it does: Proposes FloorplanNet, which matches robot-measured metric maps with building floorplans using semantic information, and applies this matching to robot localization; utilizes graph neural networks to learn node descriptors from vertex-metric graphs, enabling the matching of 3D point cloud submaps with 2D floorplans;
Follow The Rules: Online Signal Temporal Logic Tree Search for Guided Imitation Learning in Stochastic Domains
J. J. Aloor, S. Scherer
CodeOptimizationReinforcement Learning
🎯 What it does: A heuristic method combining Signal Temporal Logic (STL) rules with Monte Carlo Tree Search (MCTS) guides learners to achieve better constraint satisfaction in stochastic domains, thereby improving the performance of example-based learning (LfD) strategies.
🎯 What it does: Propose FreDSNet, a deep learning solution for achieving semantic 3D understanding of indoor environments from a single panoramic image, jointly realizing monocular depth estimation and semantic segmentation.
🎯 What it does: Developed a framework and robotic planning system that enables humans to collaborate with robots to complete canvas painting through language descriptions or images.
From Semi-supervised to Omni-supervised Room Layout Estimation Using Point Clouds
Huanhuan Gao, H. Zha
CodeSegmentationPoint CloudBenchmark
🎯 What it does: This paper proposes a new framework for the point cloud room layout estimation task, transitioning from semi-supervised to fully supervised. The core components include a quad set matching strategy, a dedicated consistency loss based on layout quadrilaterals, and an online pseudo-label collection algorithm that does not require thresholds;
Ground then Navigate: Language-guided Navigation in Dynamic Scenes
Kanishk Jain, Vineet Gandhi
CodeSegmentationAutonomous DrivingExplainability and InterpretabilityVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: In outdoor autonomous driving environments, explicitly benchmarking navigable areas using language instructions, with the model predicting corresponding segmentation masks at each moment to complete the visual language navigation task.
🎯 What it does: Proposes a simple fine-tuning method called Incremental Two-stage Fine-tuning Approach (iTFA) for incremental few-shot object detection, which separates the RoI feature extractor and classifier into base class and new class branches after base class training, and only uses a small number of new class samples to fine-tune the new class branch.
Informable Multi-Objective and Multi-Directional RRT* System for Robot Path Planning
Jiunn-Kai Huang, J. Grizzle
CodeOptimizationRobotic Intelligence
🎯 What it does: Proposed a real-time iterative system for simultaneously solving multi-objective path planning problems and determining the destination visit order.
🎯 What it does: Proposed a target-based joint calibration method for camera intrinsics and LiDAR-camera extrinsic parameters, designed a novel calibration board with four circular holes surrounding a chessboard, and constructed a cost function under reprojection constraints to solve camera intrinsics, distortion coefficients, and LiDAR-camera extrinsic parameters.
Knowledge Distillation for Feature Extraction in Underwater VSLAM
Jinghe Yang, Yen-Yu Pu
CodeData SynthesisKnowledge DistillationSimultaneous Localization and MappingImage
🎯 What it does: Propose a cross-modal knowledge distillation framework to train an underwater feature detection and matching network UFEN, and integrate it into ORB-SLAM3 to replace ORB features.
kollagen: A Collaborative SLAM Pose Graph Generator
Roberto C. Sundin, David Umsonst
CodeGenerationData SynthesisSimultaneous Localization and MappingGraph
🎯 What it does: Proposed a collaborative SLAM pose graph generator named Kollagen, which can generate reproducible pose graph datasets based on user-defined parameters.
LATITUDE: Robotic Global Localization with Truncated Dynamic Low-pass Filter in City-scale NeRF
Z. Zhu, Guyue Zhou
CodeRobotic IntelligenceNeural Radiance FieldSimultaneous Localization and MappingImage
🎯 What it does: Proposes the LATITUDE two-stage city-scale NeRF global localization framework, which includes a regressor trained on NeRF images to provide initial poses, and pose optimization achieved by minimizing the residual between observed and rendered images on the tangent plane combined with a truncated dynamic low-pass filter (TDLF).
CodeRobotic IntelligenceTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: Proposes a language-based framework that leverages pre-trained language models and visual models to encode user intent and target objects, generating trajectories applicable to different robot platforms.
R. Allen (Massachusetts Institute of Technology), D. Rus (Massachusetts Institute of Technology)
CodeComputational EfficiencyRobotic Intelligence
🎯 What it does: Proposed and implemented the LRMM model for real-time estimation of coherent risk measures in high-dimensional dynamical systems within partially observable environments.
Learning to Influence Vehicles' Routing in Mixed-Autonomy Networks by Dynamically Controlling the Headway of Autonomous Cars
Xiaoyu Ma, Negar Mehr
CodeAutonomous DrivingReinforcement Learning
🎯 What it does: In mixed automated networks, a method is proposed to influence vehicle path selection by dynamically controlling the inter-vehicle distance of autonomous vehicles, thereby reducing congestion, and training reinforcement learning strategies to achieve this goal.
Learning-Based Dimensionality Reduction for Computing Compact and Effective Local Feature Descriptors
Hao Dong, C. Stachniss
CodeRetrievalCompressionImage
🎯 What it does: Utilize a lightweight multilayer perceptron (MLP) to low-dimensionalize local feature descriptors, enhancing descriptor quality while reducing storage and computational costs, and conduct a comprehensive analysis across unsupervised, semi-supervised, and supervised settings; evaluate on tasks including visual localization, patch verification, image matching, and retrieval.
Light-Weight Pointcloud Representation with Sparse Gaussian Process
Mahmoud Ali, Lantao Liu
CodeCompressionPoint Cloud
🎯 What it does: Propose a framework that compresses high-fidelity point cloud sensor observations into a compact form using sparse Gaussian processes to achieve efficient communication and storage.
🎯 What it does: Proposes Loc-NeRF, a real-time visual robot localization method that integrates Monte Carlo localization with neural radiance fields (NeRF), utilizing a pre-trained NeRF as the map and achieving real-time localization using only RGB cameras.
🎯 What it does: Train a low-level controller to respond instantly to changes in quadrotor dynamics without requiring prior knowledge or parameter tuning.
Model Predictive Optimized Path Integral Strategies
Dylan M. Asmar, M. Kochenderfer
CodeOptimizationReinforcement Learning
🎯 What it does: Rewrite MPPI as a single joint distribution across control sequences and introduce adaptive importance sampling to improve sampling efficiency
Model- and Acceleration-based Pursuit Controller for High-Performance Autonomous Racing
Jonathan Becker, Michele Magno
CodeAutonomous DrivingOptimization
🎯 What it does: Designed and verified a model-based and acceleration-based tracking controller (MAP) for high-speed autonomous racing trajectory tracking.
🎯 What it does: Real-time monocular 3D reconstruction system supporting semantic fusion, fast motion tracking, non-rigid object deformation, and topological changes.
Monocular Visual-Inertial Odometry with Planar Regularities
Chuchu Chen, Guoquan Huang
CodePose EstimationSimultaneous Localization and MappingImage
🎯 What it does: Designed a real-time monocular visual-inertial odometry system that utilizes planar features for complete constraints through a lightweight multi-state constraint Kalman filter (MSCKF)
Multi-to-Single Knowledge Distillation for Point Cloud Semantic Segmentation
Shoumeng Qiu, Jian Pu
CodeSegmentationKnowledge DistillationPoint Cloud
🎯 What it does: Propose a multi-to-single knowledge distillation framework that focuses only on hard class instances, and employs multi-layer distillation (feature, logit, affinity) and instance-aware affinity algorithms to enhance the performance of 3D point cloud semantic segmentation for hard classes.
Neural-Kalman GNSS/INS Navigation for Precision Agriculture
Yayun Du, M. Srivastava
CodeSimultaneous Localization and MappingVideoAgriculture Related
🎯 What it does: Proposed a lightweight neural-Kalman filter, a user-friendly video processing toolbox, and a publicly available precision agriculture robot neural-inertial navigation dataset.
Obstacle avoidance using Raycasting and Riemannian Motion Policies at kHz rates for MAVs
Michael Pantic, Lionel Ott
CodeRobotic IntelligencePoint Cloud
🎯 What it does: Propose a method for real-time obstacle avoidance on voxelized maps using GPU ray casting and thousands of parallel Riemannian Motion Policies (RMP), demonstrating successful avoidance of static and dynamic obstacles on a real MAV.
🎯 What it does: Proposes a hierarchical imitation learning algorithm based on adversarial inverse reinforcement learning, which directly recovers hierarchical policies from unannotated demonstrations using the EM algorithm, introduces directional information terms to enhance causal relationships, and employs variational autoencoders (VAEs) to achieve end-to-end learning.
Orbeez-SLAM: A Real-time Monocular Visual SLAM with ORB Features and NeRF-realized Mapping
Chi-Ming Chung, Winston H. Hsu
CodeNeural Radiance FieldSimultaneous Localization and MappingImage
🎯 What it does: Developed a real-time monocular visual SLAM called Orbeez-SLAM, which combines ORB features and NeRF to achieve dense map mapping, enabling rapid adaptation to new scenes without pre-training and generating real-time dense maps.
Perturbation-Based Best Arm Identification for Efficient Task Planning with Monte-Carlo Tree Search
Daejong Jin, Kyungjae Lee
CodeOptimizationReinforcement Learning
🎯 What it does: Propose a tree search method based on perturbation-driven optimal arm identification (PBAI) to improve Monte Carlo Tree Search (MCTS) in task and motion planning, achieving better balance between exploration and exploitation while accelerating discovery of globally optimal plans.
Place Recognition under Occlusion and Changing Appearance via Disentangled Representations
Yue Chen, Xingyu Chen
CodeRecognitionImage
🎯 What it does: Propose an unsupervised method called PROCA that decomposes image representations into scene codes, appearance codes, and occlusion codes for scene recognition under occlusions and appearance variations.
Pose Relation Transformer Refine Occlusions for Human Pose Estimation
Hyung-gun Chi, K. Ramani
CodePose EstimationTransformer
🎯 What it does: Proposed a module named POse Relation Transformer (PORT) for correcting missing keypoints caused by occlusion in human pose estimation.
🎯 What it does: Propose the PredRecon framework, which utilizes autonomous path generation by drones to achieve fast and high-quality 3D reconstruction.
🎯 What it does: Propose the PriorLane framework, which utilizes an encoder-only Transformer to integrate features from pre-trained segmentation models with low-cost local prior knowledge embeddings, thereby enhancing lane detection segmentation performance.
🎯 What it does: Implemented a real-time reinforcement learning system called Remote-Local Distributed (ReLoD), which can distribute the computational tasks of Soft Actor-Critic (SAC) and Proximal Policy Optimization (PPO) between local resource-constrained computers and remote high-performance computers;
🎯 What it does: Proposes a RGB-Event fusion network called RENet to enhance the performance of moving object detection in autonomous driving environments.
🎯 What it does: This study proposes an anomaly segmentation method based on pixel-level logit variance and iterative background highlighting, utilizing logit information from pre-trained semantic segmentation networks to identify abnormal regions in urban scenes.
🎯 What it does: Propose the CoAlign framework, which utilizes proxy-target pose graph modeling and multi-scale feature fusion to enhance collaborative 3D object detection performance under pose errors.
Robust Imaging Sonar-based Place Recognition and Localization in Underwater Environments
Hogyun Kim, Younggun Cho
CodePose EstimationSimultaneous Localization and MappingImage
🎯 What it does: A robust pose recognition and loop closure method based on imaging SONAR is proposed, achieving pose estimation and loop closure correction through geometric information encoding of raw SONAR measurements, hierarchical search, adaptive translation and filling, and ICP (Iterative Closest Point).
Segregator: Global Point Cloud Registration with Semantic and Geometric Cues
Peng Yin, Lihua Xie
CodePose EstimationPoint Cloud
🎯 What it does: Proposed a global point cloud registration framework called Segregator, which efficiently constructs robust correspondences against anomalies and identifies inliers by leveraging semantic information and geometric distribution.
🎯 What it does: Proposes an online meta-reinforcement learning algorithm based on hypothesis-driven online prospective adaptation (COLA) for achieving adaptive driving in non-stationary environments.
Self-Improving Safety Performance of Reinforcement Learning Based Driving with Black-Box Verification Algorithms
Resul Dagdanov, N. K. Ure
CodeAutonomous DrivingSafty and PrivacyReinforcement Learning
🎯 What it does: Propose a self-improving AI system that enhances the safety performance of reinforcement learning-based autonomous driving through black-box verification methods
Semantic-SuPer: A Semantic-aware Surgical Perception Framework for Endoscopic Tissue Identification, Reconstruction, and Tracking
Shan Lin, Michael C. Yip
CodeClassificationObject TrackingSegmentationConvolutional Neural NetworkVideoBiomedical Data
🎯 What it does: Introduces a comprehensive surgical perception framework called Semantic-SuPer for the identification, 3D reconstruction, and tracking of tissues in endoscopic videos.
Sequential Bayesian Optimization for Adaptive Informative Path Planning with Multimodal Sensing
Joshua Ott, Mykel J. Kochenderfer
CodeOptimizationMultimodality
🎯 What it does: Proposes the problem of adaptive information path planning under multi-modal perception (AIPPMS), formulating it as a belief Markov decision process with Gaussian process beliefs; solving the problem using sequential Bayesian optimization and online planning methods.
SHAIL: Safety-Aware Hierarchical Adversarial Imitation Learning for Autonomous Driving in Urban Environments
Arec L. Jamgochian, M. Kochenderfer
CodeAutonomous Driving
🎯 What it does: Learn a high-level policy using safety-aware hierarchical adversarial imitation learning (SHAIL), selecting from a set of low-level controllers to achieve safe and human-like driving decisions in urban roundabout scenarios.
Sonicverse: A Multisensory Simulation Platform for Embodied Household Agents that See and Hear
Ruohan Gao, Jiajun Wu
CodeData SynthesisDomain AdaptationRobotic IntelligenceSimultaneous Localization and MappingWorld ModelVideoMultimodalityAudio
🎯 What it does: Propose Sonicverse, a multisensory simulation platform that can real-time render audio in 3D environments and train home agents capable of both visual and auditory perception.
🎯 What it does: Proposed an uncertainty-aware mean teacher framework for source-free unsupervised domain adaptation of 3D object detection networks under adverse weather conditions.
🎯 What it does: Proposed and implemented a self-supervised joint nighttime image enhancement and depth estimation framework without using any real labels.
🎯 What it does: Propose a non-autoregressive Transformer model for simultaneously predicting human trajectory and pose to achieve robot front-following tasks.
Structure PLP-SLAM: Efficient Sparse Mapping and Localization using Point, Line and Plane for Monocular, RGB-D and Stereo Cameras
Fangwen Shu, D. Stricker
CodeOptimizationSimultaneous Localization and MappingImage
🎯 What it does: Proposed a real-time visual SLAM system based on points, lines, and planes (PPR), achieving camera localization and sparse geometric reconstruction under multi-sensor conditions (monocular, RGB-D, stereo);
Suture Thread Spline Reconstruction from Endoscopic Images for Robotic Surgery with Reliability-driven Keypoint Detection
Neelay Joglekar, Michael C. Yip
CodeSegmentationRobotic IntelligenceBiomedical Data
🎯 What it does: Reconstructing 3D centerlines from segmented surgical image pairs using reliable keypoint detection and Minimum Variation Spline (MVS) smoothing optimization