RSS 2023 Papers with AI Summaries
Robotics: Science and Systems · 112 papers
→ RSS 2023 papers with code (22)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 112 RSS 2023 papers by keyword, author or institution —
free trial on arXivSub.
A Correct-and-Certify Approach to Self-Supervise Object Pose Estimators via Ensemble Self-Training
Jingnan Shi (Massachusetts Institute of Technology), Luca Carlone (Massachusetts Institute of Technology)
Pose EstimationOptimizationConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes an integrated self-supervised training framework that jointly trains multiple target pose estimators using discriminative observable correctness certificates and robust correctors, thereby improving pose estimation performance without requiring 3D annotations.
A Sampling-Based Approach for Heterogeneous Coalition Scheduling with Temporal Uncertainty
Andrew Messing (Georgia Institute of Technology), Seth Hutchinson (Georgia Institute of Technology)
OptimizationRobotic Intelligence
🎯 What it does: Propose a sampling-based risk-aware heterogeneous team task scheduling algorithm (CS-HSSRG) that generates task sequences satisfying user risk tolerance under time uncertainty;
Active Collaborative Localization in Heterogeneous Robot Teams
Igor Spasojevic (University of Pennsylvania), Vijay Kumar (University of Pennsylvania)
OptimizationRobotic IntelligenceSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Studies how to deploy a small number of resource-rich LiDAR robots in large-scale SWAP-constrained aerial robot teams to actively place 'artificial landmarks' and improve the positioning accuracy of Visual-Inertial Odometry (VIO).
Active Velocity Estimation using Light Curtains via Self-Supervised Multi-Armed Bandits
Siddharth Ancha (Massachusetts Institute of Technology), David Held (Massachusetts Institute of Technology)
Autonomous DrivingRobotic IntelligenceReinforcement Learning
🎯 What it does: Use programmable light curtains with dynamic occupancy grids and particle filters to actively estimate the position and velocity of obstacles in dynamic environments, and design multiple light curtain placement strategies.
Adaptive Tracking Control of Dielectric Elastomer Soft Actuators with Viscoelastic Hysteresis Compensation
Yunhua Zhao (Beihang University), Li Wen (Beihang University)
OptimizationRobotic IntelligenceTime Series
🎯 What it does: Propose a high-precision tracking control method for flexible dielectric elastomers (DEAs) based on an improved Prandtl-Ishlinskii model and adaptive inverse control.
An Efficient Multi-solution Solver for the Inverse Kinematics of 3-Section Constant-Curvature Robots
Ke Qiu (Zhejiang University), Yue Wang (Zhejiang University)
OptimizationRobotic IntelligenceTabular
🎯 What it does: Proposed an efficient multi-solution inverse kinematics solver for posture control of three-segment constant curvature soft robots.
AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System
Yuzhe Qin (UC San Diego), Dieter Fox (NVIDIA)
Pose EstimationDepth EstimationRobotic IntelligenceImageVideoMesh
🎯 What it does: Developed AnyTeleop, a generic visual-based teleoperation system supporting multiple robotic arms, finger robots, different simulators or real environments, arbitrary camera configurations, and enabling remote and collaborative teleoperation.
Autonomous Justification for Enabling Explainable Decision Support in Human-Robot Teaming
Matthew Luebbers, Bradley Hayes (University of Colorado Boulder)
Explainability and InterpretabilityRobotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes an adaptive timing framework based on the Value of Information (VOI) theory to provide explanatory justifications in a timely manner during robot-human collaborative tasks when expectations mismatch; meanwhile, four types of contrastive explanations (environment ontology/policy ontology × local/global) are defined and evaluated.
Autonomous Navigation, Mapping and Exploration with Gaussian Processes
Mahmoud Ali (Indiana University), Lantao Liu (Indiana University)
Autonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This study proposes a framework based on sparse Gaussian processes that can simultaneously accomplish autonomous navigation, mapping, and exploration in unknown environments;
Bandit Submodular Maximization for Multi-Robot Coordination in Unpredictable and Partially Observable Environments
Zirui Xu (University of Michigan), Vasileios Tzoumas (University of Michigan)
OptimizationRobotic IntelligenceReinforcement LearningTime SeriesSequential
🎯 What it does: This paper proposes the Bandit Sequential Greedy algorithm for submodular optimization in multi-robot collaboration within unpredictable and partially observable environments, leveraging bandit feedback.
Behavior Retrieval: Few-Shot Imitation Learning by Querying Unlabeled Datasets
Maximilian Du (Stanford University), Chelsea Finn (Stanford University)
Data-Centric LearningRobotic IntelligenceAuto EncoderRetrieval-Augmented Generation
🎯 What it does: Proposes a behavior retrieval method that leverages a small amount of expert data to retrieve and utilize large-scale unlabeled offline data for few-shot imitation learning.
Beyond Flat GelSight Sensors: Simulation of Optical Tactile Sensors of Complex Morphologies for Sim2Real Learning
Daniel Fernandes Gomes (University of Liverpool), Paolo Paoletti (University of Liverpool)
Data SynthesisDomain AdaptationRobotic IntelligenceConvolutional Neural NetworkImageMeshPhysics Related
🎯 What it does: This paper proposes a geodesic-based optical tactile sensor simulation method, extending the original simulation applicable only to planar GelSight sensors to curved surface GelTip sensors. In the MuJoCo environment, synthetic tactile images are generated through light propagation, elastic deformation approximation, and the Phong reflection model, followed by comparison with images collected by real GelTip sensors and evaluation of its performance in Sim2Real tasks.
Bridging Active Exploration and Uncertainty-Aware Deployment Using Probabilistic Ensemble Neural Network Dynamics
Taekyung Kim (AI Autonomy Technology Center, Agency for Defense Development), Seongil Hong (AI Autonomy Technology Center, Agency for Defense Development)
Robotic IntelligenceReinforcement LearningTabularTime Series
🎯 What it does: Propose a unified model-based reinforcement learning framework that integrates active exploration with uncertainty-aware deployment, using parallelized probabilistic ensemble neural networks for robot dynamics learning.
Causal Policy Gradient for Whole-Body Mobile Manipulation
Jiaheng Hu (University of Texas at Austin), Roberto Martín-Martín (University of Texas at Austin)
Robotic IntelligenceReinforcement LearningImage
🎯 What it does: Proposes a framework named Causal MoMa, which automatically identifies causal relationships between action dimensions and reward terms in mobile manipulation tasks, and utilizes causal policy gradients to train whole-body mobile manipulation robots. The framework can rapidly learn multi-task goals in simulation and achieve zero-shot transfer to real robots.
CCIL: Context-conditioned imitation learning for urban driving
Ke Guo (University of Hong Kong), Jia Pan (University of Hong Kong)
Autonomous DrivingTransformerReinforcement LearningGraphTabular
🎯 What it does: Propose a context-based imitation learning method that directly predicts vehicle future trajectories using context states, eliminating explicit ego information.
Centralized Model Predictive Control for Collaborative Loco-Manipulation
Flavio De Vincenti (ETH Zurich), Stelian Coros (ETH Zurich)
OptimizationRobotic IntelligencePhysics Related
🎯 What it does: This paper proposes a centralized model predictive control (MPC) framework for multiple quadruped robots equipped with single-arm manipulators to collaboratively perform ground mobility and object manipulation (collaborative local manipulation) tasks.
Cherry-Picking with Reinforcement Learning
Yunchu Zhang (Carnegie Mellon University), Siddhartha Srinivasa (University of Washington)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Utilizing reinforcement learning for fine-grained dynamic grasping tasks at the end-effector of a low-cost manipulator, the CherryBot system is designed to achieve rapid and recoverable grasping of unstable objects.
CHSEL: Producing Diverse Plausible Pose Estimates from Contact and Free Space Data
Sheng Zhong (University of Michigan), Nima Fazeli (University of Michigan)
Pose EstimationOptimizationPoint Cloud
🎯 What it does: Proposes a feasible pose estimation framework named CHSEL, which integrates tactile, free space, and target volume information to generate diverse feasible pose sets from sparse contact data.
CLIP-Fields: Weakly Supervised Semantic Fields for Robotic Memory
Nur Muhammad (Mahi)Shafiullah, Arthur Szlam
SegmentationRobotic IntelligenceVision Language ModelVision-Language-Action ModelNeural Radiance FieldContrastive LearningImageTextMultimodalityPoint Cloud
🎯 What it does: By weakly supervising with unsupervised Web pre-training models, we construct CLIP-Fields implicit semantic fields, enabling 3D scene mapping to semantic vectors, which can be used for segmentation, instance recognition, semantic search, and view localization tasks.
Co-optimization of Morphology and Behavior of Modular Robots via Hierarchical Deep Reinforcement Learning
Jieqiang Sun (Jilin University), Bo Zheng (Changchun University of Science and Technology)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: Proposed a hierarchical deep reinforcement learning based collaborative optimization framework that can simultaneously optimize the morphology and motion of modular robots, thereby enhancing their adaptability in multi-task scenarios.
CoDEPS: Online Continual Learning for Depth Estimation and Panoptic Segmentation
Niclas Vödisch, Abhinav Valada (University of Freiburg)
SegmentationDepth EstimationAutonomous DrivingConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes CoDEPS, a joint framework for monocular depth estimation and panoramic segmentation tailored for online continuous learning in robots, capable of adaptively adjusting to new environments in real-time video streams.
ConceptFusion: Open-set multimodal 3D mapping
Krishna Murthy Jatavallabhula (MIT), Antonio Torralba (MIT)
Autonomous DrivingRobotic IntelligenceConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningSimultaneous Localization and MappingImageTextMultimodalityPoint CloudAudio
🎯 What it does: Propose ConceptFusion, which can build an open-set, multi-modal 3D map online from RGB and depth images, and support query methods such as text, image, audio, and clicks.
Concurrent Constrained Optimization of Unknown Rewards for Multi-Robot Task Allocation
Sukriti Singh (Georgia Institute of Technology), Harish Ravichandar (Georgia Institute of Technology)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: The study investigates how to online and parallel optimize robot allocation in multi-robot teams when the task reward function is unknown, aiming to maximize total reward.
Convex Geometric Motion Planning on Lie Groups via Moment Relaxation
Sangli Teng (University of Michigan), Maani Ghaffari Jadidi
OptimizationComputational EfficiencyRobotic IntelligenceTabular
🎯 What it does: Discretize rigid body motion planning (dynamically constrained trajectory optimization) on SE(3) using Lie group variational integrator (LGVI), obtaining precise quadratic polynomial dynamics; subsequently formulate the entire planning problem as a low-order sparse polynomial optimization problem (POP), and solve it using Lasserre's hierarchy SDP relaxation to obtain a global optimal solution or proof of infeasibility.
Decentralization and Acceleration Enables Large-Scale Bundle Adjustment
Taosha Fan (Meta AI), Mustafa Mukadam (Meta AI)
OptimizationComputational EfficiencyPoint Cloud
🎯 What it does: Propose a fully decentralized and accelerated bundle adjustment method called DABA, achieving scalable optimization for large-scale 3D reconstruction;
Deep RL at Scale: Sorting Waste in Office Buildings with a Fleet of Mobile Manipulators
Alexander Herzog (Everyday Robots), Sergey Levine (Everyday Robots)
Data SynthesisRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningGenerative Adversarial NetworkImageSequential
🎯 What it does: In the office waste classification task, end-to-end deep reinforcement learning is used to train multiple robots to collaboratively perform waste sorting, combining simulation, scripting, online data collection, and visual masks.
Demonstrating A Walk in the Park: Learning to Walk in 20 Minutes With Model-Free Reinforcement Learning
Ilya Kostrikov (UC Berkeley), Sergey Levine (UC Berkeley)
Robotic IntelligenceReinforcement Learning
🎯 What it does: In the real world, using model-agnostic deep reinforcement learning, the A1 quadruped robot learns to walk on five different terrains within 20 minutes.
Demonstrating Arena-Web: A Web-based Development and Benchmarking Platform for Autonomous Navigation Approaches
Linh Kästner (Berlin Institute of Technology), Jens Lambrecht (Berlin Institute of Technology)
Robotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: Developed a web-based development and evaluation platform called Arena-Web, where users can create maps, scenes, network structures, reward functions, and hyperparameters directly in the browser, train DRL robot navigation models, and conduct unified evaluations against traditional planners.
Demonstrating Large Language Models on Robots
Andy Zeng (Google DeepMind), Vikas Sindhwani (Google DeepMind)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: The paper presents four robot systems based on large language models (SayCan, Socratic Models, Inner Monologue, Code as Policies), which convert natural language instructions into executable robot actions through natural language dialogue, achieving remote and local robot closed-loop demonstrations.
Demonstrating Large-Scale Package Manipulation via Learned Metrics of Pick Success
Shuai Li (Amazon Robotics), Kostas Bekris (Amazon Robotics)
OptimizationRobotic IntelligenceImageTabular
🎯 What it does: This paper deployed a pick success prediction model trained on historical production data in the Amazon Robotics 'Robin' work cell, achieving large-scale package picking and localization, processing over 200 million packages in actual operations with a 98% pick success rate.
Demonstrating Mobile Manipulation in the Wild: A Metrics-Driven Approach
Max Bajracharya (Toyota Research Institute), Mark Tjersland (Toyota Research Institute)
Explainability and InterpretabilityRobotic IntelligenceConvolutional Neural NetworkContrastive LearningSimultaneous Localization and MappingImageVideo
🎯 What it does: Proposed a complete mobile manipulation system TTT, and conducted end-to-end field tests in a real unmodified grocery store
Demonstrating RFUniverse: A Multiphysics Simulation Platform for Embodied AI
Haoyuan Fu (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerReinforcement LearningMultimodalityPhysics Related
🎯 What it does: Built RFUniverse — a Unity-based multi-physics coupled simulation platform supporting rigid bodies, multi-body dynamics, gas-liquid interactions, and heat transfer, providing complete interaction and training functions through gRPC, Python, VR, ROS-free MoveIt, and gym-style interfaces; conducted reinforcement learning tasks such as food cutting, water pushing, and towel capturing, as well as planning and control experiments for butter pushing on this platform.
DexPBT: Scaling up Dexterous Manipulation for Hand-Arm Systems with Population Based Training
Aleksei Petrenko, Viktor Makoviychuk
Hyperparameter SearchRobotic IntelligenceRecurrent Neural NetworkReinforcement Learning
🎯 What it does: Utilizing deep reinforcement learning and large-scale GPU parallel physics simulation (Isaac Gym), the Allegro Hand + Kuka arm system is trained to perform complex tasks such as grasping, throwing, and repositioning. A distributed, decentralized Population Based Training (PBT) significantly enhances exploration efficiency and hyperparameter search.
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
Cheng Chi (Columbia University), Shuran Song (Columbia University)
Robotic IntelligenceTransformerDiffusion modelVideoSequentialStochastic Differential Equation
🎯 What it does: This paper proposes a novel visual-motion policy learning framework called Diffusion Policy, which maps conditional denoising diffusion processes to a robot's action space, enabling the generation and inference of high-dimensional action sequences.
Distributed Hierarchical Distribution Control for Very-Large-Scale Clustered Multi-Agent Systems
Augustinos D Saravanos (Georgia Institute of Technology), Evangelos Theodorou (Georgia Institute of Technology)
OptimizationRobotic IntelligenceGraphOrdinary Differential Equation
🎯 What it does: The paper proposes a hierarchical distributed control framework (DHDC), which first estimates the initial and target Gaussian distributions for each layer's clique through the bottom-up DHDE, and then applies covariance steering in the top-down DHDS to perform safe and distributed trajectory planning for all agents' state distributions, applicable to million-scale robot swarms.
Dynamic-Resolution Model Learning for Object Pile Manipulation
Yixuan Wang (University of Illinois Urbana-Champaign), Jiajun Wu (Stanford University)
OptimizationComputational EfficiencyRobotic IntelligenceGraph Neural NetworkImagePoint Cloud
🎯 What it does: Proposes a dynamic resolution model learning framework that adaptively selects the number of particles for particle representation based on task progress in object stacking tasks such as robotic grasping, reorganization, and sorting, thereby generating more efficient manipulation sequences in model predictive control (MPC).
Efficient Reinforcement Learning for Autonomous Driving with Parameterized Skills and Priors
Letian Wang (University of Toronto), Steven L Waslander
Autonomous DrivingReinforcement Learning
🎯 What it does: Propose the ASAP-RL algorithm, which performs reinforcement learning in a parameterized motion skill space combined with expert prior knowledge to improve the learning efficiency and performance of autonomous driving.
Efficient volumetric mapping of multi-scale environments using wavelet-based compression
Victor Reijgwart (ETH Zurich), Lionel Ott (ETH Zurich)
CompressionComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a hierarchical voxel map framework based on wavelet compression (wavemap), achieving efficient voxel mapping and real-time updates for multi-scale environments.
Enabling Team of Teams: A Trust Inference and Propagation (TIP) Model in Multi-Human Multi-Robot Teams
Yaohui Guo (University of Michigan), Cong Shi (University of Michigan)
Robotic IntelligenceTabular
🎯 What it does: Proposed and verified the TIP (Trust Inference and Propagation) model for trust inference and propagation in teams of multiple humans and robots, evaluated through experiments involving 15 participants and two drones performing a search detection task.
Energy-based Models are Zero-Shot Planners for Compositional Scene Rearrangement
Nikolaos Gkanatsios (Carnegie Mellon University), Katerina Fragkiadaki (Carnegie Mellon University)
OptimizationRobotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose a zero-shot compositional scene rearrangement framework (SREM) that translates language instructions into target object layouts and executes them through energy models, semantic parsing, and visual language grounding.
ERASOR2: Instance-Aware Robust 3D Mapping of the Static World in Dynamic Scenes
Hyungtae Lim (KAIST), Cyrill Stachniss (University of Bonn)
Autonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Developed the ERASOR2 algorithm, which uses instance-aware dynamic object separation to achieve static 3D map construction.
Fast Monocular Visual-Inertial Initialization Leveraging Learned Single-View Depth
Nathaniel W Merrill (University of Delaware), Guoquan Huang (University of Delaware)
Pose EstimationDepth EstimationConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: Propose a linear initialization method utilizing a single-image scale-free depth network, capable of rapidly and robustly recovering the initial state of a monocular visual inertial navigation system (VINS) within an extremely short time.
Fast Traversability Estimation for Wild Visual Navigation
Jonas Frey (ETH Zurich), Marco Hutter (ETH Zurich)
SegmentationAutonomous DrivingRobotic IntelligenceTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes an online self-supervised traversability estimation method called WVN based on monocular RGB images, leveraging visual Transformer features, SLIC segmentation, and real-time graphical supervision generation to achieve rapid adaptation;
Few-shot Adaptation for Manipulating Granular Materials Under Domain Shift
Yifan Zhu (University of Illinois Urbana-Champaign), Kris Hauser (University of Illinois Urbana-Champaign)
Domain AdaptationOptimizationRobotic IntelligenceMeta LearningConvolutional Neural NetworkImageMultimodality
🎯 What it does: Proposed an adaptive mining strategy based on deep Gaussian processes and Bayesian optimization, which can rapidly improve the sampling volume with minimal online experience under extreme domain shift.
Follow my Advice: Assume-Guarantee Approach to Task Planning with Human in the Loop
Georg Schuppe (KTH Royal Institute of Technology), Jana Tumova (KTH Royal Institute of Technology)
Robotic Intelligence
🎯 What it does: For human-in-the-loop robot task planning, the authors adopt the assume-guarantee approach, generating minimally restrictive recommendations for humans by solving weakest sufficient assumptions, thereby achieving task execution while satisfying finite LTL constraints.
FurnitureBench: Reproducible Real-World Benchmark for Long-Horizon Complex Manipulation
Minho Heo (KAIST), Joseph J Lim
Robotic IntelligenceReinforcement LearningVideoMultimodalityBenchmark
🎯 What it does: Proposes a reproducible real-world furniture assembly benchmark, FurnitureBench, which includes 3D-printable furniture models, robot environment setups, demonstration data, and simulation environments.
G*: A New Approach to Bounding Curvature Constrained Shortest Paths through Dubins Gates
Satyanarayana Gupta Manyam (Infoscitex Corp), Sivakumar Rathinam (Texas A&M University)
OptimizationBenchmark
🎯 What it does: This paper addresses the curvature-constrained shortest path (CSP) problem in 2D space with a minimum turning radius constraint, introducing a new lower bound solving method called Gate* (G*). By relaxing obstacle constraints and allowing paths to jump within vertical gates, it constructs a directed acyclic graph (DAG) and computes edge lengths using the Dubins gate problem, obtaining a provable lower bound.
Gait design for limbless obstacle aided locomotion using geometric mechanics
Baxi Chong (Georgia Institute of Technology), Grigoriy Blekherman (Carnegie Mellon University)
OptimizationRobotic IntelligenceTime SeriesSequentialPhysics Related
🎯 What it does: This paper constructs a motion model for climbing robots in obstacle-rich environments based on geometric mechanics, and designs specialized gait templates for different obstacle densities (elliptical gaits for sparse environments and undulatory gaits for dense environments). Subsequently, these gaits were validated through robot experiments to demonstrate their effectiveness in obstacle-assisted locomotion.
GenAug: Retargeting behaviors to unseen situations via Generative Augmentation
Qiuyu Chen (University of Washington), Vikash Kumar (Meta AI)
Data SynthesisRobotic IntelligenceVision Language ModelDiffusion modelImageTextMesh
🎯 What it does: Leverage pre-trained text-to-image generation models to semantically augment a small amount of real robot demonstration data, generating diverse image observations (including object textures, backgrounds, clutter, etc.), and use these augmented data to train visual control policies.
Goal-Conditioned Imitation Learning using Score-based Diffusion Policies
Moritz Reuss (Karlsruhe Institute of Technology), Rudolf Lioutikov (Karlsruhe Institute of Technology)
Robotic IntelligenceTransformerDiffusion modelScore-based ModelSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A novel goal-conditioned strategy BESO is constructed using score-based diffusion models to learn multi-modal behaviors from offline play data without rewards.
GranularGym: High Performance Simulation for Robotic Tasks with Granular Materials
David R Millard, Gaurav S Sukhatme (University of Southern California)
Robotic IntelligenceReinforcement LearningPhysics Related
🎯 What it does: Developed a high-speed granular physics simulation engine called GranularGym based on CUDA GPU, capable of simulating interactions between tens of thousands of particles and arbitrary geometry rigid bodies in real-time or super-real-time speed, and providing an OpenAI Gym interface for reinforcement learning training.
Graph Attention Multi-Agent Fleet Autonomy for Advanced Air Mobility
Malintha Fernando (Indiana University), Martin Swany (Indiana University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Propose a deep multi-agent reinforcement learning framework based on heterogeneous graph attention encoding-decoding (HetGAT Enc-Dec) for coordinating drones with different capacities in dynamic demand and limited communication range scenarios for advanced air mobility (AAM) fleet operations;
HDVIO: Improving Localization and Disturbance Estimation with Hybrid Dynamics VIO
Giovanni Cioffi (University of Zurich), Davide Scaramuzza (University of Zurich)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingMultimodalityTime SeriesPhysics Related
🎯 What it does: This paper proposes HDVIO, a method that uses a hybrid dynamic model (point mass model + learning residual) in visual-inertial odometry (VIO) to simultaneously estimate drone attitude and external forces.
Hindsight States: Blending Sim & Real Task Elements for Efficient Reinforcement Learning
Simon Guist (Max Planck Institute for Intelligent Systems), Dieter Büchler (Max Planck Institute for Intelligent Systems)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Propose a method called 'Hindsight States (HiS)' used in hybrid simulation and real environment (HySR), which significantly improves sample efficiency in reinforcement learning by generating multiple virtual trajectories in parallel based on a single real trajectory and selecting useful transitions after backward relabeling.
How To Not Train Your Dragon: Training-free Embodied Object Goal Navigation with Semantic Frontiers
Junting Chen (ETH Zürich), Fisher Yu (King Abdullah University of Science and Technology)
Robotic IntelligenceConvolutional Neural NetworkTransformerVision-Language-Action ModelSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposes an untrained ObjectNav navigation method (StructNav) based on visual SLAM and structured scene representation.
Incremental Nonlinear Dynamic Inversion based Optical Flow Control for Flying Robots: An Efficient Data-driven Approach
Hann Woei Ho (Universiti Sains Malaysia), Ye Zhou (Delft University of Technology)
Computational EfficiencyRobotic IntelligenceOptical FlowImageTime Series
🎯 What it does: This paper proposes an Incremental Nonlinear Dynamic Inversion (INDI) method for vision-based micro aerial vehicle control, achieving safe landing on dynamic ground platforms.
IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality
Bingjie Tang (University of Southern California), Yashraj S Narang
Domain AdaptationRobotic IntelligenceReinforcement LearningMeshBenchmark
🎯 What it does: Developed a reinforcement learning (RL) solution for contact-rich assembly tasks from simulation to the real world, including novel algorithms, complete hardware/software systems, and reproducible tools.
InstaLoc: One-shot Global Lidar Localisation in Indoor Environments through Instance Learning
Lintong Zhang (University of Oxford), Maurice Fallon (University of Oxford)
SegmentationData SynthesisPose EstimationConvolutional Neural NetworkContrastive LearningSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Developed an indoor global localization system (InstaLoc) based on a single LiDAR scan, achieving localization against a prior map through instance-level semantic segmentation and descriptor matching.
Integrated Object Deformation and Contact Patch Estimation from Visuo-Tactile Feedback
Mark J Van der Merwe, Nima Fazeli (University of Michigan)
Robotic IntelligenceNeural Radiance FieldImageMultimodalityPoint Cloud
🎯 What it does: Studied a neural implicit field called NDCF for simultaneously predicting the geometry of deforming objects and contact surfaces, inferred through visual-tactile feedback.
Investigating the Impact of Experience on a User's Ability to Perform Hierarchical Abstraction
Nina M Moorman (Georgia Institute of Technology), Matthew Gombolay (Georgia Institute of Technology)
Robotic IntelligenceTabular
🎯 What it does: In experiments, non-expert users performed kinesthetic demonstrations using a custom interface to record subtasks, and trained robots to complete long-sequence tasks across five home domains, evaluating the impact of experience on subtask abstraction, teaching efficiency, and redundancy.
iPlanner: Imperative Path Planning
Fan Yang (ETH Zurich), Marco Hutter (ETH Zurich)
Autonomous DrivingOptimizationConvolutional Neural NetworkImage
🎯 What it does: Proposes a deep end-to-end path planning framework based on Imperative Learning, generating smooth, collision-free navigation paths through a network and differentiable trajectory optimization using single-frame depth maps.
Language-Driven Representation Learning for Robotics
Siddharth Karamcheti, Percy Liang (Toyota Research Institute)
Representation LearningRobotic IntelligenceTransformerVision-Language-Action ModelAuto EncoderContrastive LearningVideoTextBenchmark
🎯 What it does: Propose a language-driven visual representation learning framework called Voltron, aiming to learn visual representations that can capture low-level spatial details and understand high-level semantics, while building a benchmark suite across five categories of robot learning tasks;
LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning
Kenneth Shaw (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningVideo
🎯 What it does: Proposed a low-cost, easy-to-assemble, durable, and ergonomically structured robotic hand called LEAP Hand, supporting robot learning tasks.
Learning and Adapting Agile Locomotion Skills by Transferring Experience
Laura M Smith (UC Berkeley), Sergey Levine (UC Berkeley)
Domain AdaptationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: Propose a transfer learning framework called TWiRL that leverages existing controller experience to train and adapt agile movements for the quadruped robot A1, such as obstacle jumping and bipedal walking.
Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
Tony Z. Zhao (Stanford University), Chelsea Finn (Stanford University)
Robotic IntelligenceTransformerVision-Language-Action ModelAuto EncoderImageSequential
🎯 What it does: Developed a low-cost dual-arm teleoperation system ALOHA, achieving fine manipulation tasks via end-to-end visual-to-action imitation learning.
Learning-Free Grasping of Unknown Objects Using Hidden Superquadrics
Yuwei Wu (National University of Singapore), Gregory S Chirikjian
Robotic IntelligencePoint Cloud
🎯 What it does: Propose a learning-free two-stage robotic grasping method, first recovering hidden superquadrics from point clouds, then synthesizing anti-symmetric grasping poses based on their symmetry and evaluating scores.
Local object crop collision network for efficient simulation of non-convex objects in GPU-based simulators
Dongwon Son (KAIST), Beomjoon Kim (KAIST)
OptimizationComputational EfficiencyConvolutional Neural NetworkPoint CloudMeshPhysics Related
🎯 What it does: This paper proposes a Local Clipping-based Object Collision Network (LOCC) that efficiently detects collisions between non-convex objects in a GPU simulator, and integrates it into Brax to form BRAX-LOCC, achieving higher speed and better simulation accuracy under large-scale parallel simulations.
Metric-Free Exploration for Topological Mapping by Task and Motion Imitation in Feature Space
Yuhang He (University of Oxford), Chen Feng (New York University)
Robotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: Proposes a lightweight metric-free visual exploration framework called DeepExplorer, which actively constructs a topological map by utilizing task-motion planning in the image feature space, achieving efficient exploration and navigation in unknown environments.
Motion Planning (In)feasibility Detection using a Prior Roadmap via Path and Cut Search
Yoonchang Sung (University of Texas at Austin), Peter Stone (University of Texas at Austin)
Autonomous DrivingOptimizationGraph
🎯 What it does: Leveraging the prior occurrence probability of edges in a road network, two iterative path/cut search algorithms, IPC and IDPC, are proposed to rapidly detect feasibility or infeasibility in motion planning problems while significantly reducing the number of edge evaluations.
MultiSCOPE: Disambiguating In-Hand Object Poses with Proprioception and Tactile Feedback
Andrea Sipos (University of Michigan), Nima Fazeli (University of Michigan)
Pose EstimationOptimizationRobotic IntelligenceSimultaneous Localization and Mapping
🎯 What it does: Proposed the MultiSCOPE method, which significantly reduces the pose uncertainty of dual-arm grasping objects through continuous grasping collisions, robot body perception, and 6-DOF torque sensor data.
NeuSE: Neural SE(3)-Equivariant Embedding for Consistent Spatial Understanding with Objects
Jiahui Fu (MIT), John Leonard (MIT)
Data SynthesisPose EstimationSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose NeuSE—an SE(3)-equivariant neural object embedding for object SLAM, capable of maintaining consistent spatial understanding in long-term scene changes.
Non-Euclidean Motion Planning with Graphs of Geodesically-Convex Sets
Thomas B Cohn, Russ Tedrake (Massachusetts Institute of Technology)
OptimizationRobotic IntelligenceGraph
🎯 What it does: This paper proposes extending traditional GCS (Graph of Convex Sets) to Riemannian geometry, constructing a 'GGCS' (Graph of Geodesically-Convex Sets) framework to compute globally optimal, collision-safe trajectories in non-Euclidean configuration spaces that include mobile bases and continuous rotational joints.
On discrete symmetries of robotics systems: A group-theoretic and data-driven analysis
Daniel F Ordonez-Apraez (Institut de Robòtica i Informàtica Industrial), Francesc Moreno (Institut de Robòtica i Informàtica Industrial)
Robotic IntelligenceConvolutional Neural NetworkTime Series
🎯 What it does: This paper proposes a theoretical framework of Discrete Morphological Symmetry (DMS), identifying the symmetry group of a dynamic system, and utilizing these symmetries for data augmentation and constructing G-equivariant neural networks to enhance the sample efficiency and generalization ability of data-driven models.
One Policy to Dress Them All: Learning to Dress People with Diverse Poses and Garments
Yufei Wang (Carnegie Mellon University), David Held (Carnegie Mellon University)
Domain AdaptationKnowledge DistillationRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningPoint Cloud
🎯 What it does: Proposed a robot-assisted dressing system based on a single learning strategy, capable of automatically dressing individuals in various postures and multiple types of clothing using partial point cloud observations from a single camera.
Path Planning for Multiple Tethered Robots Using Topological Braids
Muqing Cao (Nanyang Technological University), Lihua Xie (Nanyang Technological University)
Robotic Intelligence
🎯 What it does: Study multi-cable robot path planning and propose a tangle-free path generation method based on the combination rule.
PATO: Policy Assisted TeleOperation for Scalable Robot Data Collection
Shivin Dass (University of Southern California), Stefanos Nikolaidis (University of Southern California)
Robotic IntelligenceRecurrent Neural NetworkReinforcement LearningAuto EncoderMultimodality
🎯 What it does: Automatically complete repetitive subtasks in robot data collection by learning hierarchical auxiliary strategies (upper-level subgoal prediction, lower-level subgoal-to-action execution), requesting human intervention only when the strategy is uncertain, thereby reducing operator workload and improving data collection efficiency.
POV-SLAM: Probabilistic Object-Aware Variational SLAM in Semi-Static Environments
Jingxing Qian (University of Toronto Institute for Aerospace Studies and University of Toronto Robotics Institute), Angela Schoellig (University of Toronto Institute for Aerospace Studies and University of Toronto Robotics Institute)
Robotic IntelligenceSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: This paper proposes POV-SLAM, a probabilistic object-aware variational SLAM framework designed for semi-static environments, capable of simultaneously estimating robot poses and object consistency while tracking changes in semi-static objects over long time spans.
Pre-Training for Robots: Offline RL Enables Learning New Tasks in a Handful of Trials
Aviral Kumar (UC Berkeley), Sergey Levine (Stanford University)
Robotic IntelligenceConvolutional Neural NetworkSupervised Fine-TuningReinforcement LearningMixture of Experts
🎯 What it does: This paper proposes the PTR (Pre-Training for Robots) framework, which pre-trains using large-scale offline multi-task robot data, and then rapidly fine-tunes with a small amount (10–15 samples) of target task demonstrations or self-collected data, achieving efficient learning of new tasks in new environments.
Precise Object Sliding with Top Contact via Asymmetric Dual Limit Surfaces
Xili Yi (University of Michigan), Nima Fazeli (University of Michigan)
OptimizationRobotic IntelligenceImageTabular
🎯 What it does: Proposed an asymmetric double limit surface model and designed an open-loop stable planning algorithm to achieve precise object sliding through top contact only.
Predefined-Time Convergent Motion Control for Heterogeneous Continuum Robots
Ning Tan (Sun Yat-sen University), Kai Huang (Sun Yat-sen University)
Robotic IntelligenceTime Series
🎯 What it does: Propose a new predefined time convergence zeroing dynamics (PTCZD) model, and based on this model, design model-free state estimators and inverse kinematics solvers, achieving a unified method for motion control of heterogeneous continuum robots.
Progressive Learning for Physics-informed Neural Motion Planning
Ruiqi Ni (Purdue University), Ahmed H Qureshi
Robotic IntelligenceConvolutional Neural NetworkPhysics Related
🎯 What it does: This study proposes a physics-informed neural motion planning method, utilizing progressive learning and viscosity terms to rewrite the Eikonal equation, achieving high-dimensional robot path planning without requiring expert trajectories.
RADIUS: Risk-Aware, Real-Time, Reachability-Based Motion Planning
Challen Enninful Adu (University of Michigan), Ram Vasudevan (University of Michigan)
Autonomous DrivingOptimization
🎯 What it does: Proposes RADIUS—a real-time trajectory planning framework based on reachability and risk constraints in dynamic environments with uncertain obstacle positions.
Reachability-based Trajectory Design with Neural Implicit Safety Constraints
Jonathan B Michaux, Ram Vasudevan (University of Michigan)
OptimizationSafty and PrivacyRobotic IntelligenceTabular
🎯 What it does: This paper proposes a signature distance function (RDF) based on reachability, which realizes real-time prediction of the distance between the robot's link reachable body and obstacles using neural implicit representations, and embeds it into a recursive time-domain trajectory planning framework to achieve safe, real-time motion planning for multi-degree-of-freedom robotic arms;
Reconfigurable Robot Control Using Flexible Coupling Mechanisms
Sha Yi (Carnegie Mellon University), Zeynep Temel (Carnegie Mellon University)
OptimizationRobotic Intelligence
🎯 What it does: This paper designs a flexible soft anchor coupling mechanism, enabling robots to easily couple and decouple without consuming additional energy, while maintaining structural stability in multi-robot configurations;
RoboNinja: Learning an Adaptive Cutting Policy for Multi-Material Objects
Zhenjia Xu (University of Massachusetts Amherst), Shuran Song (University of Massachusetts Amherst)
OptimizationRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningMesh
🎯 What it does: Developed a learning-based multi-material cutting system, RoboNinja, which utilizes interactive perception to estimate hard core positions and adaptively controls the tool to maximize the removal of soft material while preserving the hard core intact.
Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment
Huihan Liu, Yuke Zhu
Robotic IntelligenceReinforcement Learning from Human FeedbackConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningReinforcement LearningImageSequential
🎯 What it does: Propose the Sirius framework to achieve human-robot collaborative learning during deployment, using human intervention signals to weight training samples, thereby improving autonomous strategy performance in continuous deployment and reducing human workload.
Robotic Skill Acquisition via Instruction Augmentation with Vision-Language Models
Ted Xiao (Robotics at Google), Jonathan Tompson (Robotics at Google)
Robotic IntelligenceLarge Language ModelVision Language ModelVision-Language-Action ModelContrastive LearningTextTime SeriesRetrieval-Augmented Generation
🎯 What it does: This paper proposes the DIAL method, which re-annotates unlabelled robot trajectories with visual-semantic labels using a pre-trained vision-language model (CLIP), and trains a language-conditioned behavior cloning policy with these augmented instructions.
Robotic Table Tennis: A Case Study into a High Speed Learning System
David B D'Ambrosio, Laura Graesser (Google DeepMind)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningImageTime Series
🎯 What it does: This paper constructs and evaluates a high-speed learning table tennis robot system operable in real environments, covering a complete technology stack including perception, low-latency control, simulation, zero-copy transfer, and automated environment reset;
Robust and Versatile Bipedal Jumping Control through Reinforcement Learning
Zhongyu Li (University of California, Berkeley), Koushil Sreenath (University of California, Berkeley)
Robotic IntelligenceConvolutional Neural NetworkReinforcement Learning
🎯 What it does: Developed a control strategy that enables the real Cassie bipedal robot to perform multi-objective jumps (forward, backward, lateral, turning, elevation) and achieve precise landing.
Robust Safety under Stochastic Uncertainty with Discrete-Time Control Barrier Functions
Ryan Cosner, Aaron Ames
Optimization
🎯 What it does: The paper derives a probabilistic upper bound on the system's safety within finite time under unbounded random disturbances using discrete-time control barrier functions (DTCBF) and Markov inequality, and designs a safety controller considering expected constraints based on this.
ROSE: Rotation-based Squeezing Robotic Gripper toward Universal Handling of Objects
Son Tien Bui (Japan Advanced Institute of Science and Technology), Van Anh Ho (Japan Advanced Institute of Science and Technology)
Robotic Intelligence
🎯 What it does: This paper designs a soft gripper named ROSE based on rotation-induced membrane buckling, which can achieve gentle grasping of various shapes, materials, and even slippery objects through a single rotation, and demonstrates its high gripping force, load ratio, and durability.
Rotating without Seeing: Towards In-hand Dexterity through Touch
Zhao-Heng Yin (Hong Kong University of Science and Technology), Xiaolong Wang (University of California San Diego)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningTime Series
🎯 What it does: This paper proposes and implements a system named Touch Dexterity, which enables a multi-fingered hand to rotate objects in a visually impaired environment using only tactile feedback.
RT-1: Robotics Transformer for Real-World Control at Scale
Anthony Brohan (Robotics at Google), Brianna Zitkovich (Robotics at Google)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelTextMultimodality
🎯 What it does: Trained and evaluated a large-scale Transformer model RT-1 capable of handling over 700 natural language instructions and achieving multi-task operations on real robots.
SAM-RL: Sensing-Aware Model-Based Reinforcement Learning via Differentiable Physics-Based Simulation and Rendering
Jun Lv (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement LearningWorld ModelImagePhysics Related
🎯 What it does: Propose a perception-aware model-driven reinforcement learning system (SAM-RL) based on differentiable physics simulation and rendering, which can automatically update the environment model, actively select the most informative camera viewpoints, and learn policies in simulation, then map them to real robots through residual policy mapping.
Sampling-based Exploration for Reinforcement Learning of Dexterous Manipulation
Gagan Khandate (Columbia University), Matei Ciocarlie (Columbia University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes a method that utilizes sampling expansion planning (RRT) to generate a training reset distribution, enhancing the exploration efficiency of reinforcement learning in multi-fingered hand grasping. This enables precise in-hand rotation operations without external support, relying solely on tactile and proprioceptive feedback.
SAR: Generalization of Physiological Dexterity via Synergistic Action Representation
Cameron H Berg (Meta AI), Vikash Kumar (Meta AI)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This study proposes and verifies a reinforcement learning framework based on muscle synergies representation (SAR), utilizing a muscle-driven hand model to achieve complex grasping and reorientation in high-dimensional continuous control tasks, and validates it in scenarios such as robotic hands and whole-body locomotion.
Scaling Robot Learning with Semantically Imagined Experience
Tianhe Yu (Google), Fei Xia (Google)
Data SynthesisRobotic IntelligenceTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: Leverage text-to-image diffusion models to perform semantic-level image augmentation on existing robotic manipulation data, generating new objects, backgrounds, and distractions to achieve unlabeled expansion for tasks, scenes, and success detection.
Self-Supervised Lidar Place Recognition in Overhead Imagery Using Unpaired Data
Tim Y. Tang (University of Oxford), Paul M Newman (University of Oxford)
RecognitionDomain AdaptationAutonomous DrivingGenerative Adversarial NetworkContrastive LearningSimultaneous Localization and MappingImageMultimodalityPoint Cloud
🎯 What it does: Propose a self-supervised place recognition framework for LiDAR and satellite images without paired data, leveraging generated synthetic LiDAR, de-aliasing similarity matrices, and sequence alignment to automatically mine pseudo pairs for metric learning.
Self-Supervised Unseen Object Instance Segmentation via Long-Term Robot Interaction
Yangxiao Lu (University of Texas at Dallas), Yu Xiang (University of Texas at Dallas)
Object TrackingSegmentationRobotic IntelligenceOptical FlowImageVideoBenchmark
🎯 What it does: Propose a self-supervised method combining long-term robot pushing interactions, optical flow-based multi-object tracking, and video object segmentation to generate instance segmentation masks for unseen objects and fine-tune existing models.
Self-Supervised Visuo-Tactile Pretraining to Locate and Follow Garment Features
Justin Kerr (UC Berkeley), Ken Goldberg (UC Berkeley)
Object TrackingAnomaly DetectionRepresentation LearningConvolutional Neural NetworkContrastive LearningMultimodality
🎯 What it does: Proposed a self-supervised visual-tactile pre-training framework called SSVTP for learning multi-task visual-tactile representations and achieving clothing feature localization and tracking.