RSS 2025 Papers with AI Summaries
Robotics: Science and Systems Β· 163 papers
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 163 RSS 2025 papers by keyword, author or institution β
free trial on arXivSub.
A Biconvex Method for Minimum-Time Motion Planning Through Sequences of Convex Sets
Tobia Marcucci, Andrew Marchese
Optimization
π― What it does: Proposed a Biconvex method (SCS) based on bi-convex decomposition to solve the shortest time trajectory problem given a series of convex sets and velocity/acceleration constraints;
A Generic Continuous Multi-Joint Spinal Robotic System for Agile and Accurate Behaviors with GNN-MPC method
Ying Wu, Hui Cheng
Robotic IntelligenceGraph Neural NetworkTabular
π― What it does: Proposed a multi-joint continuous spinal robot (SGb-SMMS) based on ball-tooth pairs and designed a corresponding GNN-MPC control framework, verifying its performance in multiple scenarios such as load capacity, precision, flexibility, and aerial flips.
A low-cost and lightweight 6 DoF bimanual arm for dynamic and contact-rich manipulation
Jaehyung Kim, Beomjoon Kim
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement Learning
π― What it does: Developed a low-cost, lightweight 6-DOF dual-arm robot named ARMADA, capable of achieving high-speed dynamic manipulation with rich contact interactions.
A Probabilistic Measure of Multi-Robot Connectivity and Ergodic Optimal Control
Yongce Liu, Zhongqiang Ren
OptimizationRobotic Intelligence
π― What it does: This paper proposes a probabilistic connectivity metric based on robot trajectories, integrating probabilistic connectivity constraints with ergodic search objectives to formulate an optimal control problem for multi-robot connectivity maintenance, and provides corresponding theoretical optimality conditions. The numerical solution to this optimal control problem is achieved through a combination of the augmented Lagrangian method and iterative LQR (iLQR).
A Robot-Assisted Approach to Small Talk Training for Adults with ASD
Rebecca Ramnauth, Brian Scassellati
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVideoTextAudio
π― What it does: Develop and evaluate an autonomous robot in a home environment for training small talk communication skills in adults with ASD.
A Unified and General Humanoid Whole-Body Controller for Fine-Grained Locomotion
Yufei Xue, Jiangmiao Pang
Domain AdaptationRobotic IntelligenceReinforcement Learning
π― What it does: This paper proposes a unified whole-body controller, HuGWBC, which can achieve multi-parameter customizable gaits under different locomotion modes (walking, jumping, standing, leaping);
Action Flow Matching for Lifelong Learning
Alejandro Murillo-GonzΓ‘lez, Lantao Liu
Robotic IntelligenceFlow-based ModelTime SeriesSequentialOrdinary Differential Equation
π― What it does: Propose an action transformation method based on flow matching (Action Flow Matching, AFM), enabling robots to map planned actions to more realistic executions during online continuous learning, thereby continuously correcting dynamic models with significant errors.
Adaptive Locomotion on Mud through Proprioceptive Sensing of Substrate Properties
Shipeng Liu, Feifei Qian
Robotic IntelligenceTime Series
π― What it does: This paper studies a method for online estimating mud terrain properties based on self-perception (joint current/position) and adjusting the locomotion strategy of a fin-robot to avoid slipping or pulling failure on mud with different water contents.
AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body Control
Jialong Li, Xiaolong Wang
Domain AdaptationOptimizationRobotic IntelligenceTransformerReinforcement Learning
π― What it does: Propose the AMO framework, combining sim-to-real reinforcement learning with trajectory optimization to achieve real-time, adaptive whole-body control.
APEX-MR: Multi-Robot Asynchronous Planning and Execution for Cooperative Assembly
Philip Huang, Jiaoyang Li
OptimizationRobotic Intelligence
π― What it does: Propose a multi-robot asynchronous planning and execution framework APEX-MR for collaborative assembly tasks, particularly LEGO structures;
ArticuBot: Learning Universal Articulated Object Manipulation Policy via Large Scale Simulation
Yufei Wang, David Held
Robotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningDiffusion modelPoint Cloud
π― What it does: Propose ArticuBot, a visual control strategy generated through hierarchical imitation learning using 42.3k demonstrations in large-scale physics simulations, which achieves zero-shot transfer to different laboratory Franka and mobile X-Arm robots, successfully opening multiple unseen movable objects (e.g., cabinet doors, drawers, microwaves, ovens, etc.).
ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills
Tairan He, Guanya Shi
Data SynthesisDomain AdaptationRobotic IntelligenceReinforcement LearningVideoTime SeriesPhysics Related
π― What it does: Propose a two-stage ASAP framework: first, train a motion tracking strategy using human video remapped motion in simulation, then collect real-world data to train a delta action model to compensate for simulation-to-real dynamics differences, followed by fine-tuning the strategy in simulation and directly deploying it on a robot.
ASTRID: A Robotic Tutor for Nurse Training to Reduce Healthcare-Associated Infections
Peizhu Qian, Vaibhav V. Unhelkar
Pose EstimationRobotic IntelligenceVision-Language-Action ModelVideoTextMultimodality
π― What it does: This paper constructs and evaluates a robot tutoring system named ASTRID through participatory design, aimed at training nurses to replace central venous catheters (CLD) to reduce infection-related complications;
BeamDojo: Learning Agile Humanoid Locomotion on Sparse Footholds
Huayi Wang, Jiangmiao Pang
Robotic IntelligenceReinforcement LearningPoint Cloud
π― What it does: Proposed the BEAMDOJO framework, achieving agile walking and robustness for humanoid robots on sparse foothold terrains.
Behavior Synthesis via Contact-Aware Fisher Information Maximization
Hrishikesh Sathyanarayan, Ian Abraham
OptimizationRobotic IntelligenceTabularPhysics Related
π― What it does: This paper proposes to automatically synthesize robot behaviors by maximizing contact-aware Fisher information, thereby achieving efficient physical parameter learning.
Bilevel Learning for Bilevel Planning
Bowen Li, Alexander G. Gray
OptimizationRobotic IntelligenceSupervised Fine-TuningReinforcement LearningImagePoint Cloud
π― What it does: Utilizing offline demonstration data, automatically invent neural predicates and construct a neural symbolic bilevel planning framework (IVNTR), achieving zero-shot compositional generalization and long-horizon mobile robot planning.
Boxi: Design Decisions in the Context of Algorithmic Performance for Robotics
Jonas Frey, Marco Hutter
Robotic IntelligenceSimultaneous Localization and MappingMultimodalityBenchmark
π― What it does: This paper designs, implements, and evaluates the multimodal perception payload Boxi for mobile robots (such as quadruped robots), and provides a complete hardware/software open-source package and design cookbook.
Bridging Model Predictive Control and Deep Learning for Scalable Reachability Analysis
Zeyuan Feng, Somil Bansal
OptimizationSafty and PrivacyPhysics Related
π― What it does: This paper proposes a hybrid training framework that combines the approximate reachability data generated by sampling-based model predictive control (MPC) with the physics-constrained deep neural network (DeepReach), for reachability analysis and safety control of high-dimensional systems.
Bridging Perception and Action: Spatially-Grounded Mid-Level Representations for Robot Generalization
Jonathan Heewon Yang, Tingnan Zhang
Robotic IntelligenceMixture of ExpertsDiffusion model
π― What it does: This paper investigates the use of spatial perception-based mid-level representations (object-centric, motion-centric, pose-aware, depth-aware) in dual-arm dexterous manipulation to enhance the generalization performance of robotic policies, and proposes a diffusion strategy based on a hybrid expert (Mid-Level MoE) to dynamically integrate these representations, supplemented by weighted self-consistency training;
Bridging the Sim-to-Real Gap for Athletic Loco-Manipulation
Nolan Fey, Pulkit Agrawal
Domain AdaptationRobotic IntelligenceReinforcement LearningTime Series
π― What it does: Achieved sim-to-real transfer for executing high-difficulty motor manipulation tasks (throwing, weightlifting, dragging) on real hardware by constructing an unsupervised actuator network (UAN) and a two-stage training process (pre-training a motion controller, followed by fine-tuning with task rewards).
Building Rome with Convex Optimization
Haoyu Han, Heng Yang
Pose EstimationDepth EstimationOptimizationComputational EfficiencySimultaneous Localization and MappingImage
π― What it does: Built a pipeline using convex optimization for structure from motion (SfM) and bundle adjustment (BA), proposed a scalable convex bundle adjustment (SBA) framework, and implemented a GPU-accelerated solver named XM.
Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies
Chen Xu, Masha Itkina
Anomaly DetectionRobotic IntelligenceDiffusion modelFlow-based Model
π― What it does: This paper proposes FAIL-Detect, a two-stage, uncertainty estimation-based runtime failure detection framework specifically designed for generative imitation learning robot control policies.
Certifiably-Correct Mapping for Safe Navigation Despite Odometry Drift
Devansh R. Agrawal, Dimitra Panagou
Autonomous DrivingSimultaneous Localization and MappingMesh
π― What it does: Propose a framework that ensures map safety under visual-inertial odometry drift by shrinking safe regions to guarantee that the mapped free space never includes obstacles.
CLIP-RT: Learning Language-Conditioned Robotic Policies from Natural Language Supervision
Gi-Cheon Kang, Byoung-Tak Zhang
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes CLIP-RT, a vision-language-action (VLA) model that learns language-conditioned robot control policies directly using natural language supervision, and designs a language-based teleoperation data collection framework accessible to non-experts;
CodeDiffuser: Attention-Enhanced Diffusion Policy via VLM-Generated Code for Instruction Ambiguity
Guang Yin, Yunzhu Li
Explainability and InterpretabilityRobotic IntelligenceVision Language ModelDiffusion modelTextPoint Cloud
π― What it does: This paper proposes the CodeDiffuser framework, which generates 3D attention maps by using VLM to produce executable code, addressing language ambiguity in robotic manipulation tasks.
Coherence-based Approximate Derivatives via Web of Affine Spaces Optimization
Daniel Rakita, Qian Wang
OptimizationComputational Efficiency
π― What it does: This paper proposes a coherence-based approximate derivative computation method called Web of Affine Spaces Optimization (WASP), which rapidly obtains derivatives of function sequences by constructing an 'affine space web' and using closed-form KKT solutions at each step, significantly reducing the number of forward function calls.
Collaborative Object Transportation in Space via Impact Interactions
Joris Verhagen, Jana Tumova
OptimizationRobotic IntelligencePhysics Related
π― What it does: A multi-robot collaborative transportation framework based on collision interaction is proposed in zero-gravity environments, utilizing discrete impact models and spatial temporal logic (STL) specifications for trajectory planning and real-time control.
Complementarity-Free Multi-Contact Modeling and Optimization for Dexterous Manipulation
Wanxin Jin
OptimizationRobotic Intelligence
π― What it does: Propose a multi-contact model without complementary constraints and apply it to real-time multi-contact planning and control;
ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy
Yuhui Chen, Dongbin Zhao
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageMultimodality
π― What it does: For pre-trained vision-language-action (VLA) models, a two-stage reinforcement learning fine-tuning approach is implemented in real robot manipulation: first, offline Cal-ConRFT fine-tuning is performed using a small amount of demonstration data, followed by rapid convergence through online human-in-the-loop (HIL-ConRFT) interaction.
CordViP: Correspondence-based Visuomotor Policy for Dexterous Manipulation in Real-World
Yankai Fu, Shanghang Zhang
Pose EstimationRobotic IntelligenceDiffusion modelImagePoint Cloud
π― What it does: Proposed the CordViP framework, which constructs an interactive perception 3D point cloud by combining 6D object pose estimation with robot self-perception. It learns the correspondence between hands and objects and arm collaboration information through pre-training, and finally trains a visual motion strategy based on a diffusion model to achieve various real-world flexible manipulation tasks.
CREStE: Scalable Mapless Navigation with Internet Scale Priors and Counterfactual Guidance
Arthur Zhang, Amy Zhang
Autonomous DrivingRobotic IntelligenceTransformerReinforcement LearningContrastive LearningImagePoint CloudBenchmark
π― What it does: CRESTE proposes an scalable mapless navigation framework that learns structured bird's-eye view (BEV) features through distillation from visual foundation models, and achieves expert-aligned rewards by combining counterfactual inverse reinforcement learning with active reward learning, enabling safe and efficient long-range navigation in open-world urban environments.
Curating Demonstrations using Online Experience
Annie S. Chen, Chelsea Finn
Robotic IntelligenceTransformerDiffusion modelSequential
π― What it does: Propose a method called DemoβSCORE for automatically filtering demonstration data based on robot online experience. The method first trains an initial policy, generates rollβout, and trains a success/failure discriminator. Then, the discriminator is applied to the original demonstrations to eliminate unreliable strategies, thereby improving the final policy performance.
DDAT: Diffusion Policies Enforcing Dynamically Admissible Robot Trajectories
Jean-Baptiste Bouvier, Negar Mehr
OptimizationRobotic IntelligenceTransformerDiffusion modelSequential
π― What it does: Propose the DDAT framework that combines diffusion models with feasible trajectory projection to generate trajectories in real-time that satisfy robot dynamics constraints.
Debiasing 6-DOF IMU via Hierarchical Learning of Continuous Bias Dynamics
Ben Liu, Maani Ghaffari
Pose EstimationTime SeriesSequentialOrdinary Differential Equation
π― What it does: Utilizes the Neural Ordinary Differential Equation (NODE) framework for continuous dynamic modeling of IMU biases, and employs a hierarchical training strategy to first calibrate the gyroscope and then the accelerometer, achieving online bias removal;
DemoGen: Synthetic Demonstration Generation for Data-Efficient Visuomotor Policy Learning
Zhengrong Xue, Huazhe Xu
Data SynthesisRobotic IntelligenceTransformerDiffusion modelImagePoint Cloud
π― What it does: Propose DemoGen, a fully synthetic demonstration generation framework designed to significantly enhance the spatial generalization performance of visual motion policies across different object configurations.
Demonstrating a Control Framework for Physical Human-Robot Interaction Toward Industrial Applications
Bastien Muraccioli, Gentiane Venture
OptimizationRobotic IntelligenceTime Series
π― What it does: This paper proposes and implements an industrial-level physical human-robot interaction (pHRI) control framework, supporting three compliance control modes based on torque (whole-body compliance, null-space compliance, and dual compliance), and embeds safety constraints (torque, velocity, collision) into a second-order QP, achieving real-time mode switching, strict safety guarantees, and high-precision trajectory tracking.
Demonstrating Arena 5.0: A Photorealistic ROS2 Simulation Framework for Developing and Benchmarking Social Navigation
Linh KΓ€stner, Maximilian Ho-Kyoung Schreff
Robotic IntelligenceGraph Neural NetworkLarge Language ModelDiffusion modelMultimodalityBenchmark
π― What it does: Developed Arena 5.0, an integrated ROS 2 simulation framework incorporating NVIDIA Isaac Gym, for developing, evaluating, and benchmarking social navigation algorithms;
Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot
Yufeng Chi, Koushil Sreenath
Robotic IntelligenceReinforcement Learning
π― What it does: This paper designs and implements an open-source, low-cost, customizable medium-sized humanoid robot called Berkeley Humanoid Lite, and verifies seamless simulation-to-hardware bipedal walking control based on reinforcement learning, as well as hands-free bimanual control using VR + headset in headless mode on this platform.
Demonstrating CavePI: Autonomous Exploration of Underwater Caves by Semantic Guidance
Alankrit Gupta, Md Jahidul Islam
SegmentationRobotic IntelligenceConvolutional Neural NetworkSimultaneous Localization and MappingImage
π― What it does: This paper proposes and implements a lightweight autonomous underwater cave exploration AUV named CavePI, capable of achieving autonomous navigation and mapping in cave environments with GPS failure and low visibility through semantic guidance;
Demonstrating DVS: Dynamic Virtual-Real Simulation Platform for Mobile Robotic Tasks
Zijie Zheng, Long Zeng
Domain AdaptationRobotic IntelligenceTransformerSimultaneous Localization and MappingImageVideoMultimodalityPoint Cloud
π― What it does: Proposed a dynamic virtual-physical synchronization simulation platform named DVS, supporting multi-robot and multi-human dynamic scene generation, achieving bidirectional synchronization between virtual and physical environments through high-precision motion capture and ROS2, further enabling closed-loop training and real robot deployment.
Demonstrating GPU Parallelized Robot Simulation and Rendering for Generalizable Embodied AI with ManiSkill3
Stone Tao, Hao Su
Data SynthesisComputational EfficiencyRobotic IntelligenceReinforcement LearningDiffusion modelImageVideoMesh
π― What it does: Propose ManiSkill3, an open-source framework supporting GPU parallel simulation and rendering, offering 12 task categories, 20+ robot models, and integrated reinforcement learning and demonstration learning baselines.
Demonstrating LEAP Hand v2: Low-Cost, Easy-to-Assemble, High-Performance Hand for Robot Learning
Kenneth Shaw, Deepak Pathak
Robotic IntelligenceReinforcement LearningVideo
π― What it does: Developed a low-cost (around $200) dual-jointed soft-rigid hybrid structure robotic hand β LEAP Hand v2 β that can be assembled within one hour, specifically designed for robot learning research;
Demonstrating MOSART: Opening Articulated Structures in the Real World
Arjun Gupta, Saurabh Gupta
Depth EstimationOptimizationRobotic IntelligenceConvolutional Neural NetworkImage
π― What it does: Developed a modular mobile manipulator system named MOSART for zero-shot opening of door, drawers, and ovens, etc., which are hinged structures in unknown environments.
Demonstrating MuJoCo Playground
Kevin Zakka, Pieter Abbeel
Robotic IntelligenceReinforcement LearningImageTabular
π― What it does: Developed the MuJoCo Playground framework, an open-source simulation environment that implements MJX and Madrona-based batch rendering, supporting training for various robots (quadruped, humanoid, hand, robotic arm) and achieving zero-shot sim-to-real transfer.
Demonstrating Multi-Suction Item Picking at Scale via Multi-Modal Learning of Pick Success
Che Wang, Kapil Katyal
Object DetectionSegmentationDomain AdaptationRobotic IntelligenceTransformerSupervised Fine-TuningAuto EncoderImageMultimodality
π― What it does: In real warehouse environments, a model that automatically learns the success rate of multi-suction grasping was constructed by utilizing multimodal visual encoding (RGB, depth, semantic segmentation) and cross-attention mechanisms.
Demonstrating REASSEMBLE: A Multimodal Dataset for Contact-rich Robotic Assembly and Disassembly
Daniel Sliwowski, Dongheui Lee
Robotic IntelligenceDiffusion modelVideoMultimodalityBenchmarkAudio
π― What it does: This paper proposes and publicly releases the REASSEMBLE dataset, focusing on long-cycle, contact-intensive assembly and disassembly tasks, covering four core actions from NIST Task Board #1 (grasping, inserting, removing, placing).
Demonstrating Shared Force-Language Embeddings for Natural Human-Robot Communication
Ravi Tejwani, Haruhiko Asada
Representation LearningRobotic IntelligenceAuto EncoderContrastive LearningMultimodality
π― What it does: This paper proposes a cross-modal force-language embedding framework that enables bidirectional translation between force curves and natural language phrases in a shared latent space.
Demonstrating the Octopi-1.5 Visual-Tactile-Language Model
Samson Yu, Harold Soh
Representation LearningRobotic IntelligenceTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityRetrieval-Augmented Generation
π― What it does: Developed the Octopi-1.5 visual-tactile-language model and demonstrated its ability to recognize and reason about tactile inputs on the TMI gripper.
Demonstrating ViSafe: Vision-enabled Safety for High-speed Detect and Avoid
Parv Kapoor, Sebastian Scherer
Object DetectionObject TrackingAutonomous DrivingVideo
π― What it does: This paper proposes and implements a fully vision-based, low-cost high-speed drone collision avoidance system called ViSafe, and verifies its safety performance in simulation and real-world environments.
Dex1B: Learning with 1B Demonstrations for Dexterous Manipulation
Jianglong Ye, Xiaolong Wang
GenerationData SynthesisOptimizationRobotic IntelligenceConvolutional Neural NetworkAuto EncoderPoint CloudSequential
π― What it does: Constructed the DexIB dataset with 1 billion trajectories, and proposed an iterative data generation pipeline (DexSimple) combining optimization and generative models for grasping and joint manipulation tasks.
Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy
Jiayi Chen, He Wang
Data SynthesisRobotic IntelligenceConvolutional Neural NetworkPoint Cloud
π― What it does: Propose an efficient pipeline starting from a single human-annotated grasp template, capable of generating collision-free, contact-rich, and physically feasible grasp poses for any grasp type, object, and robotic hand, and construct a near ten-million-scale grasp dataset covering 31 GRASP grasp types.
DexterityGen: Foundation Controller for Unprecedented Dexterity
Zhao-Heng Yin, Mustafa Mukadam
Robotic IntelligenceReinforcement LearningDiffusion modelTime SeriesSequential
π― What it does: By pre-training diffusion generative models on large-scale simulation data, DexGen can refine coarse motion commands provided by humans or external policies into safe, precise robot finger actions, achieving unprecedented dexterous manipulation.
DexWild: Dexterous Human Interactions for In-the-Wild Robot Policies
Tony Tao, Deepak Pathak
Robotic IntelligenceTransformerDiffusion modelImageVideo
π― What it does: Developed the DexWild system, which can rapidly collect large-scale, diverse manipulation demonstrations through human operation, and jointly train with a small amount of robot demonstrations to generate finger-tip manipulation strategies that generalize across new environments, tasks, and robot body shapes.
Diffeomorphic Obstacle Avoidance for Contractive Dynamical Systems via Implicit Representations
Ken-Joel Simmoteit, Leonel Rozo
OptimizationRobotic IntelligenceSequentialOrdinary Differential Equation
π― What it does: Proposes a differential homeomorphic transformation framework based on signed distance fields (SDF) to achieve whole-body obstacle avoidance while maintaining convergence, enabling robots to safely and robustly execute skills learned by neural convergent dynamics systems (NCDS) in complex environments.
Differentiable GPU-Parallelized Task and Motion Planning
William Shen, Fabio Ramos
OptimizationComputational EfficiencyRobotic Intelligence
π― What it does: Proposes cuTAMP, a GPU-based parallel Task and Motion Planning (TAMP) framework that simultaneously performs differentiable constraint satisfaction and gradient optimization across thousands of candidate parameter batches, enabling highly constrained long-range robotic manipulation problems to be solved within seconds.
Discrete-Time Hybrid Automata Learning: Legged Locomotion Meets Skateboarding
Hang Liu, Maani Ghaffari
Robotic IntelligenceReinforcement LearningAuto EncoderSequential
π― What it does: Proposed a discrete-time hybrid automaton learning (DHAL) framework applied to quadruped robot skateboard motion, achieving automatic identification and execution of discrete modes and continuous dynamics without requiring trajectory segmentation or event marking.
Distilling Contact Planning for Fast Trajectory Optimization in Robot Air Hockey
Julius Jankowski, Sylvain Calinon
OptimizationKnowledge DistillationRobotic IntelligenceTabular
π― What it does: In the robot hitting a ping-pong ball (air hockey) task, combining offline learning-based high-level contact planning with online MPC to achieve fast real-time trajectory optimization;
DOGlove: Dexterous Manipulation with a Low-Cost Open-Source Haptic Force Feedback Glove
Han Zhang, Huazhe Xu
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningPoint Cloud
π― What it does: This paper proposes and implements DOGlove, a low-cost, open-source 21-degree-of-freedom motion capture glove that integrates cable-driven force feedback and linear resonant actuators, enabling data collection and evaluation for human-robot teleoperation, tactile perception, and imitation learning.
Doppler Correspondence: Non-Iterative Scan Matching With Doppler Velocity-Based Correspondence
Jiwoo Kim, Hyondong Oh
Pose EstimationSimultaneous Localization and MappingPoint Cloud
π― What it does: Leveraging the Doppler velocity information from 4D LiDAR/radar, a novel correspondence method (Doppler Correspondence) is proposed to achieve non-iterative point cloud registration, and it is compared with traditional methods such as ICP and DICP.
DRO: Doppler-Aware Direct Radar Odometry with Gyroscope
Cedric Le Gentil, Timothy Barfoot
Autonomous DrivingSimultaneous Localization and MappingImage
π― What it does: Proposes a Direct Radar Odometry (DRO) algorithm that directly registers millimeter-wave radar intensity images, utilizing Doppler information and continuous-time motion models to achieve radar odometry without feature extraction or point cloud generation.
Dynamic Rank Adjustment in Diffusion Policies for Efficient and Flexible Training
Xiatao Sun, Daniel Rakita
Computational EfficiencyRobotic IntelligenceConvolutional Neural NetworkDiffusion modelImage
π― What it does: This paper proposes a dynamic rank adjustment framework called DRIFT, which can change the rank of trainable parameters in real-time during training of diffusion strategies, thereby improving training efficiency while maintaining the representational power of over-parameterization.
Dynamic Safety in Complex Environments: Synthesizing Safety Filters with Poissonβs Equation
Gilbert Bahati, Aaron Ames
Robotic IntelligenceImage
π― What it does: Proposes a safety function generation method based on the Poisson equation and Dirichlet boundary conditions to construct a safety set from perceived occupancy maps, implementing it as a control barrier function (CBF) for collision avoidance in robots operating in complex dynamic environments.
Effective Sampling for Robot Motion Planning Through the Lens of Lattices
Itai Panasoff, Kiril Solovey
Robotic Intelligence
π― What it does: This paper proposes a lattice-based deterministic sampling method that generates a (0, Ξ΅)-complete sample set using Ad-lattice, thereby improving the reliability and efficiency of robot motion planning.
Efficient Hierarchical Any-Angle Path Planning on Multi-Resolution 3D Grids
Victor Reijgwart, Lionel Ott
Autonomous DrivingOptimizationComputational EfficiencyRobotic Intelligence
π― What it does: Propose a multi-resolution arbitrary-angle path planning framework called wavestar based on an octree, combining the arbitrary-angle advantage of Theta* and the efficiency of hierarchical search to quickly solve geometric shortest paths in 3D dense environments.
emg2tendon: From sEMG Signals to Tendon Control in Musculoskeletal Hands
Sagar Verma
Data SynthesisRobotic IntelligenceDiffusion modelTime SeriesSequentialBiomedical Data
π― What it does: This study extends the existing emg2pose dataset to generate the emg2tendon dataset, which includes sEMG to tendon control signals, and proposes a regression method based on the conditioned latent diffusion model (CLDM) for sEMG to tendon control signals. Subsequently, the gesture reconstruction performance is evaluated through a two-step regression (sEMGβtendonβhand joint angles) combined with physical simulation.
Enhancing Autonomous Driving Systems with On-Board Deployed Large Language Models
Nicolas Baumann, Luca Benini
Autonomous DrivingOptimizationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Propose a hybrid architecture combining a large language model (LLM) deployed locally on a vehicle with model predictive control (MPC), named DecisionxLLM and MPCxLLM, to achieve a natural language interactive autonomous driving system, with real-time inference implemented on Jetson Orin AGX.
FACTR: Force-Attending Curriculum Training for Contact-Rich Policy Learning
Jason Jingzhou Liu, Deepak Pathak
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelMultimodality
π― What it does: This paper designs a low-cost bidirectional force-sensing remote operation system and proposes the FACTR (Force-Attending Curriculum Training) method, training a multi-modal policy based on Transformer to achieve higher success rates and better generalization to unseen objects in various contact-rich tasks (box lifting, object rotation, fruit picking and placing, dough rolling);
FAST: Efficient Action Tokenization for Vision-Language-Action Models
Karl Pertsch, Sergey Levine
CompressionComputational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelMultimodalitySequential
π― What it does: Propose FAST, an action tokenizer based on Discrete Cosine Transform (DCT) and Byte Pair Encoding (BPE), to address the action tokenization problem in high-frequency robot control, and build a general-purpose version FAST+ that can be directly applied to any robot action sequence, thereby enabling efficient training of Vision-Language-Action (VLA) models.
FEAST: A Flexible Mealtime-Assistance System Towards In-the-Wild Personalization
Rajat Kumar Jenamani, Tapomayukh Bhattacharjee
Safty and PrivacyRobotic IntelligenceTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
π― What it does: Developed and evaluated FEAST, a dietary assistance robot system capable of achieving personalization in the wild, in real home environments;
FERMI: Flexible Radio Mapping with a Hybrid Propagation Model and Scalable Autonomous Data Collection
Yiming Luo, Boyu Zhou
OptimizationRobotic IntelligenceNeural Radiance FieldPoint CloudTabularPhysics Related
π― What it does: Proposed and implemented the FERMI framework for constructing radio maps in complex obstacle environments using a hybrid physical-neural network model, and achieved efficient signal strength prediction under sparse training through multi-robot autonomous data collection.
Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success
Moo Jin Kim, Percy Liang
Computational EfficiencyRobotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelMultimodality
π― What it does: For fine-tuning methods of vision-language-action models, this paper proposes an optimized fine-tuning strategy based on parallel decoding, action chunking, continuous action representation, and L1 regression objectives, significantly improving execution speed and success rate.
Flow Matching Ergodic Coverage
Max Muchen Sun, Todd Murphey
OptimizationRobotic IntelligenceFlow-based ModelMultimodalitySequential
π― What it does: This paper proposes combining Flow Matching with the empirical distribution of trajectories from observable systems to construct a new executable form of Ergodic Coverage;
Flying Hand: End-Effector-Centric Framework for Versatile Aerial Manipulation Teleoperation and Policy Learning
Guanqi He, Guanya Shi
OptimizationRobotic IntelligenceTransformerReinforcement LearningImagePoint Cloud
π― What it does: Proposes a multi-task operation framework based on a fully aerial dynamic drone equipped with a four-degree-of-freedom robotic arm, incorporating global MPC, online adaptive control, teleoperation, and imitation learning modules.
From Foresight to Forethought: VLM-In-the-Loop Policy Steering via Latent Alignment
Yilin Wu, Andrea Bajcsy
Robotic IntelligenceLarge Language ModelVision Language ModelWorld ModelMultimodality
π― What it does: Propose a VLM-in-the-Loop multi-modal generation robot strategy scheduling framework named FOREWARN, which predicts action outcomes through a world model during runtime and uses a VLM for semantic alignment and verification to select the most task-compliant low-level action plan.
Gain Tuning Is Not What You Need: Reward Gain Adaptation for Constrained Locomotion Learning
Arthicha Srisuchinnawong, Poramate Manoonpong
OptimizationRobotic IntelligenceReinforcement Learning
π― What it does: This paper proposes ROGER, a reward weight adaptation method based on body interaction, for robotic gait learning in constrained environments. It can real-time adjust the weights of rewards and penalties during learning to ensure constraint satisfaction while maximizing the primary reward.
Gait-Net-augmented Implicit Kino-dynamic MPC for Dynamic Variable-frequency Humanoid Locomotion over Discrete Terrains
Junheng Li, Quan Nguyen
OptimizationRobotic IntelligenceTabularTime Series
π― What it does: Proposed a Gait-Net-enhanced implicit dynamics MPC to enable dynamic gait planning and control for variable-frequency humanoid robots on discrete terrains.
GauSS-MI: Gaussian Splatting Shannon Mutual Information for Active 3D Reconstruction
Yuhan Xie, Jia Pan
Depth EstimationGaussian SplattingMultimodality
π― What it does: Proposed an active 3D reconstruction system called GauSS-MI based on 3D Gaussian Splatting, which uses a probabilistic model to quantify the uncertainty of each Gaussian projection and evaluates the next best view in real-time based on Shannon mutual information.
Generalizing Safety Beyond Collision-Avoidance via Latent-Space Reachability Analysis
Kensuke Nakamura, Andrea Bajcsy
Robotic IntelligenceRecurrent Neural NetworkReinforcement LearningWorld ModelImageSequential
π― What it does: Investigated a Hamilton-Jacobi (HJ) reachability analysis method based on the latent space of a generative world model, proposed a Latent Safety Filter capable of achieving safety under high-dimensional observations, and applied it to navigation, visual manipulation, and industrial robotics.
GeoDEx: A Unified Geometric Framework for Tactile Dexterous and Extrinsic Manipulation under Force Uncertainty
Sirui Chen, Rana Soltani Zarrin
OptimizationRobotic Intelligence
π― What it does: Propose a unified geometric framework named GeoDEx, which utilizes noisy measurements from tactile sensors to achieve force estimation, planning, and control under multi-point contact, enabling dexterous grasping and external manipulation tasks.
Geometric Gait Optimization for Kinodynamic Systems Using a Lie Group Integrator
Yanhao Yang, Ross Hatton
OptimizationRobotic Intelligence
π― What it does: A variational gait optimization and motion planning framework based on Lie group integrators is proposed, capable of handling 'kinodynamic' systems with both kinematic and dynamic characteristics. Complete motion planning (starting, accelerating, steady-state, turning, terminating) is achieved on roller racers, skateboards, and swimming robots under moderate Reynolds numbers.
Global Contact-Rich Planning with Sparsity-Rich Semidefinite Relaxations
Shucheng Kang, Heng Yang
OptimizationRobotic Intelligence
π― What it does: Propose a method that leverages sparsity (correlated sparsity, term sparsity, and robot-specific sparsity) to construct a high-order but sparse semi-definite relaxation (Moment-SOS) for achieving global or near-global optimal solutions in contact-rich motion planning.
Gripper Pose and Object Pointflow as Interfaces for Robotic Bimanual Manipulation
Yuyin Yang, Jiangmiao Pang
Robotic IntelligenceTransformerDiffusion modelPoint Cloud
π― What it does: This paper proposes an end-to-end continuous control framework named PPI, which enhances the spatial localization and motion planning capabilities of dual-arm robotic manipulation using two interfaces: target gripper keypose and object pointflow.
Hierarchical and Modular Network on Non-prehensile Manipulation in General Environments
Yoonyoung Cho, Beomjoon Kim
Data SynthesisDomain AdaptationRobotic IntelligenceTransformerReinforcement LearningPoint Cloud
π― What it does: Propose a non-grasping manipulation framework based on a hierarchical modular network (HAMNET) and pre-trained contact-aware representation (UNICORN), capable of moving objects to target poses in various real and simulated environments.
Hierarchical Temporal Logic Task and Motion Planning for Multi-Robot Systems
Zhongqi Wei, Changliu Liu
OptimizationRobotic IntelligenceGraph
π― What it does: Investigated a multi-robot task and motion planning framework based on hierarchical temporal logic and convex set graphs, jointly completing task allocation and shortest path search;
HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit
Qingwei Ben, Jiangmiao Pang
Robotic IntelligenceReinforcement LearningImageTabular
π― What it does: Designed and implemented a semi-autonomous humanoid teleoperation system named HOMIE, integrating reinforcement learning-based whole-body motion control, isomorphic exoskeleton arms, and motion perception gloves to form a unified cockpit, enabling humanoid robots to perform multiple chassis + operational tasks in real environments.
How to Coordinate UAVs and UGVs for Efficient Mission Planning? Optimizing Energy-Constrained Cooperative Routing with a DRL Framework
Mohammad Safwan Mondal, Pranav Bhounsule
OptimizationRobotic IntelligenceTransformerReinforcement LearningTabular
π― What it does: Proposes a multi-UAV/UGV cooperative path planning framework based on deep reinforcement learning to address energy-constrained UAV-UGV collaborative tasks.
Human2LocoMan: Learning Versatile Quadrupedal Manipulation with Human Pretraining
Yaru Niu (Carnegie Mellon University), Ding Zhao (Bosch Center for AI)
Robotic IntelligenceTransformerSupervised Fine-TuningSequential
π― What it does: Collect human demonstrations and robot body trajectories using XR headsets and remote control systems, and train a modular cross-modal Transformer model (MXT) to enable quadruped robots to perform autonomous manipulation in various tasks (single-handed/double-handed grasping, shoe tidying, shovel scooping, ball pouring, etc.).
IMLE Policy: Fast and Sample Efficient Visuomotor Policy Learning via Implicit Maximum Likelihood Estimation
Krishan Rana, Niko Suenderhauf
Computational EfficiencyRobotic IntelligenceConvolutional Neural NetworkMultimodality
π― What it does: Proposed an IMLE-based one-step generative policy (IMLE Policy) that can learn multimodal action distributions with few demonstrations and achieve real-time inference.
Implicit Neural-Representation Learning for Elastic Deformable-Object Manipulations
Jeongho Ha, Daehyung Park
Representation LearningRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningAuto EncoderContrastive LearningPoint Cloud
π― What it does: This paper proposes an elastic deformable object manipulation method based on implicit neural representation (INR-DOM), which reconstructs the complete signed distance function (SDF) from partial observations to obtain a global, unobstructed state representation of objects such as elastic bands. On this basis, dense features directly usable for control are learned through joint fine-tuning with contrastive learning and reinforcement learning.
Influence of Static and Dynamic Downwash Interactions on Multi-Quadrotor Systems
Anoop Kiran, Kenneth Breuer
Autonomous DrivingRobotic IntelligenceOptical FlowVideoTabularTime SeriesPhysics Related
π― What it does: Systematic measurement and quantification of the effects of downwash generated by the upper airframe on the thrust, torque, and velocity field of the lower airframe during close-proximity flight of multirotor drones, particularly under two configurations: vertical stacking and horizontal offset, were conducted using a six-axis torque sensor and a particle image velocimetry (PIV) system in a windless indoor environment.
Interface-level Intent Inference for Environment-agnostic Robot Teleoperation Assistance
Larisa Y.c. Loke, Brenna Argall
ClassificationRobotic IntelligenceRecurrent Neural NetworkTime SeriesBiomedical Data
π― What it does: This paper proposes an environment-agnostic robotic teleoperation assistance system that improves sip/puff breathing interface control for a 7-degree-of-freedom robotic arm through interface-level intent inference.
Interruption Handling for Conversational Robots
Shiye Cao, Chien-Ming Huang
ClassificationRecognitionRobotic IntelligenceTransformerLarge Language ModelVideoTextMultimodalityAudio
π― What it does: Designed and implemented an interruption handling system for conversational robots based on interruption intent classification, integrated into LLM-driven social robots, and validated in decision-making and debate tasks.
Is Your Imitation Learning Policy Better than Mine? Policy Comparison with Near-Optimal Stopping
David Snyder, Haruki Nishimura
Robotic IntelligenceReinforcement LearningTabular
π― What it does: Proposed a policy comparison framework called STEP based on sequential statistical testing, designed for binary success/failure evaluation in simulation and real robot tasks, achieving near-optimal stopping even with limited sample sizes;
Joint State and Noise Covariance Estimation
Kasra Khosoussi, Iman Shames
OptimizationRobotic IntelligenceSimultaneous Localization and MappingGraphBenchmark
π― What it does: This paper proposes a framework for jointly estimating the robot's state (such as pose, odometry bias, etc.) and the covariance matrix of measurement noise, and presents a block coordinate descent (BCD) algorithm that can be embedded into existing nonlinear least squares solvers;
Kinodynamic Trajectory Following with STELA: Simultaneous Trajectory Estimation & Local Adaptation
Edgar Granados, Kostas Bekris
OptimizationRobotic IntelligenceSimultaneous Localization and Mapping
π― What it does: Propose a framework called STELA that simultaneously performs trajectory estimation and local adaptation. It generates initial trajectories using sample-based motion planning and, during online operation, simultaneously estimates past trajectories and adjusts control commands in real-time via a factor graph, achieving robust tracking against dynamic uncertainties, observation noise, and model errors.
LangWBC: Language-directed Humanoid Whole-Body Control via End-to-end Learning
Yiyang Shao, Koushil Sreenath
Domain AdaptationRobotic IntelligenceReinforcement LearningVision Language ModelAuto EncoderTextSequential
π― What it does: This study proposes an end-to-end language-driven humanoid robot full-body control framework that can directly map natural language instructions to executable continuous actions.
Learned Perceptive Forward Dynamics Model for Safe and Platform-aware Robotic Navigation
Pascal Roth, Marco Hutter
Robotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningWorld ModelPoint CloudTime Series
π― What it does: Proposed and implemented a perceptual forward dynamics model (FDM) capable of predicting the robot's future states based on surrounding terrain height scans and historical body information, integrating this model into a zero-shot MPPI planner to achieve safe navigation in complex terrains.
Learning Getting-Up Policies for Real-World Humanoid Robots
Xialin He, Saurabh Gupta
Robotic IntelligenceReinforcement Learning
π― What it does: Propose a two-stage reinforcement learning framework called HUMANUP, which learns control strategies for humanoid robots to stand up from different fallen postures (prone, supine) on various terrains (flat ground, slopes, grass, rocks, snow, etc.).
Learning Humanoid Standing-up Control across Diverse Postures
Tao Huang, Jiangmiao Pang
Robotic IntelligenceReinforcement Learning
π― What it does: This paper proposes a reinforcement learning framework called HoST, enabling humanoid robots to stand up from various initial poses and different terrains (flat ground, platform, wall, slope) and can be directly deployed on real robots.
Learning Interpretable Features from Interventions
Erin Hedlund-Botti, Matthew Craig Gombolay
Explainability and InterpretabilityRobotic IntelligenceTransformerReinforcement LearningSequential
π― What it does: Proposes the LIFI framework, which leverages user interventions during robot errors to learn interpretable features, thereby adjusting robot strategies to align with individual preferences.