CoRL 2025 Papers with AI Summaries
Conference on Robot Learning · 263 papers
→ CoRL 2025 papers with code (18)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 263 CoRL 2025 papers by keyword, author or institution —
free trial on arXivSub.
“Stack It Up!”: 3D Stable Structure Generation from 2D Hand-drawn Sketch
Yiqing Xu (National University of Singapore), David Hsu (National University of Singapore)
GenerationRobotic IntelligenceDiffusion modelImageMesh
🎯 What it does: StackItUp developed a system that generates stable 3D block structures from a single 2D hand-drawn sketch and enables robotic execution.
$\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization
Kevin Black (Physical Intelligence), Ury Zhilinsky (Physical Intelligence)
Robotic IntelligenceTransformerVision-Language-Action ModelFlow-based ModelMultimodality
🎯 What it does: Developed a VLA-based π0.5 model capable of performing long-term cleaning and organizing tasks in completely new homes
$\texttt{SPIN}$: distilling $\texttt{Skill-RRT}$ for long-horizon prehensile and non-prehensile manipulation
Haewon Jung (Korea Advanced Institute of Science and Technology), Beomjoon Kim (Korea Advanced Institute of Science and Technology)
Robotic IntelligenceReinforcement LearningDiffusion model
🎯 What it does: Proposed a method to extend RRT into a skill planner called Skill-RRT, and distilled it into a deployable policy through a connector and imitation learning, to address operational tasks requiring long-term pre-grasping and non-grasping sequences.
$Door(s)$: Junction State Estimation for Efficient Exploration in Reinforcement Learning
Benjamin Fele (Jožef Stefan Institute), Jan Babic (Jožef Stefan Institute)
Reinforcement Learning
🎯 What it does: Proposes a heuristic method for door state estimation (Doorfs) to promote efficient exploration in reinforcement learning by identifying narrow passages or joints in the state space.
3DS-VLA: A 3D Spatial-Aware Vision Language Action Model for Robust Multi-Task Manipulation
Xiaoqi Li (Peking University), Hao Dong (CUHK)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImagePoint Cloud
🎯 What it does: Propose the 3DS-VLA model, enhancing 3D spatial perception on pre-trained 2D vision-language models by enabling point clouds and images to share a visual encoder, and introducing 3D spatial constraints for multi-task, multi-modal, and environment-diverse reinforcement learning;
Action-Free Reasoning for Policy Generalization
Jaden Clark (Stanford University), Suneel Belkhale (Stanford University)
Domain AdaptationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose the RAD method, which uses language reasoning from human videos to train robot policies, obtaining reasoning chains from action-agnostic data to enhance cross-entity and cross-task generalization.
ActLoc: Learning to Localize on the Move via Active Viewpoint Selection
Jiajie Li (ETH Zurich), Marc Pollefeys (ETH Zurich)
OptimizationRobotic IntelligenceTransformerSimultaneous Localization and MappingImageMultimodalityPoint Cloud
🎯 What it does: Studied an attention-based model called ActLoc for robot navigation, aiming to improve visual localization accuracy through active viewpoint selection and integrating predictions into path planning.
Adapt3R: Adaptive 3D Scene Representation for Domain Transfer in Imitation Learning
Albert Wilcox (Georgia Institute of Technology), Animesh Garg (Georgia Institute of Technology)
Domain AdaptationConvolutional Neural NetworkReinforcement LearningVision Language ModelImagePoint Cloud
🎯 What it does: Propose Adapt3R, a method that projects semantic features from pre-trained 2D vision models into 3D point clouds, and generates a single vector through position encoding and attention pooling as an observation encoder for multi-task imitation learning, achieving zero-shot transfer for new camera viewpoints and robot structures.
Adapting by Analogy: OOD Generalization of Visuomotor Policies via Functional Correspondence
Pranay Gupta, Andrea Bajcsy
Domain AdaptationRobotic IntelligenceReinforcement Learning from Human FeedbackVision-Language-Action ModelDiffusion modelImageText
🎯 What it does: Proposes a deployment-time adaptation method (Adapting by Analogy, ABA) that reuses existing ID visual-motor policies in OOD scenarios through natural language functional correspondences provided by experts, achieving task success under unknown visual conditions.
AgentWorld: An Interactive Simulation Platform for Scene Construction and Mobile Robotic Manipulation
Yizheng Zhang (Tencent Robotics X), Lei Han (Tencent Robotics X)
Robotic IntelligenceDiffusion modelBenchmark
🎯 What it does: Propose AgentWorld, an integrated interactive simulation platform for automated scene construction (layout, semantic asset placement, material configuration, and physics simulation) as well as teleoperation data collection for mobile robots, generating large-scale home manipulation datasets.
Agreement Volatility: A Second-Order Metric for Uncertainty Quantification in Surgical Robot Learning
Jordan Thompson (University of Utah), Alan Kuntz (University of Utah)
Domain AdaptationExplainability and InterpretabilityRobotic IntelligenceConvolutional Neural NetworkPoint CloudBiomedical Data
🎯 What it does: Proposes Volatility-Aware DeformerNet (VAD-Net), achieving safe handover and uncertainty space visualization in soft tissue manipulation for surgical robots by introducing a second-order uncertainty measure (agreement volatility).
AimBot: A Simple Auxiliary Visual Cue to Enhance Spatial Awareness of Visuomotor Policies
Yinpei Dai (University of Michigan), Joyce Chai (University of Michigan)
Robotic IntelligenceSupervised Fine-TuningVision-Language-Action ModelImagePoint CloudBenchmark
🎯 What it does: Propose a lightweight visual enhancement method (AimBot), which overlays the shooting line and crosshair on multi-view RGB images, directly mapping the robot end-effector (EE) spatial pose to pixel space, thereby improving the spatial perception and control performance of visuomotor policies.
AirExo-2: Scaling up Generalizable Robotic Imitation Learning with Low-Cost Exoskeletons
Hongjie Fang (Shanghai Jiao Tong University), Hao-Shu Fang (Shanghai Jiao Tong University)
Image TranslationDepth EstimationRobotic IntelligenceConvolutional Neural NetworkTransformerVision-Language-Action ModelDiffusion modelImage
🎯 What it does: Developed a low-cost exoskeleton system, AirExo-2, for large-scale collection and conversion of human demonstrations in the wild (in-the-wild), and proposed the RISE-2 strategy, which can generalize learning by leveraging 2D semantic and 3D geometric information, ultimately achieving zero-shot deployment on real dual-arm robots without requiring robot data.
AnyPlace: Learning Generalizable Object Placement for Robot Manipulation
Yuchi Zhao (University of Toronto), Animesh Garg (Georgia Institute of Technology)
Pose EstimationRobotic IntelligenceTransformerVision Language ModelDiffusion modelPoint Cloud
🎯 What it does: Proposed a two-stage object placement method called AnyPlace, which first roughly localizes the placement position using a vision-language model, and then performs fine-grained placement on local point clouds with a low-level pose prediction model.
ARCH: Hierarchical Hybrid Learning for Long-Horizon Contact-Rich Robotic Assembly
Jiankai Sun (Stanford University), Hui Li (Autodesk Research)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelBenchmark
🎯 What it does: Propose a hierarchical hybrid learning framework ARCH, integrating a low-level skill library combining motion planning and reinforcement learning with a high-level policy based on imitation learning, achieving long-horizon, high-precision, and contact-rich robotic assembly.
Articulate AnyMesh: Open-vocabulary 3D Articulated Objects Modeling
Xiaowen Qiu (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
SegmentationGenerationPose EstimationLarge Language ModelVision Language ModelGaussian SplattingMesh
🎯 What it does: Designed and implemented the ARTICULATE ANYMESH framework, which can automatically convert any 3D mesh into controllable, open-vocabulary articulated 3D objects, supporting multiple object categories;
Articulated Object Estimation in the Wild
Abdelrhman Werby (University of Freiburg), Abhinav Valada (University of Freiburg)
Pose EstimationOptimizationSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: Propose the ArtiPoint framework, which estimates 3D motion models of articulated objects in outdoor scenes with dynamic camera motion, occlusions, and partial visibility through depth point tracking and factor graph optimization.
AT-Drone: Benchmarking Adaptive Teaming in Multi-Drone Pursuit
Yang Li (University Of Manchester), Wei Pan (University Of Manchester)
Robotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: Propose the AT-Drone benchmark, integrating configurable simulation, real drone deployment, distributed training framework, and standard evaluation for adaptive collaboration research in multi-drone pursuit.
ATK: Automatic Task-driven Keypoint Selection for Robust Policy Learning
Yunchu Zhang (University of Washington), Abhishek Gupta (University of Washington)
Pose EstimationDomain AdaptationKnowledge DistillationReinforcement LearningImagePoint Cloud
🎯 What it does: Propose a task-driven automatic keypoint selection method called ATK, which enhances the robustness and cross-domain transfer ability of visual control policies by minimizing the keypoint set.
AutoEval: Autonomous Evaluation of Generalist Robot Manipulation Policies in the Real World
Zhiyuan Zhou (University Of California Berkeley), Sergey Levine (University Of California Berkeley)
Robotic IntelligenceVision Language ModelImageVideoTextMultimodality
🎯 What it does: Developed the AutoEval system, achieving automated, continuous physical evaluation of general robotic manipulation policies, significantly reducing human effort.
BEHAVIOR Robot Suite: Streamlining Real-World Whole-Body Manipulation for Everyday Household Activities
Yunfan Jiang (Stanford University), Li Fei-Fei (Stanford University)
Robotic IntelligenceConvolutional Neural NetworkTransformerVision-Language-Action ModelDiffusion modelMultimodalityPoint Cloud
🎯 What it does: This paper proposes BEHAVIOR ROBOT SUITE (BRS), which includes a low-cost whole-body teleoperation interface JoyLo and a full-body visual-motor attention strategy WB-VIMA based on Transformers, to perform five complex daily tasks requiring dual-arm collaboration, precise navigation, and large end-effector movements in real home environments.
Belief-Conditioned One-Step Diffusion: Real-Time Trajectory Planning with Just-Enough Sensing
Gokul Puthumanaillam (University of Illinois Urbana-Champaign), Melkior Ornik (University of Illinois Urbana-Champaign)
Autonomous DrivingOptimizationReinforcement LearningDiffusion modelMultimodality
🎯 What it does: This paper proposes the Belief-Conditioned One-Step Diffusion (B-COD) framework, which simultaneously generates short-term trajectories and predicts local localization errors based on trajectory dispersion in a single forward pass of the diffusion model. This prediction serves as a risk metric to drive a constrained Soft Actor-Critic (SAC) policy to switch multimodal sensors online, achieving 'just-enough' perception by minimizing energy consumption while meeting the localization error budget.
BEVCalib: LiDAR-Camera Calibration via Geometry-Guided Bird’s-Eye View Representation
Weiduo Yuan (University of Southern California), Hang Qiu (University of California, Riverside)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkTransformerImageMultimodalityPoint Cloud
🎯 What it does: Developed BEVCALIB, a target-agnostic LiDAR-camera extrinsic calibration model that leverages bird's-eye-view (BEV) features to predict extrinsic parameters from raw data in one go.
Beyond Constant Parameters: Hyper Prediction Models and HyperMPC
Jan Węgrzynowski (Poznan University of Technology), Krzysztof Walas
Autonomous DrivingOptimizationRobotic IntelligenceRecurrent Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: Propose a Hyper Prediction Model (HyperPM) that predicts the time-varying trajectory of future dynamic model parameters using neural networks, and embeds it into Model Predictive Control (MPC) to form HyperMPC, thereby improving long-term prediction and control performance for partially observable systems.
Bipedal Balance Control with Whole-body Musculoskeletal Standing and Falling Simulations
Chengtian Ma (Tsinghua University), Yanan Sui (Tsinghua University)
Data SynthesisRobotic IntelligenceReinforcement LearningBiomedical Data
🎯 What it does: This study utilizes a full-body musculoskeletal model with 700 muscles to propose an unsupervised hierarchical control method (HBC), achieving high-fidelity simulation of human bipedal static standing and falling. It systematically collects and analyzes dynamic data on balance and falling under various physiological conditions, including healthy, injured, and exoskeleton-assisted scenarios.
BranchOut: Capturing Realistic Multimodality in Autonomous Driving Decisions
Hee Jae Kim (Boston University), Eshed Ohn-Bar (Boston University)
Data SynthesisAutonomous DrivingTransformerDiffusion modelNeural Radiance FieldMultimodalityBenchmark
🎯 What it does: Proposed BranchOut, a diffusion planner based on Gaussian Mixture Model (GMM), capable of generating multimodal driving trajectories from visual information, and constructed a human-computer interaction-based simulation environment to collect multimodal benchmark data.
Capability-Aware Shared Hypernetworks for Flexible Heterogeneous Multi-Robot Coordination
Kevin Fu (Georgia Institute of Technology), Harish Ravichandar (Georgia Institute of Technology)
Robotic IntelligenceRecurrent Neural NetworkReinforcement Learning
🎯 What it does: Designed a soft parameter sharing neural network architecture called CASH for heterogeneous multi-robot collaborative learning, which can dynamically generate decoder parameters based on robot capabilities, achieving diverse behaviors and zero-shot generalization.
CARE: Enhancing Safety of Visual Navigation through Collision Avoidance via Repulsive Estimation
Joonkyung Kim (Sogang University), Changjoo Nam (Carnegie Mellon University)
Depth EstimationRobotic IntelligenceImage
🎯 What it does: Propose the CARE module to enhance visual navigation safety
CaRL: Learning Scalable Planning Policies with Simple Rewards
Bernhard Jaeger (University of Tübingen), Andreas Geiger (University of Tübingen)
Autonomous DrivingReinforcement LearningBenchmark
🎯 What it does: Studied the scalability of using reinforcement learning (RL) for planning in autonomous driving, proposing the CaRL method that uses only a route completion reward, and trained and evaluated on CARLA and nuPlan.
CASPER: Inferring Diverse Intents for Assistive Teleoperation with Vision Language Models
Huihan Liu, Yuke Zhu (University Of Texas At Austin)
Robotic IntelligenceLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: Propose CASPER, an assisted remote manipulation system based on a vision-language model (VLM), capable of real-time inferring user intent in the background and autonomously executing multi-step mobile manipulation tasks after user confirmation.
CDP: Towards Robust Autoregressive Visuomotor Policy Learning via Causal Diffusion
Jiahua Ma (Sun Yat-sen University), Ruimao Zhang (Sun Yat-sen University)
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelImageSequential
🎯 What it does: Propose a Causal Diffusion Policy (CDP) based on causal transformers, which enhances temporal context by incorporating historical action sequences into action generation and reduces computational overhead in autoregressive inference through a cache-sharing mechanism.
CHD: Coupled Hierarchical Diffusion for Long-Horizon Tasks
Ce Hao (National University of Singapore), Harold Soh (National University of Singapore)
GenerationOptimizationRobotic IntelligenceDiffusion model
🎯 What it does: Proposes a Coupled Hierarchical Diffusion (CHD) framework that jointly generates high-level subgoals and low-level trajectories for long-horizon planning;
CLAMP: Crowdsourcing a LArge-scale in-the-wild haptic dataset with an open-source device for Multimodal robot Perception
Pranav N. Thakkar (Cornell University), Tapomayukh Bhattacharjee (Cornell University)
RecognitionRobotic IntelligenceConvolutional Neural NetworkTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposed the CLAMP device, CLAMP dataset, and CLAMP model for large-scale multimodal tactile data collection and material/texture identification to support robots' multi-body perception and manipulation in real-world environments.
CLASS: Contrastive Learning via Action Sequence Supervision for Robot Manipulation
Sung-Wook Lee (University of Virginia), Yen-Ling Kuo (University of Virginia)
Representation LearningRobotic IntelligenceConvolutional Neural NetworkContrastive LearningImageSequential
🎯 What it does: Propose the CLASS (Contrastive Learning via Action Sequence Supervision) method, which learns visual representations through supervised contrastive learning, aggregating states with similar subsequent actions in the latent space.
CLONE: Closed-Loop Whole-Body Humanoid Teleoperation for Long-Horizon Tasks
Yixuan Li (Peking University), Siyuan Huang (Beijing Institute of Technology)
Robotic IntelligenceMixture of ExpertsSimultaneous Localization and MappingPoint CloudTime SeriesStochastic Differential Equation
🎯 What it does: Designed and implemented the CLONE system, achieving full-body teleoperation of humanoid robots using only head and hand poses from a mixed reality headset, and addressed全身 coordination and localization drift issues during long-term tasks.
ClutterDexGrasp: A Sim-to-Real System for General Dexterous Grasping in Cluttered Scenes
Zeyuan Chen (Peking University), Hao Dong (Peking University)
Domain AdaptationRobotic IntelligenceReinforcement LearningDiffusion modelPoint Cloud
🎯 What it does: Proposed the ClutterDexGrasp framework to achieve zero-shot closed-loop target-oriented dexterous grasping from simulation to real-world, achieving high success rates in cluttered scenes.
Co-Design of Soft Gripper with Neural Physics
Sha Yi (University of California San Diego), Xiaolong Wang (University of California San Diego)
OptimizationPoint CloudMesh
🎯 What it does: This paper proposes a soft gripper co-design framework that jointly optimizes the block stiffness distribution of the gripper and the grasping posture, utilizing a neural physical agent trained in simulation for fast differentiable evaluation and optimization; finally, multiple stiffness configurations of the gripper are realized through 3D printing and their performance is validated on hardware.
CogniPlan: Uncertainty-Guided Path Planning with Conditional Generative Layout Prediction
Yizhuo Wang (National University of Singapore), Guillaume Adrien Sartoretti (National University of Singapore)
OptimizationRobotic IntelligenceGraph Neural NetworkReinforcement LearningGenerative Adversarial NetworkImageGraph
🎯 What it does: Propose CogniPlan, a framework that utilizes multiple feasible layout predictions combined with a graph attention network for path planning in unknown environments.
COLLAGE: Adaptive Fusion-based Retrieval for Augmented Policy Learning
Sateesh Kumar (University of Texas at Austin), Roberto Martín-Martín (University of Texas at Austin)
RetrievalTransformerReinforcement LearningMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the COLLAGE method, which enhances few-shot imitation learning performance through multi-modal retrieval and adaptive weighted fusion of data.
COMBO-Grasp: Learning Constraint-Based Manipulation for Bimanual Occluded Grasping
Jun Yamada (University of Oxford), Ingmar Posner (University of Oxford)
Knowledge DistillationRobotic IntelligenceReinforcement LearningDiffusion modelMesh
🎯 What it does: Research a dual-arm collaborative robotic grasping method to address the challenge of grasping occluded objects.
ComposableNav: Instruction-Following Navigation in Dynamic Environments via Composable Diffusion
Zichao Hu (University of Texas at Austin), Joydeep Biswas (University of Texas at Austin)
Robotic IntelligenceReinforcement LearningDiffusion model
🎯 What it does: Studied generating robot navigation trajectories that meet multiple specifications based on natural language instructions in dynamic environments.
Constrained Style Learning from Imperfect Demonstrations under Task Optimality
Kehan Wen (ETH Zurich), Marco Hutter (ETH Zurich)
OptimizationRobotic IntelligenceReinforcement LearningTime Series
🎯 What it does: This paper proposes the ConsMimic framework, which safely learns robot styles from imperfect demonstrations under constrained Markov decision processes (CMDP) through adaptive Lagrange multipliers, while ensuring task performance remains close to optimal.
Constraint-Aware Diffusion Guidance for Robotics: Real-Time Obstacle Avoidance for Autonomous Racing
Hao Ma (ETH Zurich), Michael Muehlebach (Max Planck Institute for Intelligent Systems)
Autonomous DrivingRobotic IntelligenceDiffusion modelScore-based ModelSequentialStochastic Differential Equation
🎯 What it does: Proposes the Constraint-Aware Diffusion Guidance (CoDiG) framework, which incorporates a differentiable obstacle function into the reverse diffusion process to enable real-time safe trajectory generation based on diffusion models, achieving dynamic obstacle avoidance in a mini unmanned racing car.
Constraint-Preserving Data Generation for One-Shot Visuomotor Policy Generalization
Kevin Lin (Stanford University), Jeannette Bohg (Stanford University)
Data SynthesisOptimizationRobotic IntelligenceDiffusion modelImage
🎯 What it does: Propose a data generation framework called CP-Gen, which can generate diverse demonstrations containing new object geometries and poses using only a single expert demonstration, for training closed-loop visual motion policies.
Contrastive Forward Prediction Reinforcement Learning for Adaptive Fault-Tolerant Legged Robots
Yangqing Fu (Shanghai Jiao Tong University), Yue Gao (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement LearningContrastive Learning
🎯 What it does: Propose a reinforcement learning framework combining contrastive learning and forward prediction to achieve adaptive fault-tolerant control for legged robots under joint damage scenarios.
ControlVLA: Few-shot Object-centric Adaptation for Pre-trained Vision-Language-Action Models
Puhao Li (Tsinghua University), Siyuan Huang (State Key Lab of General Artificial Intelligence, BIGAI)
Robotic IntelligenceMeta LearningTransformerSupervised Fine-TuningVision-Language-Action ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes the ControlVLA framework, achieving robot manipulation learning under few-shot (10–20 demonstrations) conditions by integrating pre-trained Vision-Language-Action (VLA) models with object-based representations.
CoRI: Communication of Robot Intent for Physical Human-Robot Interaction
Junxiang Wang (Carnegie Mellon University), Zackory Erickson (Carnegie Mellon University)
GenerationPose EstimationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose the CoRI (Communication of Robot Intent) pipeline, which can automatically generate natural language intent communication from a robot's visual perception and motion planning, covering overall goals, motion details, and user collaboration, applicable to physical human-robot interaction scenarios;
Cost-aware Discovery of Contextual Failures using Bayesian Active Learning
Anjali Parashar (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
Autonomous DrivingRobotic IntelligenceLarge Language ModelImage
🎯 What it does: Propose a cost-aware Bayesian active learning framework that constructs a probabilistic surrogate model for multi-modal contextual failure using expert feedback, and achieves diverse and efficient failure discovery through active exploration.
Cross-Sensor Touch Generation
Samanta Rodriguez (University of Michigan), Nima Fazeli (University of Michigan)
GenerationData SynthesisDepth EstimationDomain AdaptationConvolutional Neural NetworkDiffusion modelMultimodality
🎯 What it does: Propose two cross-sensor touch generation methods that enable the output of one touch sensor to be converted into the output of another sensor to achieve model transfer.
Crossing the Human-Robot Embodiment Gap with Sim-to-Real RL using One Human Demonstration
Tyler Ga Wei Lum (Stanford University), Jeannette Bohg (Stanford University)
Pose EstimationDepth EstimationDomain AdaptationRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningVideo
🎯 What it does: By using only a segment of human RGB-D video demonstration, fine-grained manipulation policies are trained in simulation via reinforcement learning, enabling zero-shot transfer to real robots.
CUPID: Curating Data your Robot Loves with Influence Functions
Christopher Agia (Stanford University), Jeannette Bohg (Stanford University)
Data-Centric LearningRobotic IntelligenceReinforcement LearningDiffusion modelBenchmark
🎯 What it does: This paper proposes CUPID, a data screening and subset selection method for robot imitation learning based on influence functions, which can quantify the impact of each demonstration data on the policy's closed-loop expected return and perform data pruning or augmentation accordingly;
D-CODA: Diffusion for Coordinated Dual-Arm Data Augmentation
I-Chun Arthur Liu (University of Southern California), Daniel Seita (University of Southern California)
SegmentationGenerationData SynthesisOptimizationRobotic IntelligenceConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: For dual-arm eye-camera imitation learning, D-CODA is proposed to generate perspective-consistent wrist images and corresponding actions via diffusion models, achieving offline data augmentation.
D-Cubed: Latent Diffusion Trajectory Optimisation for Dexterous Deformable Manipulation
Jun Yamada (University of Oxford), Ingmar Posner (University of Oxford)
OptimizationRobotic IntelligenceDiffusion modelScore-based ModelAuto EncoderVideoSequentialBenchmark
🎯 What it does: Studied a method called D‑Cubed that utilizes latent diffusion models (LDM) and gradient-free CEM-guided trajectory optimization for solving long-horizon manipulation tasks with multi-fingered robotic hands on flexible objects.
Data Retrieval with Importance Weights for Few-Shot Imitation Learning
Amber Xie, Joey Hejna
RetrievalRobotic IntelligenceBenchmark
🎯 What it does: Proposes a new importance-weighted retrieval method (IWR) for few-shot imitation learning from small-scale demonstration datasets, by enhancing the limited demonstration dataset through extracting relevant samples from a large prior dataset.
Decentralized Aerial Manipulation of a Cable-Suspended Load Using Multi-Agent Reinforcement Learning
Jack Zeng (Delft University of Technology), Sihao Sun (Delft University of Technology)
Robotic IntelligenceReinforcement LearningAgentic AITabular
🎯 What it does: Proposed and implemented the first fully decentralized, real-time executable method for full-attitude control of UAVs协同绳索悬挂负载 (UAVs协同绳索悬挂负载) on multiple MAVs, achieving zero-shot transfer experiments from simulation to the real world;
Deep Reactive Policy: Learning Reactive Manipulator Motion Planning for Dynamic Environments
Jiahui Yang (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
Knowledge DistillationRobotic IntelligenceTransformerPoint Cloud
🎯 What it does: Propose Deep Reactive Policy (DRP), a vision-motion neural network policy based on point clouds, capable of achieving collision-free goal reaching in dynamic, partially observable environments; the core model IMPACT is a behavior cloning network based on Transformer, pre-trained on 10M expert trajectories generated by cuRobo, further enhanced through student-teacher distillation and geometric fabrics (Geometric Fabrics) for static obstacle avoidance, and finally integrated with DCP-RMP end-to-end local dynamic obstacle reaction module.
DemoSpeedup: Accelerating Visuomotor Policies via Entropy-Guided Demonstration Acceleration
Lingxiao Guo (Shanghai Qi Zhi Institute), Huazhe Xu (Tsinghua University)
Anomaly DetectionComputational EfficiencyRobotic IntelligenceDiffusion modelVideoSequential
🎯 What it does: Propose the DemoSpeedup method, which segments visual motion demonstrations using self-supervised action entropy estimation and adaptively downsamples based on entropy values to accelerate the execution of visual motion policies.
DEQ-MPC : Deep Equilibrium Model Predictive Control
Swaminathan Gurumurthy (Carnegie Mellon University), Zachary Manchester (Carnegie Mellon University)
OptimizationRobotic IntelligenceRecurrent Neural NetworkTime Series
🎯 What it does: Proposes a deep equilibrium control framework named DEQ-MPC, which jointly models neural networks and MPC solvers as a fixed-point problem;
Dexplore: Scalable Neural Control for Dexterous Manipulation from Reference Scoped Exploration
Sirui Xu (University of Illinois Urbana-Champaign), Wei Yang (NVIDIA)
Knowledge DistillationRepresentation LearningRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAuto EncoderImageVideoPoint Cloud
🎯 What it does: Proposes DEXPLORE, a framework that treats human MoCap references as soft constraints, unifying trajectory tracking and policy learning through adaptive spatial range and reinforcement learning, and distills the learned state controller into a generative visual controller based solely on single-view depth maps, achieving dexterous manipulation on real robotic hands.
DexSkin: High-Coverage Conformable Robotic Skin for Learning Contact-Rich Manipulation
Suzannah Wistreich (Stanford University), Jiajun Wu (Stanford University)
Robotic IntelligenceReinforcement LearningDiffusion model
🎯 What it does: Propose a soft capacitive tactile skin called DexSkin and integrate it into parallel grippers to enhance the learning and execution capabilities of robots in tactile-rich environments.
DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation
Mengda Xu (Stanford University), Shuran Song (Stanford University)
SegmentationData SynthesisRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningVision-Language-Action ModelDiffusion modelImageMultimodality
🎯 What it does: This paper proposes the DexUMI framework, which utilizes a wearable arm exoskeleton to directly map human hand movements to operations of multi-fingered robot hands, thereby achieving seamless transfer across different robot hands.
DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control
Junjie Wen (Midea Group), Feifei Feng (Midea Group)
Robotic IntelligenceMixture of ExpertsVision-Language-Action ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Propose the DexVLA framework, enabling robots to perform diverse, cross-platform fine operations through a vision-language-action model, supporting complex long-term tasks.
Diffusion Dynamics Models with Generative State Estimation for Cloth Manipulation
Tongxuan Tian (University of California San Diego), Hao Su (University of California San Diego)
OptimizationRobotic IntelligenceTransformerDiffusion modelPoint CloudTime Series
🎯 What it does: Propose a unified Diffusion framework named UniClothDiff, which employs Transformer-based diffusion models to perform fabric state estimation (DPM) and dynamics prediction (DDM), and integrates them into Model Predictive Control (MPC) to achieve real-world robotic folding tasks.
Diffusion-Guided Multi-Arm Motion Planning
Viraj Parimi (Massachusetts Institute of Technology), Brian C. Williams (Massachusetts Institute of Technology)
Robotic IntelligenceDiffusion model
🎯 What it does: Proposed DG-MAP, a multi-arm motion planning framework that integrates MAPF ideas with two dedicated conditional diffusion models to address collision and scalability issues in multi-arm collaboration.
Disentangled Multi-Context Meta-Learning: Unlocking Robust and Generalized Task Learning
Seonsoo Kim (Agency for Defense Development), Seongil Hong (Agency for Defense Development)
Robotic IntelligenceMeta LearningTabular
🎯 What it does: Proposed a Disentangled Multi-Context Meta-Learning (DMCM) framework that learns separate context vectors for each task factor and updates only the relevant vectors internally during inner-loop optimization;
Distilling On-device Language Models for Robot Planning with Minimal Human Intervention
Zachary Ravichandran (University of Pennsylvania), Vijay Kumar (University of Pennsylvania)
Knowledge DistillationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the PRISM framework, which distills a small, device-deployable robot planner from a large language model using synthetic task and environment data.
Distributed Upload and Active Labeling for Resource-Constrained Fleet Learning
Oguzhan Akcin (University of Texas at Austin), Sandeep P. Chinchali (University of Texas at Austin)
Federated LearningRobotic IntelligenceConvolutional Neural NetworkTransformerVision Language ModelImageMultimodalityPoint CloudAudio
🎯 What it does: Propose a two-stage distributed data upload and active annotation framework (DUAL), enabling multiple robots to efficiently select and upload information in communication and storage-constrained environments, while performing active annotation on the cloud based on a global annotation budget;
Divide, Discover, Deploy: Factorized Skill Learning with Symmetry and Style Priors
Rafael Cathomen (ETH Zurich), Marco Hutter (ETH Zurich)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: Propose a modular unsupervised skill learning framework based on user-defined state decomposition, optional USD algorithms, symmetry, and style priors, capable of training safe, interpretable skills transferable directly to real-world quadruped robots in simulation;
DiWA: Diffusion Policy Adaptation with World Models
Akshay L Chandra (University of Freiburg), Abhinav Valada (University of Freiburg)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelWorld ModelVideoSequential
🎯 What it does: This paper proposes the DiWA framework, which uses a world model for offline learning to perform fully offline reinforcement learning fine-tuning on pre-trained diffusion policies, enabling robot skill adaptation without any physical interaction.
Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving
Mingyi Wang (Autolab, Westlake University), Kaicheng Yu (Autolab, Westlake University)
Autonomous DrivingTransformerLarge Language ModelReinforcement LearningTime SeriesSequential
🎯 What it does: Systematically evaluate and transfer core modules of large language models (LLMs) (tokenizer, position encoding, pre-training, post-training, and inference-time computation) for motion generation tasks in autonomous driving.
DreamGen: Unlocking Generalization in Robot Learning through Video World Models
Joel Jang, Linxi Fan
Data SynthesisRobotic IntelligenceTransformerSupervised Fine-TuningPrompt EngineeringVision-Language-Action ModelDiffusion modelWorld ModelVideoTextBenchmark
🎯 What it does: Developed a four-stage pipeline called DREAMGEN, which leverages a video world model to generate synthetic training data, recovers pseudo actions through inverse dynamics or latent action models, and ultimately trains a visuomotor control policy to achieve generalization across multiple behaviors and environments.
Dynamics-Compliant Trajectory Diffusion for Super-Nominal Payload Manipulation
Anuj Pasricha (University of Colorado Boulder), Alessandro Roncone (University of Colorado Boulder)
OptimizationRobotic IntelligenceConvolutional Neural NetworkDiffusion modelSequential
🎯 What it does: By training a trajectory generator based on a denoising diffusion model, executable motion trajectories are generated in a single step within the joint angle, velocity, and acceleration spaces that satisfy given load constraints (maximum torque, velocity, acceleration, collision), thereby enabling robots to carry payloads far exceeding the nominal load without exceeding hardware limits.
Efficient Evaluation of Multi-Task Robot Policies With Active Experiment Selection
Abrar Anwar (University of Southern California), Jesse Thomason (University of Southern California)
Robotic Intelligence
🎯 What it does: Studied a cost-aware active experimental selection framework for efficiently evaluating the performance distribution of multi-task robot policies.
Elucidating the Design Space of Torque-aware Vision-Language-Action Models
Zongzheng Zhang (Beijing Academy Of Artificial Intelligence), Hao Zhao (Beijing Academy Of Artificial Intelligence)
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Building upon existing Vision-Language-Action (VLA) models, the paper systematically explores incorporating joint torque signals as physical perception inputs, designs multiple fusion strategies, and ultimately proposes three key design principles;
Embrace Contacts: humanoid shadowing with full body ground contacts
Ziwen Zhuang (Tsinghua University), Hang Zhao (Tsinghua University)
Robotic IntelligenceTransformerReinforcement LearningVideoSequential
🎯 What it does: A unified full-body contact motion interface was constructed, and zero-shot sim-to-real reinforcement learning policies were trained in GPU-accelerated physics simulations, enabling real-time execution of full-body motion trained in simulation on the real Unitree G1 humanoid robot.
Enabling Long(er) Horizon Imitation for Manipulation Tasks by Modeling Subgoal Transitions
Shivam Jain (Indian Institute of Technology Delhi), Rohan Paul (Indian Institute of Technology Delhi)
Robotic IntelligenceTransformerReinforcement LearningSequentialBenchmark
🎯 What it does: Proposed two transformer-based strategies (ST-GPT and SGPT) to achieve long-horizon imitation learning by learning subgoal transition signals, validated on both simulated and real robots.
EndoVLA: Dual-Phase Vision-Language-Action for Precise Autonomous Tracking in Endoscopy
CHI KIT NG, Hongliang Ren (Chinese University of Hong Kong)
Object TrackingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelImageMultimodalityBiomedical Data
🎯 What it does: Propose EndoVLA, a vision-language-action model for autonomous tracking in continuous robotic endoscopy.
Ensuring Force Safety in Vision-Guided Robotic Manipulation via Implicit Tactile Calibration
Lai Wei (University of California San Diego), Ruimao Zhang (Sun Yat-sen University)
Robotic IntelligenceSupervised Fine-TuningDiffusion modelMultimodality
🎯 What it does: Propose a safety state planning method SafeDiff based on diffusion models, which addresses the problem of harmful forces caused by visual errors in constrained trajectory tasks such as door opening by adding real-time tactile calibration on visually guided end-effector sequences.
Enter the Mind Palace: Reasoning and Planning for Long-term Active Embodied Question Answering
Muhammad Fadhil Ginting (Field AI), Shayegan Omidshafiei (Field AI)
Robotic IntelligenceLarge Language ModelAgentic AIVision Language ModelVision-Language-Action ModelImageVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper studies the Long-term Active Embodied Question Answering (LA-EQA) task, proposing the Mind Palace Exploration method, which organizes the robot's long-term observations into multi-instance scene graphs, combining memory retrieval with active exploration to answer spatiotemporal questions.
Estimating Value of Assistance for Online POMDP Robotic Agents
Yuval Goshen (Technion Israel Institute of Technology), Sarah Keren (Technion Israel Institute of Technology)
Robotic IntelligenceBenchmark
🎯 What it does: This paper proposes a framework for evaluating the Value of Assistance (VOA) in online POMDP planning, and designs three low-computation-cost heuristic methods.
Extracting Visual Plans from Unlabeled Videos via Symbolic Guidance
Wenyan Yang (Aalto University), Joni Pajarinen (University of Trento)
Object DetectionRepresentation LearningRobotic IntelligenceTransformerReinforcement LearningContrastive LearningVideo
🎯 What it does: Vis2Plan proposes a symbol-guided visual planning framework that extracts object-centric symbols from unlabeled demonstrations using pre-trained visual foundation models, constructs symbolic graphs for search, generates physically reachable visual subgoals, and drives low-level control to complete multi-stage robotic tasks.
exUMI: Extensible Robot Teaching System with Action-aware Task-agnostic Tactile Representation
Yue Xu (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Innovation Institute)
Representation LearningRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelAuto EncoderMultimodality
🎯 What it does: Developed the exUMI handheld device and the TPP pre-training framework, achieving scalable multi-modal data collection and action-aware tactile representation learning.
Eye, Robot: Learning to Look to Act with a BC-RL Perception-Action Loop
Justin Kerr (UC Berkeley), Angjoo Kanazawa (UC Berkeley)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningContrastive LearningVideo
🎯 What it does: Studied a mechanically rotatable artificial eye (EyeRobot) that achieves hand-eye coordination for grasping and placing tasks in a large workspace through a closed-loop system combining behavior cloning (BC) and reinforcement learning (RL).
Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning
Yunsheng Tian (MIT), Wojciech Matusik (MIT)
Robotic IntelligenceReinforcement LearningMeshBenchmark
🎯 What it does: Designed and implemented a dual-arm robotic system named Fabrica, achieving a complete automated assembly process from CAD models to multi-part objects, including hierarchical planning, automatic fixture generation, and a general insertion strategy, with multi-step assembly completed on a real robot.
FACET: Force-Adaptive Control via Impedance Reference Tracking for Legged Robots
Botian Xu (Tsinghua University), Huazhe Xu (Tsinghua University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Train simulation models using reinforcement learning to enable quadruped, biped, and arm-equipped robots to achieve adjustable compliance and high force output through a virtual mass-spring-damper system in response to external forces.
Fail2Progress: Learning from Real-World Robot Failures with Stein Variational Inference
Yixuan Huang (University Of Utah), Tucker Hermans (University Of Utah)
Data SynthesisDomain AdaptationRobotic IntelligenceSupervised Fine-TuningPoint CloudTabular
🎯 What it does: This paper proposes the Fail2Progress method, which utilizes Stein variational inference (SVI) to rapidly generate diverse training data for real-world failure cases in simulated environments, thereby fine-tuning a pre-trained skill effect model to improve the success rate of long-cycle mechanical operations.
Fast Flow-based Visuomotor Policies via Conditional Optimal Transport Couplings
Andreas Sochopoulos (University of Edinburgh), Sethu Vijayakumar (Honda Research Institute Europe)
OptimizationComputational EfficiencyRobotic IntelligenceFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose a flow matching strategy (COT Policy) leveraging conditional optimal transport (COT) coupling to achieve high-speed inference for robot vision-action generation.
FastUMI: A Scalable and Hardware-Independent Universal Manipulation Interface with Dataset
Zhaxizhuoma (Shanghai Jiao Tong University), Xuelong Li (Institute of AI, China Telecom Corp Ltd)
Depth EstimationRobotic IntelligenceRecurrent Neural NetworkTransformerVision-Language-Action ModelVideoSequential
🎯 What it does: This paper designs the FastUMI system, achieving hardware decoupling, replacing complex VIO with commercial trackers, integrating a data collection ecosystem, and releasing 15,000 real demonstrations across 24 tasks.
FetchBot: Learning Generalizable Object Fetching in Cluttered Scenes via Zero-Shot Sim2Real
Weiheng Liu (Institute of Automation, Chinese Academy of Sciences), He Wang (Peking University)
Depth EstimationDomain AdaptationRobotic IntelligenceTransformerDiffusion modelImage
🎯 What it does: Proposes the FetchBot sim-to-real framework for generalizable object grasping in cluttered scenes.
Few-Shot Neuro-Symbolic Imitation Learning for Long-Horizon Planning and Acting
Pierrick Lorang (Tufts University), Matthias Scheutz (Austrian Institute of Technology)
Robotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackGraph Neural NetworkDiffusion modelImageSequential
🎯 What it does: Propose a neural symbolic imitation learning framework that can simultaneously learn low-level continuous control policies and high-level symbolic planning models from a few raw demonstrations, achieving planning and execution for long-horizon tasks.
FFHFlow: Diverse and Uncertainty-Aware Dexterous Grasp Generation via Flow Variational Inference
Qian Feng (Amigos Robots), Alois Knoll (Technical University of Munich)
GenerationRobotic IntelligenceFlow-based ModelPoint Cloud
🎯 What it does: Proposes FFHFlow—a multi-finger grasp generation framework based on Normalizing Flows, capable of generating diverse and reliable grasp poses under partial point cloud observations, while quantifying and self-reflecting on shape uncertainty.
First Order Model-Based RL through Decoupled Backpropagation
Joseph Amigo (New York University), Ludovic Righetti (New York University)
Reinforcement LearningWorld ModelBenchmark
🎯 What it does: Propose a model-based reinforcement learning framework called DMO, where the prerequisite is decoupling trajectory generation from gradient computation; utilize a GPU-accelerated high-fidelity simulator to generate trajectories, and compute gradients using a learned differentiable dynamics model, thereby achieving efficient first-order gradient optimization.
FLARE: Robot Learning with Implicit World Modeling
Ruijie Zheng (NVIDIA), Linxi Fan (NVIDIA)
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelFlow-based ModelWorld ModelVideoMultimodalityBenchmark
🎯 What it does: Propose the FLARE framework, introducing implicit world modeling in diffusion/flow-matching robot policies by aligning the latent embeddings of future observations to the hidden states of the denoising network, enabling the policy to predict and reason about future states while generating actions.
FlashBack: Consistency Model-Accelerated Shared Autonomy
Luzhe Sun (Toyota Technological Institute at Chicago), Matthew Walter
Knowledge DistillationRobotic IntelligenceReinforcement LearningDiffusion modelOrdinary Differential Equation
🎯 What it does: This paper proposes a shared autonomous framework CSA based on a consistency model, utilizing ODE knowledge distillation to achieve single-step denoising in real-time, assisting human-machine collaborative control.
FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Flow Models
Moritz Reuss (Karlsruhe Institute of Technology), Rudolf Lioutikov (Karlsruhe Institute of Technology)
Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelFlow-based ModelRectified FlowMultimodality
🎯 What it does: Designed an efficient vision-language-action (VLA) strategy called FLOWER, which can accomplish 190 tasks with less than 1B parameters and a pre-training cost of only 200 GPU hours.
Focusing on What Matters: Object-Agent-centric Tokenization for Vision Language Action models
Rokas Bendikas (UCL), Pietro Mazzaglia
Object DetectionSegmentationComputational EfficiencyTransformerLarge Language ModelVision-Language-Action ModelImageText
🎯 What it does: Reconstruct the visual input of Vision-Language-Action models using an object-agent-centric tokenization method, significantly reducing the number of visual tokens and improving training efficiency.
FOMO-3D: Using Vision Foundation Models for Long-Tailed 3D Object Detection
Anqi Joyce Yang (Waabi University of Toronto), Raquel Urtasun (Waabi University of Toronto)
Object DetectionDepth EstimationAutonomous DrivingTransformerImageMultimodalityPoint Cloud
🎯 What it does: Propose FOMO-3D, a multi-modal 3D detection framework that integrates 2D detection and depth information from the visual foundation model OWL and Metric3D to address the challenge of long-tail object detection.
Force-Modulated Visual Policy for Robot-Assisted Dressing with Arm Motions
Alexis Yihong Hao (Carnegie Mellon University), Zackory Erickson (Carnegie Mellon University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Developed a robot-assisted dressing system capable of adaptively and safely wearing long-sleeved clothing during arm movement.
From Real World to Logic and Back: Learning Generalizable Relational Concepts For Long Horizon Robot Planning
Naman Shah (Arizona State University), Siddharth Srivastava (Arizona State University)
Robotic IntelligenceWorld ModelSequential
🎯 What it does: This paper proposes the LAMP (Learning Abstract Models for Planning) method, which can automatically invent symbolic relational concepts and world models from a small number of unlabeled, unsegmented motion trajectories, enabling robots to achieve zero-shot reasoning and planning in unseen, more complex long-horizon tasks.
From Space to Time: Enabling Adaptive Safety with Learned Value Functions via Disturbance Recasting
Sander Tonkens (University of California San Diego), Sylvia Lee Herbert
Time SeriesSequential
🎯 What it does: Propose the SPACE2TIME framework, achieving adaptive safety filtering by offline learning of a time-varying safety value function in spatially varying perturbation environments.