RSS 2025 Papers — Page 2
Robotics: Science and Systems · 163 papers
Learning to Act Anywhere with Task-centric Latent Actions
Qingwen Bu, Hongyang Li
Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelAuto EncoderVideoTextMultimodality
🎯 What it does: Built a unified visual-language-action framework called UniVLA, which can learn task-centered latent actions through unsupervised learning across different robots, tasks, and perspectives, achieving cross-system general control.
Leveling the Playing Field: Carefully Comparing Classical and Learned Controllers for Quadrotor Trajectory Tracking
Pratik Kunapuli, Vijay Kumar
OptimizationRobotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: A fair experimental protocol for comparing reinforcement learning (RL) and geometric control (GC) methods in trajectory tracking for quadrotors and fixed-arm aerial manipulators is proposed, and the best performance of both controller types is evaluated based on this protocol.
LiDAR Registration with Visual Foundation Models
Niclas Vödisch, Davide Scaramuzza
Pose EstimationDomain AdaptationAutonomous DrivingTransformerContrastive LearningSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Leverage the visual foundation model DINOv2 to extract high-dimensional features from surround-view images, serving as descriptors for LiDAR point clouds. Features are mapped to LiDAR points via point-to-pixel projection, followed by traditional registration algorithms (RANSAC, ICP, or TEASER++) to achieve 6DoF scan-to-map registration in long-term, cross-seasonal scenarios.
Manual2Skill: Learning to Read Manuals and Acquire Robotic Skills for Furniture Assembly Using Vision-Language Models
Chenrui Tie, Lin Shao
Robotic IntelligenceGraph Neural NetworkTransformerVision Language ModelImageTextPoint Cloud
🎯 What it does: Propose the Manual2Skill framework based on a visual language model, enabling robots to learn from abstract instructions and automatically complete furniture assembly.
Map Space Belief Prediction for Manipulation-Enhanced Mapping
Joao Marcos Correia Marques, Kris Hauser
Robotic IntelligenceVision-Language-Action ModelSimultaneous Localization and MappingImage
🎯 What it does: Propose a manipulation-enhanced mapping (MEM) framework based on POMDP, which implements joint Bayesian inference of observations and manipulation actions through calibrated Bayesian updating (CNABU) learned by neural networks, achieving zero-shot transfer on simulation and real robots.
Meta-Learning Online Dynamics Model Adaptation in Off-Road Autonomous Driving
Jacob Levy, Patrick Spieler
Autonomous DrivingOptimizationMeta LearningRecurrent Neural NetworkSimultaneous Localization and MappingSequential
🎯 What it does: Propose a meta-learning based online dynamic model adaptation framework, combining Kalman filters to real-time update vehicle dynamics, enabling safe and efficient driving of autonomous off-road vehicles on unknown terrains.
MISO: Multiresolution Submap Optimization for Efficient Globally Consistent Neural Implicit Reconstruction
Yulun Tian, Nikolay Atanasov
OptimizationComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed the MISO framework, which utilizes multi-resolution subgraphs for hierarchical optimization of neural implicit SDF, enabling efficient globally consistent 3D reconstruction and SLAM;
Morpheus: A Neural-driven Animatronic Face with Hybrid Actuation and Diverse Emotion Control
Zongzheng Zhang, Hao Zhao
GenerationPose EstimationRobotic IntelligenceTransformerImageVideoMeshAudio
🎯 What it does: Proposed a hybrid-driven anthropomorphic facial system called Morpheus, which can map emotional speech in real-time to diverse expressions through neural networks, and utilize a self-built model to map expressions to control signals of 33 motors, achieving realistic facial animations.
NaVILA: Legged Robot Vision-Language-Action Model for Navigation
An-Chieh Cheng, Xiaolong Wang
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelImageVideoTextPoint Cloud
🎯 What it does: Propose NaVILA, a two-layer framework that combines intermediate language instructions generated by vision-language models (VLM) with low-level visual reinforcement learning (RL) strategies, enabling multi-legged robot navigation based on single-view RGB instructions.
Neural Inertial Odometry from Lie Events
Royina Karegoudra Jayanth, Daniel Gehrig
Pose EstimationConvolutional Neural NetworkTime SeriesSequential
🎯 What it does: Propose a neural inertial odometry method based on Lie group events, generating event sequences using pre-integrated SE(3) trajectories to achieve robustness to sampling rates and motion variations.
Novel Demonstration Generation with Gaussian Splatting Enables Robust One-Shot Manipulation
Sizhe Yang, Jiangmiao Pang
Data SynthesisRobotic IntelligenceTransformerGaussian SplattingImage
🎯 What it does: Propose the RoboSplat method, which uses a single real demonstration combined with multi-view images to reconstruct and automatically edit using a 3D Gaussian distribution, generating diverse visual samples to train a visual motion controller.
On the Surprising Robustness of Sequential Convex Optimization for Contact-Implicit Motion Planning
Yulin Li, Heng Yang
OptimizationRobotic IntelligenceSequentialBenchmark
🎯 What it does: Proposed a pure primal sequential convex optimization framework CRISP to solve mathematical programming and complementarity constraints in contact-implicit motion planning, with theoretical convergence guarantees; simultaneously implemented a high-performance C++ solver and conducted experiments on six benchmark tasks involving complex contact dynamics.
Online Competitive Information Gathering for Partially Observable Trajectory Games
Mel Krusniak, Forrest John Laine
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: An online active information acquisition method for partially observable trajectory games (POSG) is proposed, which combines model predictive game playing (MPGP) with particle filtering to estimate joint states and observation history, and generates strategies by solving Nash equilibrium via gradient at each time step.
Optimal Interactive Learning on the Job via Facility Location Planning
Shivam Vats, George Konidaris
OptimizationRobotic Intelligence
🎯 What it does: Propose the COIL (Cost-Optimal Interactive Learning) framework, which proactively determines three types of interactive queries (skill teaching, preference elicitation, human execution) in a multi-task environment using facility location planning methods to minimize human workload.
Particle-Grid Neural Dynamics for Learning Deformable Object Models from RGB-D Videos
Kaifeng Zhang, Yunzhu Li
GenerationGaussian SplattingVideo
🎯 What it does: This paper proposes a particle-grid neural dynamics framework that achieves full-dense dynamic prediction of deformable objects and generates 3D action-conditioned videos by learning particle representations and grid velocity fields directly from multi-view RGB-D videos.
PartInstruct: Part-level Instruction Following for Fine-grained Robot Manipulation
Yifan Yin, Tianmin Shu
Robotic IntelligenceReinforcement LearningVision Language ModelVision-Language-Action ModelDiffusion modelMultimodalityPoint CloudBenchmark
🎯 What it does: Constructed and released the PartInstruct benchmark, containing 3D part annotations for 14 categories of daily objects, 10,000 expert demonstrations, and 1,302 fine-grained instruction tasks, along with the PartGym simulation environment for training and evaluating vision-language robot manipulation models.
Physics-Driven Data Generation for Contact-Rich Manipulation via Trajectory Optimization
Lujie Yang, Russ Tedrake
GenerationData SynthesisOptimizationRobotic IntelligenceDiffusion modelVideoTime SeriesPhysics Related
🎯 What it does: Propose a physics-driven trajectory optimization data generation pipeline, which first collects a small number of human demonstrations via VR, then performs kinematic reproduction and trajectory optimization to generate a large number of dynamically feasible and contact-rich robot demonstrations.
PIN-WM: Learning Physics-INformed World Models for Non-Prehensile Manipulation
Wenxuan Li, Kai Xu
Domain AdaptationRobotic IntelligenceReinforcement LearningGaussian SplattingWorld ModelImageVideoPhysics Related
🎯 What it does: This paper proposes a physics-informed world model (PIN-WM) that can end-to-end identify three-dimensional rigid dynamics and rendering parameters from RGB images using only a small number of visual interaction trajectories (random pushing actions) without task labels, and applies this physical simulation model to learn non-grasping manipulation (pushing/ flipping) strategies, achieving Sim2Real transfer without fine-tuning in the real environment.
PINGS: Gaussian Splatting Meets Distance Fields within a Point-Based Implicit Neural Map
Yue Pan, Cyrill Stachniss
Autonomous DrivingOptimizationGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposes a LiDAR-visual SLAM system called PINGS, which jointly constructs continuous distance fields and high-quality Gaussian splatting radiance fields, achieving geometric consistency and loop closure correction through point-based implicit neural graphs.
PP-Tac: Paper Picking Using Omnidirectional Tactile Feedback in Dexterous Robotic Hands
Pei Lin, Ziyuan Jiao
Robotic IntelligenceTransformerDiffusion modelMultimodality
🎯 What it does: Developed a system named PP-Tac, integrating a multi-fingered robotic hand with a high-resolution hemispherical tactile sensor (R-Tac), to achieve grasping of thin, flexible paper-like objects.
Prompting with the Future: Open-World Model Predictive Control with Interactive Digital Twins
Chuanruo Ning, Wei-Chiu Ma
Robotic IntelligenceVision Language ModelGaussian SplattingVideoTextMesh
🎯 What it does: Achieved model predictive control (MPC) for open-world robot tool manipulation by integrating interactive digital twins with vision-language models (VLM);
Provably-Safe, Online System Identification
Bohao Zhang, Ram Vasudevan
OptimizationRobotic Intelligence
🎯 What it does: Developed an online, provably safe system identification framework capable of real-time estimation of the inertial parameters of unknown payloads at the robot's end-effector under constraints and obstacle avoidance.
RAMEN: Real-time Asynchronous Multi-agent Neural Implicit Mapping
Hongrui Zhao, Negar Mehr
Simultaneous Localization and MappingImageMultimodalityPoint CloudBenchmark
🎯 What it does: Propose the RAMEN framework to realize real-time asynchronous multi-agent neural implicit mapping.
RAPID: Robust and Agile Planner Using Inverse Reinforcement Learning for Vision-Based Drone Navigation
Minwoo Kim, Hyongdong Oh
Autonomous DrivingRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningAuto EncoderImage
🎯 What it does: This study proposes a visual UAV path planning framework called RAPID based on inverse reinforcement learning, achieving high-speed and safe flight in dense obstacle environments.
Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation
Han Xue, Cewu Lu
Robotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkVision-Language-Action ModelDiffusion modelImageMultimodalityTime Series
🎯 What it does: Proposed the TactAR low-cost AR haptic feedback system and a two-level visual-haptic imitation learning framework with Reactive Diffusion Policy (RDP), addressing the conflict between high-frequency haptic closed-loop control and complex trajectory modeling in contact-rich tasks.
Resolving Conflicting Constraints in Multi-Agent Reinforcement Learning with Layered Safety
Jason Jangho Choi, Claire Tomlin
Graph Neural NetworkReinforcement Learning
🎯 What it does: Propose the Layered Safe MARL framework, combining multi-agent reinforcement learning with a layered safety filter to address conflict constraints in multi-robot navigation.
Riemannian Direct Trajectory Optimization of Rigid Bodies on Matrix Lie Groups
Sangli Teng, Maani Ghaffari
Optimization
🎯 What it does: Propose a direct optimization framework for rigid-body trajectory based on matrix Lie groups, achieving efficient optimization under full dynamic constraints using Lie group variational integrators and Riemannian interior-point methods;
RLDG: Robotic Generalist Policy Distillation via Reinforcement Learning
Charles Xu, Sergey Levine
Knowledge DistillationRobotic IntelligenceReinforcement LearningDiffusion modelBenchmark
🎯 What it does: This paper proposes the RLDG method, which first trains a dedicated RL policy to generate a high-quality dataset, and then fine-tunes the robot's base models (such as OpenVLA, Octo) to enhance their performance and generalization ability in dexterous manipulation tasks.
RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation
Kun Wu, Jian Tang
Robotic IntelligenceLarge Language ModelVision-Language-Action ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Created and released the RoboMIND multi-robot manipulation dataset, containing 107k trajectories, 479 tasks, and 96 object types, covering four categories of robots, failure cases, and fine-grained linguistic annotations.
RoboPanoptes: The All-Seeing Robot with Whole-body Dexterity
Xiaomeng Xu, Shuran Song
Robotic IntelligenceTransformerDiffusion modelVideo
🎯 What it does: Designed and implemented a multi-eyed full-body visual robot, RoboPanoptes, capable of achieving multi-contact and full-body manipulation through full-body cameras.
Robot Data Curation with Mutual Information Estimators
Joey Hejna, Dorsa Sadigh
Robotic IntelligenceAuto EncoderImageSequential
🎯 What it does: This work proposes a robot demonstration data filtering method called DemInf based on mutual information.
Robot Learning with Super-Linear Scaling
Marcel Torne Villasevil, Abhishek Gupta
Knowledge DistillationRobotic IntelligenceReinforcement LearningVision-Language-Action ModelNeural Radiance FieldContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: Propose a real-to-sim-to-real data collection pipeline (CASHER) based on crowd-sourced 3D scanning, combining simulation-based self-reinforcement learning with human demonstrations to achieve general strategy training across multiple environments, and enabling zero-shot and few-shot fine-tuning.
RoboVerse: A Unified Platform, Benchmark and Dataset for Scalable and Generalizable Robot Learning
Haoran Geng, Pieter Abbeel
Data SynthesisRobotic IntelligenceLarge Language ModelReinforcement LearningVision-Language-Action ModelDiffusion modelBenchmark
🎯 What it does: Built a unified robot learning platform called ROBOVERSE, including multi-simulator interfaces, a massive synthetic dataset, and cross-task evaluation benchmarks.
Robust Peg-in-Hole Assembly under Uncertainties via Compliant and Interactive Contact-Rich Manipulation
Yiting Chen, Kaiyu Hang
Robotic IntelligenceBenchmark
🎯 What it does: Designed and implemented a vision-free, contact-based robotic hole insertion method based on the 'manipulation funnel' framework, which gradually eliminates perception and execution uncertainties through reversible contact and environmental constraints to achieve precision hole insertion.
ROMAN: Open-Set Object Map Alignment for Robust View-Invariant Global Localization
Mason Peterson, Jonathan P. How
Robotic IntelligenceGraph Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: Propose the ROMAN method, leveraging open-set semantic segmentation, CLIP semantic embeddings, geometric shape features, and gravity direction prior to create and align viewpoint-invariant object maps for 3D segments, achieving global localization for robots under opposing or extreme viewpoint changes.
RUKA: Rethinking the Design of Humanoid Hands with Learning
Anya Zorin, Lerrel Pinto
Data SynthesisRobotic IntelligenceRecurrent Neural NetworkSequential
🎯 What it does: Proposed and implemented a low-cost, open-source human hand-shaped robotic hand called RUKA based on 3D printing and standard parts, achieving precise grasping through learning-driven methods.
Safe Beyond the Horizon: Efficient Sampling-based MPC with Neural Control Barrier Functions
Ji Yin, Panagiotis Tsiotras
Autonomous DrivingOptimization
🎯 What it does: Propose a sampling-based model predictive control (NS-VIMPC) based on variational inference, achieving safety guarantees beyond the prediction horizon by learning discrete-time control barrier functions (DCBF) and incorporating resampling (RBR) during the control sequence sampling process.
SafeMimic: Towards Safe and Autonomous Human-to-Robot Imitation for Mobile Manipulation
Arpit Bahety, Roberto Martín-Martín
Safty and PrivacyRobotic IntelligenceReinforcement LearningVision Language ModelVision-Language-Action ModelVideoTextPoint Cloud
🎯 What it does: Propose the SAFEMIMIC framework, enabling robots to safely and autonomously learn multi-step mobile manipulation tasks from a single third-person human video.
Safety with Agency: Human-Centered Safety Filter with Application to AI-Assisted Motorsports
Donggeon David Oh (Princeton University), Jaime Fernández Fisac
Autonomous DrivingReinforcement LearningTabular
🎯 What it does: Proposed a human-machine centered safety filter (HCSF) for shared autonomous systems, ensuring vehicle safety in high-risk racing environments without compromising human operational autonomy.
SATA: Safe and Adaptive Torque-Based Locomotion Policies Inspired by Animal Learning
Li Peizhuo, Guillaume Adrien Sartoretti
Domain AdaptationRobotic IntelligenceReinforcement Learning
🎯 What it does: Designed and implemented a torque-based quadrupedal walking control framework (SATA), achieving zero-shot transfer from simulation to real robots without additional tuning by incorporating a biomechanical muscle model and a progressive growth mechanism.
Self-supervised Multi-future Occupancy Forecasting for Autonomous Driving
Bernard Lange, Mykel Kochenderfer
Autonomous DrivingTransformerDiffusion modelGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: Propose a self-supervised multi-future occupancy grid prediction framework LOPR, which leverages LiDAR-generated occupancy grids together with RGB images, maps, and planning trajectories to predict possible future scenarios.
Sense and Sensibility: What makes an social robot convincing to high-school students?
Pablo Gonzalez-Oliveras, Ali Reza Majlesi
Robotic IntelligenceLarge Language ModelTextMultimodality
🎯 What it does: In Swedish high school students, researchers let students interact with a Furhat robot, answer eight circuit true/false questions, and examined the effect of the robot's different confidence levels (determined, balanced, uncertain) on students' final answers.
Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation
Abhiram Maddukuri, Yuke Zhu
Data SynthesisDomain AdaptationRobotic IntelligenceDiffusion modelImage
🎯 What it does: This paper proposes a co-training method that combines real-world and simulated data to train visual manipulation policies, and provides a directly applicable practical recipe.
Sketch-to-Skill: Bootstrapping Robot Learning with Human Drawn Trajectory Sketches
Peihong Yu, Pratap Tokekar
Robotic IntelligenceReinforcement LearningAuto EncoderImage
🎯 What it does: This paper proposes the SKETCH-TO-SKILL framework, which generates 3D robot trajectories from human-drawn 2D sketches. It rapidly learns manipulation tasks such as grasping and placing by combining behavior cloning pre-training with TD3-based reinforcement learning guided by a discriminator.
SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation
Shengjie Wang, Yang Gao
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision Language ModelDiffusion modelImageVideo
🎯 What it does: This work proposes the Semantic Keypoint Imitation Learning (SKIL) framework, which automatically extracts semantic keypoints as sparse observations using a visual foundation model, and combines them with a diffusion policy action head to achieve long-horizon, high-precision grasping and placement tasks with only a small number of demonstrations (around 30).
Solving Multi-Agent Safe Optimal Control with Distributed Epigraph Form MARL
Songyuan Zhang, Chuchu Fan
OptimizationGraph Neural NetworkReinforcement Learning
🎯 What it does: Proposed a distributed minimax form multi-agent safe reinforcement learning method (Def-MARL), which can solve multi-agent safe optimal control problems under strict zero constraints; meanwhile, theoretical proofs are provided to enable distributed execution, and its effectiveness is verified in simulation and real Crazyflie drone swarms.
SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models
Delin Qu, Xuelong Li
Depth EstimationRepresentation LearningRobotic IntelligenceTransformerVision-Language-Action ModelMultimodality
🎯 What it does: Built a new vision-language-action (VLA) model called SpatialVLA that can simultaneously process images, language instructions, and robot actions in a single model.
STDArm: Transfer Visuomotor Policy From Static Data Training to Dynamic Robot Manipulation
Yifan Duan, Yanyong Zhang
Domain AdaptationRobotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkTransformerDiffusion modelSimultaneous Localization and MappingVideo
🎯 What it does: Proposed the STDArm framework, which transfers visual motion policies trained in static environments to dynamic platforms (e.g., drones, quadruped robots), achieving high-frequency action correction and platform motion compensation.
Superfast Configuration-Space Convex Set Computation on GPUs for Online Motion Planning
Peter Werner, Daniela Rus
OptimizationRobotic IntelligencePoint CloudMeshGraph
🎯 What it does: Leveraging GPU parallel computing, the EI-ZO zeroth-order algorithm is proposed to rapidly construct a sequence of convex polyhedrons (SCS) with probabilistic collision safety, which is integrated with dynamic roadmaps (DRM) to form a complete online motion planning pipeline, achieving high-quality trajectory planning in perception-changing environments.
Tactile sensing enables vertical obstacle negotiation for elongate many-legged robots
Juntao He, Daniel Goldman
Robotic IntelligenceTime Series
🎯 What it does: Designed and tested an extensible multi-legged robot based on tactile antennas, achieving vertical climbing over obstacles up to five times its height.
Towards Uncertainty Unification: A Case Study for Preference Learning
Shaoting Peng, Katherine Rose Driggs-Campbell
OptimizationRobotic IntelligenceReinforcement Learning from Human Feedback
🎯 What it does: This paper proposes a preference learning framework called UUPL, which unifies human preference uncertainty with robot self-uncertainty. The framework mainly improves learning performance through user-specific confidence calibration, improvements to the posterior mean and covariance of Gaussian processes, and a weighted Gaussian Mixture Model (GMM) based on uncertainty.
Uncertainty-Aware Trajectory Prediction via Rule-Regularized Heteroscedastic Deep Classification
Kumar Manas, Nadja Klein
Autonomous DrivingLarge Language ModelImageTextPoint CloudRetrieval-Augmented Generation
🎯 What it does: Propose SHIFT, a trajectory prediction framework that combines uncertainty quantification with soft traffic rule constraints, converting trajectory prediction into a classification task and employing HetSNGP for heteroscedastic modeling.
Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks
Jiazhao Zhang, He Wang
Computational EfficiencyRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This study proposes Uni-NaVid, an end-to-end vision-language-action model based on RGB video streams and language instructions, capable of uniformly executing four tasks: audio-visual navigation, goal retrieval, embedded question answering, and human following.
Unified Video Action Model
Shuang Li, Shuran Song
GenerationRobotic IntelligenceTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: Propose a Unified Video and Action model (UVA) that jointly learns latent representations for videos and actions, and achieves efficient action prediction through decoupled diffusion inference; the model can be trained with masks to support multiple functions (policy, forward/reverse dynamics, video generation).
Unified World Models: Coupling Video and Action Diffusion for Pretraining on Large Robotic Datasets
Chuning Zhu, Abhishek Gupta
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelWorld ModelVideoSequential
🎯 What it does: Propose Unified World Models (UWM), a unified diffusion model framework capable of simultaneously learning actions, future observations, inverse dynamics, and video generation within the same network, enabling pre-training and fine-tuning of robot policies.
Users and Wizards in Conversations: How WoZ Interface Choices Define Human-Robot Interactions
Ekaterina Torubarova, Andre Pereira
Robotic IntelligenceVideoTextMultimodalityAudio
🎯 What it does: The study compares the effects of three Wizard-of-Oz (WoZ) interfaces (restricted perceptual GUI, unlimited perceptual GUI, VR remote perception) on user experience, robot dialogue fluency, and operator workload during ethical dilemma conversations with the Furhat robot.
V-HOP: Visuo-Haptic 6D Object Pose Tracking
Hongyu Li, Srinath Sridhar
Object TrackingPose EstimationTransformerMultimodalityPoint Cloud
🎯 What it does: Propose a 6D object pose tracking framework V-HOP that integrates visual and tactile information.
Verti-Bench: A General and Scalable Off-Road Mobility Benchmark for Vertically Challenging Terrain
Tong Xu, Xuesu Xiao
Data SynthesisAutonomous DrivingImageBenchmark
🎯 What it does: Developed a high-fidelity multi-physics engine benchmark called Verti-Bench, providing 100 vertical rugged terrain environments and 1,000 navigation tasks to evaluate off-road mobility performance.
Vib2Move: In-hand Object Reconfiguration via Fingertip Micro-vibrations
Xili Yi, Nima Fazeli
Pose EstimationRobotic IntelligenceImage
🎯 What it does: Achieve repositioning of planar objects in arbitrary orientations within a standard parallel gripper by utilizing fingertip micro-vibration and gravity, proposing vibration-modulated friction, sliding motion model, vision-based closed-loop control, and heuristic subgoal planning, with experimental validation conducted on a real robot.
ViTaSCOPE: Visuo-tactile Implicit Representation for In-hand Pose and Extrinsic Contact Estimation
Jayjun Lee, Nima Fazeli
Pose EstimationRobotic IntelligenceImageMultimodalityPoint CloudMesh
🎯 What it does: This paper proposes ViTaSCOPE, a unified neural implicit representation framework that simultaneously estimates the pose of objects inside the hand and external contact positions using vision and high-resolution tactile shear fields.
Vysics: Object Reconstruction Under Occlusion by Fusing Vision and Contact-Rich Physics
Bibit Bianchini, Michael Posa
GenerationRobotic IntelligenceVideoPoint CloudPhysics Related
🎯 What it does: Propose the Vysics method, which integrates visual tracking with physics-based contact dynamics, generating complete geometric and inertial models of objects using RGBD videos and robot proprioception.
You Only Teach Once: Learn One-Shot Bimanual Robotic Manipulation from Video Demonstrations
Huayi Zhou, Kui Jia
Robotic IntelligenceConvolutional Neural NetworkDiffusion modelVideoPoint Cloud
🎯 What it does: This paper proposes the YOTO framework, which extracts fine-grained hand motion trajectories from a single human demonstration video (using stereo cameras) and injects them into dual-arm robots, enabling one-time teaching to achieve long-duration bimanual collaborative operations; through keyframe extraction, motion masks, and automated demonstration augmentation with auto-rolling/geometric transformations, it rapidly generates massive training samples; combined with a customized bimanual diffusion strategy BiDP (employing SIM(3)-equivariant networks, point cloud observations, and discrete key poses), it achieves efficient vision-motion imitation learning.
π₀: A Vision-Language-Action Flow Model for General Robot Control
Kevin Black, Ury Zhilinsky
Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelFlow-based ModelMultimodality
🎯 What it does: Propose a vision-language-action model To for general robot control, based on pre-trained vision-language models (VLM) and combined with flow-matching action experts.