A5100566831- Robot Manipulation and Learning
- Reinforcement Learning in Robotics
- 3D Shape Modeling and Analysis
- Robotic Path Planning Algorithms
- Direction-of-Arrival Estimation Techniques
- Neural Networks Stability and Synchronization
- Manufacturing Process and Optimization
- Soft Robotics and Applications
- Advanced Vision and Imaging
- Stability and Control of Uncertain Systems
- Multimodal Machine Learning Applications
- Adhesion, Friction, and Surface Interactions
- Robotic Mechanisms and Dynamics
- Human Pose and Action Recognition
- Speech and Audio Processing
- Advanced Adaptive Filtering Techniques
- Stability and Controllability of Differential Equations
- Advanced Data Processing Techniques
- Dynamics and Control of Mechanical Systems
- Muscle activation and electromyography studies
- Advanced Sensor and Control Systems
- 3D Surveying and Cultural Heritage
- Computational Physics and Python Applications
- Modular Robots and Swarm Intelligence
- Teleoperation and Haptic Systems
National University of Singapore
2023-2024
University at Albany, State University of New York
2022-2024
Shandong University of Science and Technology
2019-2024
Nanjing University of Science and Technology
2024
Yancheng Institute of Technology
2023
Shanghai Jiao Tong University
2023
Bank of Canada
2023
Lanzhou Jiaotong University
2023
Wuhan Textile University
2023
Stanford University
2017-2022
We aim to endow a robot with the ability learn manipulation concepts that link natural language instructions motor skills. Our goal is single multi-task policy takes as input instruction and an image of initial scene outputs motion trajectory achieve specified task. This has generalize over different environments. insight we can approach this problem through learning from demonstration by leveraging large-scale video datasets humans performing actions. Thereby, avoid more time-consuming...
We introduce a large-scale 3D shape understanding benchmark using data and annotation from ShapeNet object database. The consists of two tasks: part-level segmentation shapes reconstruction single view images. Ten teams have participated in the challenge best performing outperformed state-of-the-art approaches on both tasks. A few novel deep learning architectures been proposed various representations report techniques used by each team corresponding performances. In addition, we summarize...
We aim to endow a robot with the ability learn manipulation concepts that link natural language instructions motor skills.Our goal is single multi-task policy takes as input instruction and an image of initial scene outputs motion trajectory achieve specified task.This has generalize over different environments.Our insight we can approach this problem through Learning from Demonstration by leveraging large-scale video datasets humans performing actions.Thereby, avoid more time-consuming...
Given two consecutive RGB-D images, we propose a model that estimates dense 3D motion field, also known as scene flow. We take advantage of the fact in robot manipulation scenarios, scenes often consist set rigidly moving objects. Our jointly (i) segmentation into an unknown but finite number objects, (ii) trajectories these objects and (iii) object employ hourglass, deep neural network architecture. In encoding stage, RGB depth images undergo spatial compression correlation. decoding...
The ability to perform in-hand manipulation still remains an unsolved problem; having this capability would allow robots sophisticated tasks requiring repositioning and reorienting of grasped objects. In work, we present a novel non-anthropomorphic robot grasper with the manipulate objects by means active surfaces at fingertips. Active are achieved spherical rolling fingertips two degrees freedom (DoF) - pivoting motion for surface reorientation continuous moving object. A further DoF is in...
Humans possess an extraordinary ability to understand and execute complex manipulation tasks by interpreting abstract instruction manuals. For robots, however, this capability remains a substantial challenge, as they cannot interpret instructions translate them into executable actions. In paper, we present Manual2Skill, novel framework that enables robots perform assembly guided high-level manual instructions. Our approach leverages Vision-Language Model (VLM) extract structured information...
Autonomous part assembly is a challenging yet crucial task in 3D computer vision and robotics. Analogous to buying an IKEA furniture, given set of parts that can assemble single shape, intelligent agent needs perceive the geometry, reason propose pose estimations for input parts, finally call robotic planning control routines actuation. In this paper, we focus on estimation subproblem from side involving geometric relational reasoning over geometry. Essentially, generative predict 6-DoF...
To achieve a successful grasp, gripper attributes such as its geometry and kinematics play role important the object geometry. The majority of previous work has focused on developing grasp methods that generalize over novel but are specific to certain robot hand. We propose UniGrasp, an efficient data-driven synthesis method considers both inputs. UniGrasp is based deep neural network architecture selects sets contact points from input point cloud object. proposed model trained large dataset...
Learning contact-rich, robotic manipulation skills is a challenging problem due to the high-dimensionality of state and action space as well uncertainty from noisy sensors inaccurate motor control. To combat these factors achieve more robust manipulation, humans actively exploit contact constraints in environment. By adopting similar strategy, robots can also manipulation. In this paper, we enable robot autonomously modify its environment thereby discover how ease skill learning....
Building general-purpose robots to perform a diverse range of tasks in large variety environments the physical world at human level is extremely challenging. According [1], it requires robot learning be sample-efficient, generalizable, compositional, and incremental. In this work, we introduce systematic framework called SAGCI-system towards achieving these above four requirements. Our system first takes raw point clouds gathered by camera mounted on robot's wrist as inputs produces initial...
Model-based reinforcement learning (MBRL) is recognized with the potential to be significantly more sample efficient than model-free RL.How an accurate model can developed automatically and efficiently from raw sensory inputs (such as images), especially for complex environments tasks, a challenging problem that hinders broad application of MBRL in real world.In this work, we propose sensingaware model-based system called SAM-RL.Leveraging differentiable physics-based simulation rendering,...
Robotic dexterous grasping is a challenging problem due to the high degree of freedom (DoF) and complex contacts multi-fingered robotic hands. Existing deep re-inforcement learning (DRL) based methods leverage human demonstrations reduce sample complexity dimensional action space with grasping. However, less attention has been paid hand-object interaction representations for high-level generalization. In this paper, we propose novel geometric spatial representation, named DexRep, capture...
We present ClothesNet: a large-scale dataset of 3D clothes objects with information-rich annotations. Our consists around 4400 models covering 11 categories annotated features, boundary lines, and keypoints. ClothesNet can be used to facilitate variety computer vision robot interaction tasks. Using our dataset, we establish benchmark tasks for perception, including classification, line segmentation, keypoint detection, develop simulated environments robotic tasks, rearranging, folding,...