最近總結了基於視覺的機器人抓取的相關論文及代碼，同步於github。根據抓取的類別，基於視覺的機器人抓取可以分爲兩類：2D平面抓取以及6D空間抓取。這個頁面總結了近年來涉及到的各種各樣的方法，其中大部分都使用了深度學習。總結的中文框架結構如下：
0.綜述論文
1. 2D平面抓取
這類方法將抓取，用平面內旋轉的矩形框表示；
1.1基於RGB或者RGB-D的方法
採用基於目標檢測類的方法，直接得到2D抓取位姿；如果有depth，可以增加depth通道；
1.2基於深度圖的方法
這類方法不是端到端進行，而是首先生成候選的抓取位置，之後對候選進行評分，並選擇質量最優的進行抓取；
1.3目標物體定位
如果能夠提前將目標物體檢測或分割出來，則能夠提供一個較好的輸入；
1.3.1 目標2D檢測
1.3.2 目標2D實例分割
2. 6D空間抓取
這類方法將抓取，用空間抓取器的位置和朝向表示；輸入是單視角的點雲數據;
2.1目標物體的3D定位
可以藉助2D定位結果得到對應的depth區域，得到點雲；也可以直接在3D空間進行定位；
2.1.1 3D檢測
這類方法包含3小類，基於RGB，基於RGB+點雲，基於純點雲；主要在無人車領域；
2.1.2 3D實例分割
基於很多3D深度學習網絡；
2.2 物體的6D位姿估計
這類方法適用於已有3D模型的情況；假定已有3D模型上已知抓取位置，知道物體姿態後，就可以進行抓取；
2.2.1基於RGB-D圖像的方法
這類方法主要是結合RGB圖像和Depth圖像，進行特徵提取匹配等；包括四類a.基於對應的方法；b.基於模板的方法；c.基於投票的方法；d.基於迴歸的方法；
2.2.2 基於點雲的方法
一般都需要進行粗匹配得到粗位姿，再使用細匹配如ICP方法得到精確位姿；包括3類：a.基於RANSAC的方法；b.基於3D特徵的方法；c.基於深度學習的方法；
2.3基於深度學習的方法
這類方法適用於不存在3D模型的情況，但可以擁有大量類似物體組成的抓取3D模型庫；
2.3.1從部分視角點雲直接回歸6DoF抓取姿態
2.3.2 抓取支撐能力估計
2.3.3 形狀補全輔助的抓取
2.3.4 深度圖補全
3.抓取遷移
3.1在形狀部分之間遷移
3.2非剛性形狀匹配
3.2.1 非剛性配準
3.2.2 形狀對應
3.3 3D部件分割
4.靈巧手
5.模擬到真實
6.多源信息
7.模仿學習
8.強化學習
9.領域專家
具體相關論文如下：
Vision-based Robotic Grasping: Papers and Codes
According to the kinds of grasp, the methods of vision-based robotic grasping can be roughly divided into two kinds, 2D planar grasp and 6DoF Grasp. This repository summaries these methods in recent years, which utilize deep learning mostly. Before this summary, previous review papers are also reviewed.
0. Review Papers
[arXiv] 2019-A Review of Robot Learning for Manipulation- Challenges, Representations, and Algorithms, [paper]
[arXiv] 2019-Vision-based Robotic Grasping from Object Localization, Pose Estimation, Grasp Detection to Motion Planning: A Review, [paper]
[MTI] 2018-Review of Deep Learning Methods in Robotic Grasp Detection, [paper]
[ToR] 2016-Data-Driven Grasp Synthesis - A Survey, [[paper](Data-Driven Grasp Synthesis - A Survey)]
[RAS] 2012-An overview of 3D object grasp synthesis algorithms - A Survey, [paper]
1. 2D Planar Grasp
Grasp Representation:
The grasp is represented as an oriented 2D box, and the grasp is constrained from one direction.
1.1 RGB or RGB-D based methods
This kind of methods directly regress the oriented 2D box from RGB or RGB-D images. When using RGB-D images, the depth image is regarded as an another channel, which is similar with RGB-based methods.
2019:
[arXiv] Efficient Fully Convolution Neural Network for Generating Pixel Wise Robotic Grasps With High Resolution Images, [paper]
[arXiv] A Single Multi-Task Deep Neural Network with Post-Processing for Object Detection with Reasoning and Robotic Grasp Detection, [paper]
[arXiv] SilhoNet: An RGB Method for 6D Object Pose Estimation, [paper]
[ICRA] Multi-View Picking: Next-best-view Reaching for Improved Grasping in Clutter, [paper] [code]
2018:
[arXiv] Real-Time, Highly Accurate Robotic Grasp Detection using Fully Convolutional Neural Networks with High-Resolution Images, [paper]
[arXiv] Real-world Multi-object, Multi-grasp Detection, [paper]
[ICRA] Robotic Pick-and-Place of Novel Objects in Clutter with Multi-Affordance Grasping and Cross-Domain Image Matching, [paper] [code]
2017:
[IROS] Robotic Grasp Detection using Deep Convolutional Neural Networks, [paper]
2016:
[ICRA] Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours, [paper]
2015:
[ICRA] Real-time grasp detection using convolutional neural networks, [paper] [code]
2014:
[IJRR] Deep Learning for Detecting Robotic Grasps, [paper]
Datasets:
Cornell dataset, the dataset consists of 1035 images of 280 different objects.
1.2 Depth-based methods
This kind of methods utilized an indirectly way to obtain the grasp pose, which contains grasp candidate generation and grasp quality evaluation. The candidate grasp with the highly score will be selected as the final grasp.
2019:
[ICRA] Mechanical Search: Multi-Step Retrieval of a Target Object Occluded by Clutter, [paper]
[ICRA] MetaGrasp: Data Efficient Grasping by Affordance Interpreter Network, [paper]
[ICIRS] GlassLoc: Plenoptic Grasp Pose Detection in Transparent Clutter, [paper]
2018:
[RSS] Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach, [paper]
[BMVC] EnsembleNet Improving Grasp Detection using an Ensemble of Convolutional Neural Networks, [paper]
2017:
[RSS] Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics, [paper] [code]
Dataset:
Dex-Net, a synthetic dataset of 6.7 million point clouds, grasps, and robust analytic grasp metrics generated from thousands of 3D models.
Jacquard Dataset, Jacquard: A Large Scale Dataset for Robotic Grasp Detection” in IEEE International Conference on Intelligent Robots and Systems, 2018, [paper]
1.3 Target object localization in 2D
In order to provide a better input to compute the oriented 2D box, or generate the candidates, the targe object’s mask should be computed. The current deep learning-based 2D detection or 2D segmentation methods could assist.
1.3.1 2D detection:
2019:
[arXiv] Object Detection in 20 Years A Survey, [paper]
2018:
[arXiv] YOLOv3: An Incremental Improvement, [paper] [code]
2016:
[CVPR] You only look once: Unified, real-time object detection, [paper] [code]
[TPAMI] Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, [paper] [code]
[ECCV] SSD: Single Shot MultiBox Detector, [paper] [code]
1.3.2 2D segmentation:
2019:
[CASE] Deep Workpiece Region Segmentation for Bin Picking, [paper]
2017:
[ICCV] Mask r-cnn, [paper] [code]
[IROS] SegICP: Integrated Deep Semantic Segmentation and Pose Estimation, [paper]
2. 6DoF Grasp
Grasp Representation:
The grasp is represented as 6DoF pose in 3D domain, and the gripper can grasp the object from various angles. The input to this task is 3D point cloud from RGB-D sensors, and this task contains two stages. In the first stage, the targe object should be extracted from the scene. In the second stage, if there exist an existing 3D model, the 6D pose of the object could be computed. If there exists no 3D models, the 6DoF grasp pose will be computed from some other methods.
2.1 Target object extraction in 3D
The staightforward way is to conduct 2D dection or segmentation, and utilize the point cloud from the corresponding depth area. This part is already related in section 1.3. In the following, only 3D detection and 3D instance segmentation will be summarized.
2.1.1 3D detection
This kind of methods can be divided into three kinds: RGB-based methods, point cloud-based methods, and fusion methods which consume images and point cloud. Most of these works are focus on autonomous driving.
a. RGB-based methods
Most of this kind of methods estimate depth images from RGB images, and then conduct 3D detection.
2019:
[IROS] Look Further to Recognize Better: Learning Shared Topics and Category-Specific Dictionaries for Open-Ended 3D Object Recognition, [paper]
[arXiv] Task-Aware Monocular Depth Estimation for 3D Object Detection, [paper]
[CVPR] Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving, [paper] [code]
[AAAI] MonoGRNet: A Geometric Reasoning Network for 3D Object Localization, [paper] [code]
[ICCV] Accurate Monocular 3D Object Detection via Color-Embedded 3D Reconstruction for Autonomous Driving, [paper]
[ICCV] M3D-RPN: Monocular 3D Region Proposal Network for Object Detection, [paper]
[ICCVW] Monocular 3D Object Detection with Pseudo-LiDAR Point Cloud, [paper]
[arXiv] Monocular 3D Object Detection and Box Fitting Trained End-to-End Using Intersection-over-Union Loss, [paper]
[arXiv] Monocular 3D Object Detection via Geometric Reasoning on Keypoints, [paper]
b. Point cloud-based methods
This kind of methods purely utilize the 3D point cloud data.
2019:
[NeurIPSW] Patch Refinement – Localized 3D Object Detection, [paper]
[CoRL] End-to-End Multi-View Fusion for 3D Object Detection in LiDAR Point Clouds, [paper]
[ICCV] Deep Hough Voting for 3D Object Detection in Point Clouds, [paper] [code]
[arXiv] Part-A2 Net: 3D Part-Aware and Aggregation Neural Network for Object Detection from Point Cloud, [paper]
[ICCV] STD: Sparse-to-Dense 3D Object Detector for Point Cloud, [paper]
[CVPR] PointPillars: Fast Encoders for Object Detection from Point Clouds, [paper]
[arXiv] StarNet: Targeted Computation for Object Detection in Point Clouds, [paper]
2018:
[CVPR] PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud, [paper] [code]
[CVPR] PIXOR: Real-time 3D Object Detection from Point Clouds, [paper] [code]
[ECCVW] Complex-YOLO: Real-time 3D Object Detection on Point Clouds, [paper] [code]
[ECCVW] YOLO3D: End-to-end real-time 3D Oriented Object Bounding Box Detection from LiDAR Point Cloud, [paper]
c. Fusion methods
This kind of methods utilize both rgb images and depth images/point clouds. There exist early fusion methods, late fusion methods, and dense fusion methods.
2019:
[arXiv] Adaptive and Azimuth-Aware Fusion Network of Multimodal Local Features for 3D Object Detection, [paper]
[arXiv] Frustum VoxNet for 3D object detection from RGB-D or Depth images, [paper]
[CVPR] Multi-Task Multi-Sensor Fusion for 3D Object Detection, [paper]
2018:
[CVPR] Frustum PointNets for 3D Object Detection from RGB-D Data, [paper] [code]
[ECCV] Deep Continuous Fusion for Multi-Sensor 3D Object Detection, [paper]
[IROS] Joint 3D Proposal Generation and Object Detection from View Aggregation, [paper] [code]
[CVPR] PointFusion: Deep Sensor Fusion for 3D Bounding Box Estimation, [paper]
[ICRA] A General Pipeline for 3D Detection of Vehicles, [paper]
2017:
[CVPR] Multi-View 3D Object Detection Network for Autonomous Driving, [paper] [code]
2.1.2 3D instance segmentation
2019:
[arXiv] A survey on Deep Learning Advances on Different 3D Data Representations, [paper]
[CoRL] The Best of Both Modes: Separately Leveraging RGB and Depth for Unseen Object Instance Segmentation, [paper] [code]
2018:
[CVPR] SPLATNet: Sparse Lattice Networks for Point Cloud Processing, [paper] [code]
[arXiv] PointSIFT: A SIFT-like Network Module for 3D Point Cloud Semantic Segmentation, [paper]
2017:
[CVPR] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation, [paper] [code]
[NeurIPS] PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space, [paper] [code]
2.2 6D object pose estimation (Exist 3D models)
2.2.1 RGB-D based methods
This kind of methods can be divided into four kinds, which are corresponding-based methods, template-based methods, voting-based methods and regression-based methods.
[ECCVW] A Summary of the 4th International Workshop on Recovering 6D Object Pose, [paper]
a. Corresponding-based methods
2019:
[CVPR] Segmentation-driven 6D Object Pose Estimation, [paper]
2018:
[arXiv] Estimating 6D Pose From Localizing Designated Surface Keypoints, [paper]
2017:
[ICRA] 6-DoF Object Pose from Semantic Keypoints, [paper]
2012:
[3DIMPVT] 3D Object Detection and Localization using Multimodal Point Pair Features, [paper]
b. Template-based methods
2019:
[arXiv] Real-time Background-aware 3D Textureless Object Pose Estimation, [paper]
2012:
[ACCV] Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes, [paper]
c. Voting-based methods
2018:
[TPAMI] Robust 3D Object Tracking from Monocular Images Using Stable Parts, [paper]
2014:
[ECCV] Learning 6d object pose estimation using 3d object coordinate, [paper]
[ECCV] Latent-class hough forests for 3d object detection and pose estimation, [paper]
d. Regression-based methods
1) Directly way
2019:
[arXiv] Learning Object Localization and 6D Pose Estimation from Simulation and Weakly Labeled Real Images, [paper]
[ICHR] Refining 6D Object Pose Predictions using Abstract Render-and-Compare, [paper]
[CVPR] Normalized Object Coordinate Space for Category-Level 6D Object Pose and Size Estimation, [paper] [code]
[CVPR] Densefusion: 6d object pose estimation by iterative dense fusion, [paper] [code]
[arXiv] Deep-6dpose: recovering 6d object pose from a single rgb image, [paper]
2018:
[ECCV] Implicit 3D Orientation Learning for 6D Object Detection From RGB Images, [paper] [code]
[ECCV] DeepIM:Deep Iterative Matching for 6D Pose Estimation
[paper] [code]
[RSS] Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes, [paper] [code]
[IROS] Robust 6D Object Pose Estimation in Cluttered Scenes using Semantic Segmentation and Pose Regression Networks, [paper]
2017:
[ICCV] SSD-6D: Making rgb-based 3d detection and 6d pose estimation great again, [paper] [code]
2) Indirectly way (Firstly regress feature points and use PnP methods)
2019:
[CVPR] PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation, [paper] [code]
2018:
[CVPR] Real-time seamless single shot 6d object pose prediction, [paper] [code]
2017:
[ICCV] BB8: a scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth, [paper]
Datasets:
YCB Datasets: The YCB Object and Model Set: Towards Common Benchmarks for Manipulation Research, IEEE International Conference on Advanced Robotics (ICAR), 2015 [paper]
T-LESS Datasets: T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-less Objects, IEEE Winter Conference on Applications of Computer Vision (WACV), 2017 [paper]
2.2.2 3D point cloud
The partial-view point cloud will be aligned to the complete shape in order to obtain the 6D pose. Generally, coarse registration should be conduct firstly to provide an intial alignment, and dense registration methods like ICP (Iterative Closest Point) will be conducted to obtain the final 6D pose.
a. Ransac-based methods
2014:
[SGP] Super 4PCS Fast Global Pointcloud Registration via Smart Indexing, [paper] [code]
b. 3D feature-based methods
2017:
[CVPR] 3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions, [paper] [code]
2016:
[arXiv] Lessons from the Amazon Picking Challenge, [paper]
[arXiv] Team Delft’s Robot Winner of the Amazon Picking Challenge 2016, [paper]
c. Deep learning-based methods
2019:
[arXiv] Deep Closest Point: Learning Representations for Point Cloud Registration, [paper] [code]
[arXiv] PCRNet: Point Cloud Registration Network using PointNet Encoding, [paper]
2017:
[ICRA] Multi-view Self-supervised Deep Learning for 6D Pose Estimation in the Amazon Picking Challenge, [paper] [code]
2.3 Deep learning-based methods (No existing 3D models)
In this situation, there exist no 3D models, an the 6-DoF grasps are estimated from available partial data. This can be implemented by directly estimating from partial view point cloud, or indirectly estimating after shape completion.
2.3.1 Estimating 6-DoF grasps from partial view point cloud
2019:
[arXiv] Learning to Generate 6-DoF Grasp Poses with Reachability Awareness, [paper]
[CoRL] S4G: Amodal Single-view Single-Shot SE(3) Grasp Detection in Cluttered Scenes, [paper]
[ICCV] 6-DoF GraspNet: Variational Grasp Generation for Object Manipulation, [paper]
[ICRA] PointNetGPD: Detecting Grasp Configurations from Point Sets, [paper] [code]
2017:
[IJRR] Grasp Pose Detection in Point Clouds, [paper] [code]
2.3.2 Grasp affordance
2019:
[IROS] Learning Grasp Affordance Reasoning through Semantic Relations, [paper]
[arXiv] Automatic pre-grasps generation for unknown 3D objects, [paper]
[IECON] A novel object slicing based grasp planner for 3D object grasping using underactuated robot gripper, [paper]
2018:
[arXiv] Workspace Aware Online Grasp Planning, [paper]
2.3.3 Shape completion assisted grasp
2019:
[arXiv] ClearGrasp- 3D Shape Estimation of Transparent Objects for Manipulation, [paper]
[arXiv] kPAM-SC: Generalizable Manipulation Planning using KeyPoint Affordance and Shape Completion, [paper] [code]
[arXiv] Data-Efficient Learning for Sim-to-Real Robotic Grasping using Deep Point Cloud Prediction Networks, [paper]
[arXiv] Inferring Occluded Geometry Improves Performance when Retrieving an Object from Dense Clutter, [paper]
[IROS] Robust Grasp Planning Over Uncertain Shape Completions, [paper]
[arXiv] Multi-Modal Geometric Learning for Grasping and Manipulation, [paper]
2018:
[ICRA] Learning 6-DOF Grasping Interaction via Deep Geometry-aware 3D Representations, [paper]
[IROS] 3D Shape Perception from Monocular Vision, Touch, and Shape Priors, [paper]
2016:
[IROS] Shape Completion Enabled Robotic Grasping, [paper]
2.3.4 Depth completion
2019:
[ICCV] Depth Completion from Sparse LiDAR Data with Depth-Normal Constraints, [paper]
[arXiv] Image-based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era, [paper]
[arXiv] Real-time Vision-based Depth Reconstruction with NVidia Jetson, [paper]
[IROS] Self-supervised 3D Shape and Viewpoint Estimation from Single Images for Robotics, [paper]
[arXiv] Mesh R-CNN, [paper]
2018:
[NeurIPS] Learning to Reconstruct Shapes from Unseen Classes, [paper] [code]
[ECCV] Learning Shape Priors for Single-View 3D Completion and Reconstruction, [paper] [code]
[CVPR] Deep Depth Completion of a Single RGB-D Image, [paper] [code]
3. Grasp Transfer
3.1 Grasp transfer between shape parts
2019:
[arXiv] Using Synthetic Data and Deep Networks to Recognize Primitive Shapes for Object Grasping, [paper]
[ICRA] Transferring Grasp Configurations using Active Learning and Local Replanning, [paper]
2017:
[AIP] Fast grasping of unknown objects using principal component analysis, [paper]
2015:
[RAS] Category-based task specific grasping, [paper]
3.2 Non-rigid shape matching
3.2.1 Non-rigid registration
2018:
[RAL] Transferring Category-based Functional Grasping Skills by Latent Space Non-Rigid Registration, [paper]
[RAS] Learning Postural Synergies for Categorical Grasping through Shape Space Registration, [paper]
[RAS] Autonomous Dual-Arm Manipulation of Familiar Objects, [paper]
3.2.2 Shape correspondence
2019:
[arXiv] Unsupervised cycle-consistent deformation for shape matching, [paper]
[arXiv] ZoomOut: Spectral Upsampling for Efficient Shape Correspondence, [paper]
3.3 3D part segmentation
2019:
[CVPR] PartNet: A Recursive Part Decomposition Network for Fine-grained and Hierarchical Shape Segmentation, [paper] [code]
[C&G] Autoencoder-based part clustering for part-in-whole retrieval of CAD models, [paper]
4. Deterous Grippers
2019:
[ScienceRobotics] On the choice of grasp type and location when handing over an object, [paper]
[arXiv] Solving Rubik’s Cube with a Robot Hand, [paper]
[IJARS] Fast geometry-based computation of grasping points on three-dimensional point clouds, [paper] [code]
[arXiv] Learning better generative models for dexterous, single-view grasping of novel objects, [paper]
[arXiv] DexPilot: Vision Based Teleoperation of Dexterous Robotic Hand-Arm System, [paper]
[arXiv] Generating Grasp Poses for a High-DOF Gripper Using Neural Networks, [paper]
[arXiv] Deep Dynamics Models for Learning Dexterous Manipulation, [paper]
[CVPR] Learning joint reconstruction of hands and manipulated objects, [paper]
[CVPR] H+O: Unified Egocentric Recognition of 3D Hand-Object Poses and Interactions, [paper]
[IROS] Efficient Grasp Planning and Execution with Multi-Fingered Hands by Surface Fitting, [paper]
[arXiv] Efficient Bimanual Manipulation Using Learned Task Schemas, [paper]
[ICRA] High-Fidelity Grasping in Virtual Reality using a Glove-based System, [paper] [code]
5. Simulation to Reality
2019:
[arXiv] Self-supervised 6D Object Pose Estimation for Robot Manipulation, [paper]
[arXiv] Accept Synthetic Objects as Real-End-to-End Training of Attentive Deep Visuomotor Policies for Manipulation in Clutter, [paper]
[RSSW] Generative grasp synthesis from demonstration using parametric mixtures, [paper]
2018:
[RSS] Learning Task-Oriented Grasping for Tool Manipulation from Simulated Self-Supervision, [paper]
[CoRL] Deep Object Pose Estimation for Semantic Robotic Grasping of Household Objects, [paper]
[arXiv] Multi-Task Domain Adaptation for Deep Learning of Instance Grasping from Simulation, [paper]
2017:
[arXiv] Using Simulation and Domain Adaptation to Improve Efficiency of Deep Robotic Grasping, [paper]
6. Multi-source
2019:
[ICRA] Making Sense of Vision and Touch: Self-Supervised Learning of Multimodal Representations for Contact-Rich Tasks, [paper]
[CVPR] ContactDB: Analyzing and Predicting Grasp Contact via Thermal Imaging, [paper] [code]
2018:
[arXiv] Learning to Grasp without Seeing, [paper]
7. Learning from Demonstration
2019:
[IROS] Learning Actions from Human Demonstration Video for Robotic Manipulation, [paper]
[RSSW] Generative grasp synthesis from demonstration using parametric mixtures, [paper]
2018:
[arXiv] Deep Imitation Learning for Complex Manipulation Tasks from Virtual Reality Teleoperation, [paper]
8. Reinforcement Learning
2019:
[arXiv] A Deep Learning Approach to Grasping the Invisible, [paper]
[arXiv] Knowledge Induced Deep Q-Network for a Slide-to-Wall Object Grasping, [paper]
[arXiv] Quantile QT-Opt for Risk-Aware Vision-Based Robotic Grasping, [paper]
[arXiv] Adaptive Curriculum Generation from Demonstrations for Sim-to-Real Visuomotor Control, [paper]
[arXiv] Reinforcement Learning for Robotic Manipulation using Simulated Locomotion Demonstrations, [paper]
[arXiv] Self-Supervised Sim-to-Real Adaptation for Visual Robotic Manipulation, [paper]
[CoRL] ROBEL: Robotics Benchmarks for Learning with Low-Cost Robots, [paper] [code]
[RSS] End-to-End Robotic Reinforcement Learning without Reward Engineering, [paper]
[arXiv] Learning to combine primitive skills: A step towards versatile robotic manipulation, [paper]
[CoRL] A Survey on Reproducibility by Evaluating Deep Reinforcement Learning Algorithms on Real-World Robots, [paper] [code]
[ICCAS] Deep Reinforcement Learning Based Robot Arm Manipulation with Efficient Training Data through Simulation, [paper]
[CVPR] CRAVES: Controlling Robotic Arm with a Vision-based Economic System, [paper] [code]
[Report] A Unified Framework for Manipulating Objects via Reinforcement Learning, [paper]
2018:
[IROS] Learning Synergies between Pushing and Grasping with Self-supervised Deep Reinforcement Learning, [paper] [code]
[CoRL] QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation, [paper]
[arXiv] Deep Reinforcement Learning for Vision-Based Robotic Grasping: A Simulated Comparative Evaluation of Off-Policy Methods, [paper]
[arXiv] Pick and Place Without Geometric Object Models, [paper]
2017:
[arXiv] Deep Reinforcement Learning for Robotic Manipulation-The state of the art, [paper]
2016:
[IJRR] Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning, [paper]
2013:
[IJRR] Reinforcement learning in robotics: A survey, [paper]
9. Experts:
Andy Zeng(Princeton University & Google Brain Robotics): 3D Deep Learning, Robotic Grasping
Andreas ten Pas(Northeastern University): Robotic Grasping, Deep Learning, Simulation-based Planning
Cewu Lu(SJTU): Machine Vision
Charles Ruizhongtai Qi(Facebook Artificial Intelligence Research (FAIR)): 3D Deep Learning
Danfei Xu(Stanford University): Robotics, Computer Vision
Deter Fox(Nvidia & University of Washington): Robotics, Artificial intelligence, State Estimation
Fei-Fei Li(Stanford University): Computer Vision
Guofeng Zhang(ZJU): 3D Vision, SLAM
Hao Su(UC San Diego): 3D Deep Learning
Jeannette Bohg(Stanford University): perception for autonomous robotic manipulation and grasping
Jianping Shi(SenseTime): Computer Vision
Kaiming He(Facebook AI Research (FAIR)): Deep Learning
Kai Xu(NUDT): Graphics, Geometry
Ken Goldberg(UC Berkeley): Robotics
Marc Pollefeys(Microsoft & ETH): Computer Vision
Pascal Fua(INRIA): Computer Vision
Peter K. Allen.(Columbia University): Robotic Grasping, 3-D vision, Modeling, Medical robotics
Raquel Urtasun(Uber ATG & University of Toronto): AI for self-driving cars, Computer Vision, Robotics
Ruigang Yang(Baidu): Computer Vision, Robotics
Sergey Levine(UC Berkeley): Reinforcement Learning
Shuran Song(Columbia University), 3D Deep Learning, Robotics
Silvio Savarese(Stanford University): Computer Vision
Song-Chun Zhu(UCLA): Computer Vision
Thomas Funkhouser(Princeton University): Geometry, Graphics, Shape
Vicient Lepetit(University of Bordeaux): Machine Learning, 3D Vision
Xiaogang Wang(Chinese University of Hong Kong): Deep Learning, Computer Vision
Xiaozhi Chen(DJI): Deep learning
Yan Xinchen(Uber ATG): Deep Representation Learning, Generative Modeling
Yu Xiang(Nvidia): Robotics, Computer Vision