University of Pennsylvania. From Perception and Reasoning to Grasping

Transcription

1 University of Pennsylvania GRASP LAB PR2GRASP: From Perception and Reasoning to Grasping Led by Maxim Likhachev Kostas Daniilides Vijay Kumar Katherine J. Kuchenbecker Jianbo Shi Daniel D. Lee Mark Yim Camillo Jose Taylor Student Representatives: Mike Phillips Benjamin Cohen Cody Phillips Soonkyum Kim

2 @ GRASP Lab

3 Research Proposal Planning for Navigation in Dynamic Environments Pennochio Project to Establish Telepresence through PR2 Kostas Daniilides Planning and Controls for two-arm Manipulation Jianbo Shi & Maxim Likhachev Visual Localization and Pose Estimation of Objects for Grasping Katherine Kuchenbecker & Maxim Likhachev Tracking People in Cluttered Spaces for Navigation in Dynamic Environment Maxim Likhachev Transferring Natural Handheld Objects between PR2 and Human and other Robots Mark Yim Planning for Autonomous Opening of Spring-Loaded Doors CJ Taylor & Mark Yim Support for Modular End Effectors on PR2 Maxim Likhachev Vijay Kumar & Maxim Likhachev Learning salient features for better perceptual processing by PR2 Dan Lee

4 Planning for Autonomous Opening of Spring-Loaded Doors Maxim Likhachev Motivation: Spring loaded doors are all around us! So the PR2 can use fire exits in case of an emergency Help save old people & children Save Itself Goal: A principled approach to: plan a complete motion to open a spring loaded door using the arm(s) and base of the robot collision-free trajectory for the complete robot body can be used on any door or cabinet provides completeness & suboptimality guarantees

5 Planning for Autonomous Opening of Spring-Loaded Doors Benjamin Cohen, Maxim Likhachev Planning for Autonomous Door Opening with a Mobile Manipulator Sachin Chitta, Benjamin Cohen, Maxim Likhachev ICRA 2010

6 Transferring Natural Handheld Objects between PR2 and Human and other Robots Joe Romano, Katherine Kuchenbecker In order to move robots into the home they need to interact naturally with people. Passing objects to a human, and receiving objects passed from a human, constitute a basic but important robot skill. It is important to follow social conventions: passing at the right speed passing or accepting when the human is ready...without violating the mechanical constraints of the interaction: letting go when the human is ready applying and releasing stable grasps We plan to use tactile cues, such as contact acceleration signals, to detect important events in the interaction.

7 Pennochio Project to Establish Telepresence through PR2 CJ Taylor, Mark Yim + Explore using the PR2 for immersive teleoperation Leverage inexpensive Head Mounted Displays Slave Pan Tilt head to real-time motion capture Map motion of user onto motion of base and arms Develop hand held manipulative units to control motion of the grippers and to provide haptic feedback Archive motion and sensor data for use in teaching by example systems Explore use of low cost Markerless Motion capture Some Related Projects: + MARIONET UT Austin Interaction Lab - USC

8 Support for Modular End Effectors on PR2 Mark Yim Goal: Change end effectors to suit tasks Issues: What types of end effectors would be useful? Gravity compensation in arm requires same moment/mass from end-effector Low level software requires modification to handle change in hardware state High level ROS PR2 interface needs to support endeffector types

9 Support for Modular End Effectors on PR2 Mark Yim

10 Planning for Navigation in Dynamic Environments Mike Phillips, Maxim Likhachev Motivation Dynamic Obstacles are all around us! Household tasks require navigation around people and pets Problem Given: Map of the static environment Pose of the robot and goal Predicted trajectories of dynamic obstacles Output: A time parameterized path that gets the robot safely to the goal

11 Planning for Navigation in Dynamic Environments Mike Phillips, Maxim Likhachev A common approach is to treat dynamic obstacles as static and replan often This is fast but lacks optimality and completeness Another approach adds a time dimension to the search space Is optimal and complete but very slow due to increased dimensionality Our approach exploits the idea that dynamic obstacles generally occupy a small fraction of the environment We build planners that only use a time dimension when relevant Most of the space is free of dynamic obstacles This allows for fast planning times while still guaranteeing optimality and completeness

12 Planning for Navigation in Dynamic Environments Mike Phillips, Maxim Likhachev Time-bounded lattice for efficient planning in dynamic environment Aleksandr Kushleyev, Maxim Likhachev ICRA 2009

13 Planning for Navigation in Dynamic Environments Mike Phillips, Maxim Likhachev This planner will be run on the PR2 Planning times are all < 1 sec

14 Tracking People in Cluttered Spaces for Navigation in Dynamic Environment Jianbo Shi, Maxim Likhachev People are all around us! Household tasks require robots to identify and track people Tracking people can become more robust to occlusion if planning is used to predict trajectories people may follow

15 Tracking People in Cluttered Spaces for Navigation in Dynamic Environment Jianbo Shi, Maxim Likhachev

16 Visual Localization and Pose Estimation of Objects for Grasping Cody Phillips, Alex Toshev, Kostas Daniilidis Object recognition/localization 3D pose estimation Most approaches texture-based Exists problem objects Glasses, bottles, cups Shiny, transparent objects Shape-based approach Recognize by shape Learn from 3D model library

17 Visual Localization and Pose Estimation of Objects for Grasping Cody Phillips, Alex Toshev, Kostas Daniilidis Grasping By Shape (Single Image) Extract 2D views from 3D models Compute shape descriptors Hypothesize object and pose Shapes Match? Recover Pose!

18 Visual Localization and Pose Estimation of Objects for Grasping Cody Phillips, Alex Toshev, Kostas Daniilidis Grasping By Shape (Video) Multiple object camera views Combine evidence from views Refine hypothesis space Hypotheses Geometrically Consistent

19 Planning & Controls for two-arm Manipulation Soonkyum Kim, Vijay Kumar, Maxim Likhachev Goal: Perform manipulation tasks that require two arm coordination Challenges: Maintain and control contacts between object and end effector Maintain force closure Rolling/sliding could occur Redundant system

20 Planning & Controls for two-arm Manipulation Soonkyum Kim, Vijay Kumar, Maxim Likhachev Cooperative Quasi-Static Planar Manipulation with Multiple Robots. Quentin J. Lindsey, Michael Shomin, and Vijay Kumar. IDETC 2010

21 Planning & Controls for two-arm Manipulation Soonkyum Kim, Vijay Kumar, Maxim Likhachev A general paradigm for control: Control for trajectory + control for contact Nonlinear feedback controller Planning: Partition of the configuration space Construct compact graph Search-based planning

22 Thank You!