UC Berkeley Personal Robotics-w-updated-vision-slide · 2010. 7. 6. · Hartmann, Michael Jordan,...
Transcript of UC Berkeley Personal Robotics-w-updated-vision-slide · 2010. 7. 6. · Hartmann, Michael Jordan,...
UC Berkeley Personal Robotics
Presenter: Arjun Singh
Team leads: Pieter Abbeel, Ruzena Bajcsy, Trevor Darrell, Ken Goldberg, Bjoern
Hartmann, Michael Jordan, Dan Klein, Jitendra Malik, Stuart Russell, Claire Tomlin
Main themes
• Hierarchical planning
• Perception
• Manipulation of deformable objects
– End-to-end laundry– End-to-end laundry
• Learning from demonstrations
– Assembly
Hierarchical Planning for
Mobile Manipulation
• Objective 1: speed & scale up
– Apply recent results in efficient, guaranteed-
optimal "angelic” hierarchical planning • Commit to provably optimal (or “good-enough”) high-level plans
• Prune provably suboptimal high-level plans
[Marthi, Russell & Wolfe 2007, 2008]
• Objective 2: uncertainty
– Incorporate partial observability,
information-gathering, hierarchical plan repair
Hierarchical planning example
Preliminary results
[Jason Wolfe, Bhaskara Marthi, Stuart Russell]
Perception
• Visual object recognition
• Pose regression for grasping
• Detecting people
New Local Features for Visual Object
RecognitionKarayev, Fritz, Fidler, Bradski, Darrell
• Handle transparency
• Learn higher-level representations
• Statistically modeled• Statistically modeled
Advanced Methods for Pose
Regression and GraspingSong, Gu, Malik, Darrell
• Category-level pose estimation using latent HOG descriptors
• Discriminatively-trained variant (Gu)
• Combine with local grasp point detection for better grasping in
cases where category-level knowledge is relevant
Detecting peopleMalik, Darrell, Bajcsy
Manipulation of deformable objectsAbbeel
• Large configuration spaces
– Perceptual challenges: estimation of configuration and/or grasp points
– Manipulation challenges: planning in high-dimensional spacesdimensional spaces
• Current directions
– Visual and manipulation primitives:
• Corner detection, edge tracing, bottom most point detection, …
– Simple “worst-case” simulators
• Practical landmark goal: end-to-end laundry
Preliminary results
[autonomous, 100x]
Learning from demonstrationsAbbeel, Goldberg, Hartmann
• Programming robots can be time-consuming
• Often significantly faster and simpler to
provide demonstrations
• Application areas:• Application areas:
– Robot locomotion
– Autonomous helicopter flight
– Manipulation
• Practical landmark goal: teaching basic assembly
Learning from demonstrations
Summary
• Hierarchical planning
• Perception
Pranav Shah
• Deformable objects
• Learning from demonstrations
Hyun Oh Song
Marco
Cusumano-Towner
Arjun
Singh
Shervin JavdaniJudy Hoffman