
Motion Planning
Motion planning for mobile manipulation has progressed significantly in the last few years. Our focus in this area has been on generation of fast, efficient and safe motion plans in unstructured environments. To this end, we have worked on three types of motion planners, each of which has its advantages in different situations. We have also worked on infrastructure for benchmarking all these types of motion planners to determine the best type of planner to use in different situations.
Randomized Planners - Arguably the best set of planners to use in unstructured environments because of their performance and ease of implementation. They are currently our "goto" planners for manipulation tasks like pick and place. We have also done work on improving the performance of randomized planners when planning with constraints by constructing, a priori, an approximate representation of the constraint manifold for use online.
Search-based (A*) like Planners - Traditionally used for lower dof motion planning problems (e.g mobile base navigation), we were able to adapt them for use in higher-dimensional planning problems like manipulation as well. Their advantage is that they are deterministic and provide guarantees on the sub-optimality bounds of the solutions that are found. In addition, the paths found by these types of planners tend to be consistent, i.e. similar types of start and goal problems result in similar types of solutions. When combined with caching, we have shown that search-based planners can catch up with their randomized counterparts in terms of performance, allowing them to be used even for high-dimensional problems like whole-body manipulation.
Trajectory Optimization-based Planners - Collision avoidance is often not the only constraint that a motion planning problem will involve. Smoothness, torque and dynamic constraints also need to be addressed. We developed trajectory optimization based planners like STOMP to address such problems. STOMP can take into account arbitrary types of constraints and can give back trajectories that are directly executable on a robot with minimal post-processing. We have used STOMP for pick and place tasks and for dynamic balancing tasks with the PR2 robot.
Related Publications
Demonstrations
Motion planning for mobile manipulation has progressed significantly in the last few years. Our focus in this area has been on generation of fast, efficient and safe motion plans in unstructured environments. To this end, we have worked on three types of motion planners, each of which has its advantages in different situations. We have also worked on infrastructure for benchmarking all these types of motion planners to determine the best type of planner to use in different situations.
Randomized Planners - Arguably the best set of planners to use in unstructured environments because of their performance and ease of implementation. They are currently our "goto" planners for manipulation tasks like pick and place. We have also done work on improving the performance of randomized planners when planning with constraints by constructing, a priori, an approximate representation of the constraint manifold for use online.
Search-based (A*) like Planners - Traditionally used for lower dof motion planning problems (e.g mobile base navigation), we were able to adapt them for use in higher-dimensional planning problems like manipulation as well. Their advantage is that they are deterministic and provide guarantees on the sub-optimality bounds of the solutions that are found. In addition, the paths found by these types of planners tend to be consistent, i.e. similar types of start and goal problems result in similar types of solutions. When combined with caching, we have shown that search-based planners can catch up with their randomized counterparts in terms of performance, allowing them to be used even for high-dimensional problems like whole-body manipulation.
Trajectory Optimization-based Planners - Collision avoidance is often not the only constraint that a motion planning problem will involve. Smoothness, torque and dynamic constraints also need to be addressed. We developed trajectory optimization based planners like STOMP to address such problems. STOMP can take into account arbitrary types of constraints and can give back trajectories that are directly executable on a robot with minimal post-processing. We have used STOMP for pick and place tasks and for dynamic balancing tasks with the PR2 robot.
Related Publications
- "Learning to Plan for Constrained Manipulation from Demonstrations", Phillips, Mike., Hwang, Victor., Chitta, Sachin., and Likhachev, Maxim, Robotics Science and Systems (RSS), Berlin, Germany, 2013 [PDF]
- "Anytime Incremental Planning with E-Graphs", Phillips, Mike., Dornbush, Andrew., Chitta, Sachin., and Likhachev, Maxim, IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, 2013 [PDF]
- "E-Graphs: Bootstrapping Planning with Experience Graphs", Phillips, Michael., Cohen, Benjamin., Chitta, Sachin., and Likhachev, Maxim, Robotics Science and Systems (RSS), Sydney, Australia, 2012 [PDF]
- "Search-based Planning for Dual-arm Manipulation with Upright Orientation Constraints", Cohen, Benjamin., Chitta, Sachin., and Likhachev, Maxim, IEEE International Conference on Robotics and Automation, Minneapolis, Minnesota, 2012 [PDF]
- "A Generic Infrastructure for Benchmarking Motion Planners", Cohen, Benjamin., Sucan, Ioan A.., and Chitta, Sachin, IEEE/RSJ International Conference on Intelligent Robots and Systems, Vilamoura, Algarve, Portugal, 2012 [PDF]
- "Motion Planning With Constraints Using Configuration Space Approximations", Sucan, Ioan A.., and Chitta, Sachin, IEEE/RSJ International Conference on Intelligent Robots and Systems, Vilamoura, Algarve, Portugal, 2012 [PDF]
- "STOMP: Stochastic Trajectory Optimization for Motion Planning", Kalakrishnan, Mrinal., Chitta, Sachin., Theodorou, Evangelos., Pastor, Peter., and Schaal, Stefan, International Conference on Robotics and Automation, Shanghai, China, 2011 [PDF]
- "Search-Based Planning for Manipulation with Motion Primitives", Cohen, Benjamin., Chitta, Sachin., and Likhachev, Maxim, IEEE International Conference on Robotics and Automation, Anchorage, Alaska, 2010 [PDF]
Demonstrations
Ioan Sucan's early demonstrations with OMPL and the PR2. This was the start of Arm Navigation and OMPL
|
Ben Cohen's Dual-Arm Planning Using SBPL
|
Mrinal Kalakrishnan's Willow Garage Video on STOMP
|
Tobias Kunz's Willow Garage Intern Video
|