
- RRT Connect is a sampling based planning algorithm used for planning of multi-arm manipulators.
- Instead of naively choosing a tree to extend(switching between the trees), we propose a multi-armed bandit setting where the action space is selecting the tree to extend.
- The agent choosing the tree is given a higher reward if we don’t make significant progress(we want to expand it more) and given a lower reward if it makes significant progress.
- Tested two multi armed algorithms : Upper Confidence Bound and Dynamic Thompson Sampling.
- We were able to decrease the planning time of a simulated multi-arm manipulator by 30 percent by augmenting out tree switching mechanism.
Technical Report Presentation Video