Implementation of RRT

[jih189]

Is it possible that RRT-like planners will emerge in the future? If not, what might be the reasons?

Here is a work on RRT with CUDA:
[1] J. Bialkowski, S. Karaman and E. Frazzoli, "Massively parallelizing the RRT and the RRT," 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, San Francisco, CA, USA, 2011.

[balakumar-s]

We have a GPU accelerated implementation of PRM* that uses BIT* like exploration when adding nodes. See page 12 of https://curobo.org/reports/curobo_report.pdf.

However, we found that trajectory optimization gave us minimum-time paths faster than our PRM* planner. We still use the PRM* planner to initialize trajectory optimization when linear interpolated seeds fail to solve the problem.

[jih189]

Thank you for replying. Do you trajectory optimization is faster in terms of looking for a feasible solution? Or optimal solution?

[balakumar-s]

Trajectory optimization is better in finding an optimal solution.

Trajectory optimization is also faster in finding a feasible solution. This is because our graph planner implementation is not fully torch compiled (or cuda graph captured). So graph planner ends up calling many small python operations.