Planning through Contact with Smoothing

Pang, Terry, Lu

Summary of Paper Draft Feedback

  • Randomized smoothing --> smoothing
    • A "force at a distance" formulation should also work, and should be faster.
    • It would be nice to establish the correspondence between smoothing scheme and smoothing distribution.
  • Remove comparison between extension with traj-opt and extension with projection.
    • Consequences of using traj-opt, such as fewer nodes and less dancing, are expected and should not be the focus.
  • Mention how long it takes to find trajectories.
    • Not mentioning them makes people think that planning is slow.
  • Comparison with more established baselines, e.g. "standard RRT". 
    • One of R3T's baselines: "Randomized Kinodynamic Planning" https://journals.sagepub.com/doi/pdf/10.1177/02783640122067453
  • Address the non-standard definition of reachability in our paper: why is \(\mathcal{R}\) a function of \(u\)?

Reduce hydro-planing with multi-step rollouts.

  • We divide the \(h\) used for RRT into finer steps (e.g. \(h_s = h / 10\)), and roll out using \(h_s\) for 10 steps. 
  • Need to scale down the mass matrix accordingly. Here the mass matrix should be interpreted as the damping experienced by the object from the environment.

Pros: less hyrdoplaning.

Cons:

  • wrong model used for planning, which hurts the performance of RRT. 
  • Motion looks more "choppy", as motions that rely on hydro-planing to move the object are no longer effective.

 

 

Taking a step back, we think the gap between hardware and the assumptions made in this work that made the plan aphysical is better addressed as future work, because it is hard to evaluate the pros and cons of different strategies (better plan or better controller?).