IROS TODOs

[pangtao22]

(Very rough) paper structure

• Intro angles (in addition to / in conjunction with #6):
  • Existing planning through contact approaches that based on exact description of reachable sets cannot deal with the exponential (Yifan has a paper that says it's linear/polynomial?) number of contact modes. But through bundling, the importance of each piece in the dynamics can be characterized by the samples that land in those pieces. In practice, we can get a bundled model that's useful for planning with a small, fixed computational budget.
  • Enforcing 2nd order Newtonian dynamics leads to stiff (high-Lipchitz?) constraints, and as a result, small h (time step? step size) and large T (horizon? number of knot points?)

TODOs

• Dynamic step size based on prediction error.
• Contact sampler, which biases samples during bundling towards more mode changes.
• Terry's RRT.
• Lu's RRT.

[hjsuh94]

Also, here are list of "enemies" to keep aware of:

1. Papers that reason about reachable sets based on linearization.

   • R3T: https://groups.csail.mit.edu/robotics-center/public_papers/Wu20.pdf
   • Motion Cones: https://arxiv.org/abs/1810.00219
   • (LQRTrees? Ani's funnels?)

2. Papers that reason about reachable sets based on mode enumeration.

   • https://arxiv.org/abs/2105.14431

3. Papers that help out the optimization process "too much" by assuming lot of structure. If we do decide to go with the finger teleporting assumption, I would prioritize on how our approach is different from theirs given that fingers teleport

   • TrajectoTree: https://arxiv.org/pdf/2109.14088.pdf
   • Contact Reasoning: https://ieeexplore.ieee.org/iel7/9187508/9196508/09197409.pdf

[hjsuh94]

@pangtao22 I do remember that while doing iRS-MPC paper, Russ wanted to disambiguate whether the trajopt efficiency comes from quasisdynamic or bundling (which led to a painful week implementing everything on second-order systems in Drake again).

Maybe this will happen this time too? I really think both are necessary for long-term steps :sweat:

[pangtao22]

I remember that damping has to be tuned really carefully in drake for the auto-diffed derivatives to make sense... and Russ later kind of forgot about asking for it..., and focused more on the connection to RL?