CODEWITHSUNDEEP
simulationcontactsroboticsdrake
jmlee0106
jmlee0106

Reputation: 11

Hydroelastic contact and robust contact generation

I have a keen interest in physics simulations and robust contact generation for robotics problems. I'm aware of Drake's hydroelastic contact module [1,2], which provides an elegant approach for modeling contact accurately with arbitrary geometries.

My specific inquiry pertains to a more precise discussion on accuracy and robustness in contact modeling. Based on my experience and observations in related papers, it's evident that the point contact model isn't robust enough to simulate a wide range of scenarios. For example, simulating stable box stacking using a single point from GJK-EPA or MPR algorithms poses challenges.

However, to my best knowledge, there are still some options to enhance stability & reflect area-dependent effects e.g.,

  1. generating contact manifold using SAT algorithm+clipping 2) employing persistent manifold by caching contact points from previous time steps 3) introducing perturbation geometry pose to get multiple contact points 4) utilizing SDFs, etc.

In my implementation tests, all of these methods seem to work well in box stacking scenarios. However, I'm interested in a more in-depth discussion beyond the assessment of "seem to work well."

I have a few specific questions:

  1. How can we evaluate the advantages and disadvantages of methods for contact generation, in the context of robotics, especially in manipulation, planning, and control through contact?

  2. Are there established benchmark scenarios that can be used to compare the accuracy and robustness of different contact generation methods?

  3. How significant is this issue? I'm interested in this topic, but I think works that describe the effectiveness of robust/accurate contact generation on robotic tasks in the real world are almost non-existent.

Regarding to 1, I think hydroelastic contact model is less heuristic which gives some advantages in (automatic) differentiation and model-based planning & control. But any insights or guidance will be very helpful.

Sorry for a long question. Thank you in advance.

[1] R. Elandt et al, A pressure field model for fast, robust approximation of net contact force and moment between nominally rigid objects, IROS 2019.

[2] J. Masterjohn et al, Velocity Level Approximation of Pressure Field Contact Patches, RAL 2022.

Upvotes: 1

Views: 218

Answers (1)

Sherm
Sherm

Reputation: 678

At the risk of being trite, the best model depends on what you are trying to do with it. There are tradeoffs in performance, accuracy, differentiability, difficulty of specifying the model, and complexity of code. Point contact is at one end and nonlinear finite element contact at the other. I have seen some attempts to quantify contact model differences and found them entirely unconvincing. In Drake we are focused on "what gets the job done" for control through robotic manipulation. You can read a little about how we landed on hydroelastic contact in this blog post.

Upvotes: 1

Related Questions