search

pangtao22 / planning_through_contact

MIT License
42 stars 8 forks source link

A More General Contact Sampler #31

Open hjsuh94 opened 2 years ago

hjsuh94 commented 2 years ago

Currently, the contact sampler are exploiting a lot of knowledge about the ball geometry and finger configuration.

How can we get sampling strategies that are more general?

hjsuh94 commented 2 years ago

One "qualitative characteristic" we want is that we should not limit ourselves to only doing pinch grasps since otherwise, one could come up with hand-written finger gaiting strategies that would do the job, and we want our solution to be more general than that showing contact-rich behavior that RRT can automatically find.

It would also be worthwhile to think through what kind of systems we can show (#14) to be persuasive to people that hand-written policies would simply not do the job.

hjsuh94 commented 2 years ago

If dynamics evaluation is cheap, one way of sampling contacting configurations is to simply look at the B matrices that result from different samples and choose the best one.