search

erwincoumans / tiny-differentiable-simulator

Tiny Differentiable Simulator is a header-only C++ and CUDA physics library for reinforcement learning and robotics with zero dependencies.
Apache License 2.0
1.23k stars 130 forks source link

tds joint issue #91

Closed erwincoumans closed 3 years ago

erwincoumans commented 3 years ago

simulating floating base, or fixed base (with x,y,z translation + spherical joint, to mimic floating base), has some issue. Attached is a repro case, and pybullet comparison. tds_joint_issues.zip

ErikGartner commented 3 years ago

Hi Erwin!

I've noticed this issue as well, for example the using the humanoid model with a floating base very quickly accelerates until velocity blows up the simulation, but using a completely fixed base seems fine.

So far I've been unable to track down the issue. Do you have any guess to what may be the cause?

erwincoumans commented 3 years ago

Hi Erik, not yet. I first guessed it was due to floating base, but the issue also happens with fixed base + additional degrees of freedom to make an object move. It is also not specific to spherical joints. @eric-heiden is wrapping up a paper for a deadline end of this month, he offered to help tracking it down after that.

erwincoumans commented 3 years ago

See description in this commit: https://github.com/erwincoumans/tiny-differentiable-simulator/commit/288d822d84f6e3d0438b07740bf14ec64b0106e6

So for now, the workaround for this Issue #91 is either:

1) use a fixed base and manually add xyz trans and xyz rotational joints to the fixed base or 2) use a floating base, and make sure the base doesn't have any inertial offset. If there was an offset, move the collision and visual shape of the base in the opposite direction.

erwincoumans commented 3 years ago

Although things have improved, there are still some failing cases, see attached zip file

1) floating base has an issue with mass matrix (not positive definite) at impact (line 48)

2) fixed base with xyz translational, xyz-rotational joint, similar issue

Assertion failed: is_positive_definite_b, file D:\dev\tds\tiny-differentiable-simulator\src\mb_constraint_solver.hpp, line 238

tiny_urdf_parser_opengl_example.zip

eric-heiden commented 3 years ago

102 fixes the fixed-base ant leg contact simulation. The issue was in the URDF to multi-body conversion when the inertia matrix is rotated. Directly rotating the 3x3 inertia matrix is incorrect (can become asymmetric), instead I create a Transform with the defined rotation and then apply that to the RBI of the base and links.

The floating-base ant leg is still behaving incorrectly, I am looking into it.

ErikGartner commented 3 years ago

Thanks! I tested #102 and it improves the simulation. :-) I'll continue experimenting with it and see how far I get.

erwincoumans commented 3 years ago

some fixes for floating base issues here: https://github.com/google-research/tiny-differentiable-simulator/pull/163

erwincoumans commented 3 years ago

Floating base should work reliably now, also with spherical joints. Make sure not to have any inertial center of mass offset at the base link. Closing for now.