sim2sim from v0.2 to v0.3

[ax-anoop]

For a while we've been using v0.2 for locomotion policy on a G1 (very similar code base as the example quadruped code you provide). Asides from a few random bugs v0.2 has been fairly stable, and sim2sim with mujoco has been very consisten.

However, jumping to v0.3 the same policies that use to work in v0.2 and mujoco, no longer work well in v0.3.

Any idea why? What has changed in the physics and or XML / URDF parsing that could cause this ?

[duburcqa]

The main objectives with v0.3 was to improve stability and conformance wrt mujoco. Every behaviour changes that were introduced serve this exact purpose. Unit tests have been implemented to ensure that we are matching Mujoco at machine-precision accuracy for a diverse set of robots.

jumping to v0.3 the same policies that use to work in v0.2 and mujoco, no longer work well in v0.3. Any idea why?

I don't have enough information to draw any conclusions. Many aspects of the physics are now working differently in v0.3, i.e. a new collision detection algorithm has been implemented (which matches Mujoco), default simulation parameters have been updated to match Mujoco, XML / URDF has been refactored to align with Mujoco, and finally, some missing features used in Mujoco have been implemented, notably joint friction. The only thing that is different wrt Mujoco is handling of non-convex geometries. Previously, like Mujoco, they were replaced by their corresponding convex hull. Now a convex decomposition is applied if the convex hull is a bad approximation of the original geometry. This option can be disabled manually if desired.

[ax-anoop]

I need to do more tests on my side, so let me follow up. In general this is my favourity simulator :) sim2real worked on our first runs with G1.

What about in terms of speed ? Another thing I noticed was the training time was slower, though I need to do much more testing and see maybe there's some optimizations to be made.

But as is, my current code base is running at about half the speed, would this make sense at all since physics may have gotten more accurate ?

[duburcqa]

But as is, my current code base is running at about half the speed

Try disabling self-collision using enable_self_collision option.

If it does not help, I would be very interested in getting access to your "benchmark", so that we can iterate on it and find the regression. We expect a 20% slowdown because of the improvement of the physics, not 100%...

[ax-anoop]

This worked as expected, so previously self collisions were disabled by default ?

[duburcqa]

Yes exactly.

[ax-anoop]

Question, why match mujoco physics ?

[ax-anoop]

Is there any options to select "old" physics ?

[duburcqa]

Question, why match mujoco physics ?

Because it is exactly with this objective in min that Genesis was initially designed. Ultimately, it is just a starting point. Once we are happy with our ability to reproduce Mujoco, then it would be time to move forward and start implementing our own physics to overcome the known limitations of Mujoco.

I'm not going to explain all the advantages and drawbacks of this approach. One very concrete advantage is making our life much easier in terms of software validation: no need to write 1000 unit tests checking the validity of the physics, a few integration tests asserting that all quantities are matching mujoco is sufficient.