Discrepancy between Fwd and Inv Dynamics in CUDA

[michaellin]

Inverse and forward dynamics functions should agree with each other. We were testing this assumption in the following:

tau_pred = compute_inverse_dynamics(q, qd, qdd)
qdd_pred = compute_forward_dynamics(q, qd, tau_pred)

We expected qdd and qdd_pred to be very close to each other.

When we do this experiment with device=cpu, the L2 error is in the 10^-5 order. But when we use device=cuda, we get L2 errors around 0.12 which is unexpected.

I have a simple script that minimally reproduces this problem here

@fmeier @gsutanto @exhaustin

[exhaustin]

Hello @michaellin!
I tried running your scripts but am unable to reproduce the discrepancy between running forward and inverse dynamics on CPU and CUDA.
After fixing the random seeds, I got an average L2 loss of 1.679e-5 on CPU and 1.683e-5 with CUDA. The differences negligible, and are probably due to different rng schemes for CPU vs CUDA tensors.
For reference I am using Pytorch 1.13.0 with CUDA 11.6.2
Can you try using the same dependencies and see if the issue persists?