Closed tfederico closed 4 years ago
They should be very similar, I tried using the same training code from Jason Peng, some differences include: 1) PyBullet uses a larger time step 2) PyBullet uses a URDF definition of the character, but it should be identical 3) PyBullet currently doesn't enable the root orientation reward in training, so direction may drift. 4) PyBullet is easier to install (pip3 install pybullet)
There could be small other unintended differences. We could use the original policies from DeepMimic directly in PyBullet, but a while back some of them started to perform poorly (didn't have time to track down why) so I retrained some of the policies.
I created a plain Gym environment wrapper to train the PyBullet deep mimic, but didn't train it fully yet, so for now we rely on Jason't training code, which is included. So you can use: See here for examples how to train deep mimic in PyBullet: https://github.com/bulletphysics/bullet3/issues/2304
Closing for now, just reply if you have other questions. Also, note that Jason Peng did an internship with us and we released another imitation environment based on PyBullet for our quadruped robots. See https://github.com/google-research/motion_imitation That one uses Stable Baselines PPO to train.
Looks amazing, thanks! I am currently interested in human-like characters, but I will definitively take a look at the paper :)
out of curiosity, is there any reason why you chose VideoPose3D instead of using HMR or HMMR like in the Skills from Video paper?
@tfederico There is no video involved in PyBullet deep_mimic.
sorry, I must have confused with some other repo I was looking at
On Wed, 24 Jun 2020 at 20:39, erwincoumans notifications@github.com wrote:
@tfederico https://github.com/tfederico There is no video involved in PyBullet deep_mimic.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/bulletphysics/bullet3/issues/2871#issuecomment-649026647, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADU47BR2T77O433I22RLQLDRYJI67ANCNFSM4OCZSLQA .
Hi,
which are the current differences between the DeepMimic implementation in pybullet and the original one?