robfiras / loco-mujoco

Imitation learning benchmark focusing on complex locomotion tasks using MuJoCo.
MIT License
539 stars 46 forks source link

need support: unable to reproduce satisfactory training results #23

Open leyongloh opened 5 months ago

leyongloh commented 5 months ago

Hi Firas, I really appreciate your great work on this benchmark, it saves us a lot of time in our research project. Currently, I am trying to reproduce the walking and running examples demonstrated by the gif in Readme. However, I was not able to get a satisfactory result using the examples/imitation_learning training pipeline for HumanoidTorque4Ages environment (my models failed even after I tried different configurations). I only able to get a good result with the HumanoidTorque.walk environment. I trained the model with arms and feet moving due to the requirements of my project, I also tried using only box_feet for the HumanoidTorque4Ages but was not successful. Have been working on the training for almost a month, if you could provide the working pretrained model for HumanoidTorque4Ages that will be super helpful, would appreciate any support/help, thanks a lot!

robfiras commented 4 months ago

Hi @leyongloh, sorry for my late reply and thanks for using LocoMuJoCo! HumanoidTorque4Ages is a challenging environment, as it is randomly changing the humanoid. I have a different project in which I actually use this environment to train an agent suitable across different humanoid sizes (check it out here). But this was without arms. The gif in the readme just shows the motion capture, not a trained policy. Due to time constraints, I did not have time to solve this environment with arms. So, while I can not directly provide you with an agent, I am more than willing to help you solving this env! Drop me an email (click on "contact me" here), maybe we can have a meeting.