Difficulties when Training RL Policies

[PeideHuang]

I tried to run the RL training scripts for multiple tasks such as Stabilize, Reach and Grasp, and Insert by
python3 main/rl/train.py task=<task_name> sim_device=cuda:<gpu_id> rl_device=cuda:<gpu_id> graphics_device_id=<gpu_id>
However, none of the RL agents successfully learn to complete the tasks even after long time (an example for ReachAndGraspSingle shown below). I used the num_envs in the default task config file. Are there any hyperparameters I need to tune?

[yunfanjiang]

For most tasks we also added task-specific curricula (such as warm up over object geometries and action space), reward functions provided in the code correspond to those in the last curriculum stages. But for the task Stabilize there should be some learning signal without the curriculum. Could you provide some learning curves for that task so I can take a look?

[user432]

Hello @yunfanjiang, thanks for open-sourcing your code. I would also like to ask the same question that if we should be tuning any parameters before running the code. I ran the InsertSingle task as it is and here are the plots: