Pierre Aumjaud

Yes that's correct, the joints are forced to a new position rather than being applied a position or torque. This is a design choice made by [Replab](https://github.com/bhyang/replab) from which the...

For information, I have just pushed an update to the WidowX environments where the joints are position-controlled (as opposed to being forced to the new position), see [here](https://github.com/PierreExeter/rl_reach/blob/2e95b5963dbd037b9541dbea54a6cfb3514e46df/code/gym_envs/widowx_env/envs/widowx_env.py#L615). You can...

Did you manage to train a model successfully with position-controlled joints? If so, please could you share how you implemented it? It seems that I cannot get the model to...

> When moving with position control the joint values that one gets by calling getjointState can slightly vary from the values that we put into the position control function. This...

> One more thought: I believe that your threshold for sparse rewards is too small. Especially if you train wihtout HER the model might not see enought reward signals. I...

> Another thought: when training her, the reward for sampled goals is calculated using the "compute_reward" function defined in the environment. Because I did not specify the axis along which...

Hi Stefan, I did a quick check and I didn't encounter any problem in training a model with DDPG. I'm not sure what you mean by "fold", can you attach...

I don't have time for a case by case troubleshooting but I can suggest a few things: - check that you can train with DDPG and the most simple env...