In addition to video rendering

[chan-yuu]

Excellent work

python ppo.py --env_id="PickCube-v1" --evaluate --checkpoint=./runs/PickCube-v1__ppo__1__1765370681/final_ckpt.pt --num_eval_envs=1 --num-eval-steps=1000

I got a lot of video files, I wonder if it is possible to use render for real-time visualization in addition to video rendering when evaluating. For example, env.render() method.

[StoneT2000]

We don't recommend doing real time GUI rendering during training really. If you really want to, it is better to just write another script that loads a PPO checkpoint and then renders in a window for you.

If you really want to record videos and open a render window, during the eval loop of PPO add eval_envs.render_human()

[chan-yuu]

I see. Thank you.

I had problems running interactive controls:

python interactive_panda.py -e "PickCube-v1" 

(Segmentation fault (core dumped))

PandaArmMotionPlanningSolver failed to initialize properly, What am I supposed to do about it?

[chan-yuu]

I solved this problem by reinstalling mplib

pip uninstall mplib
pip install mplib==0.1.1

My research is mainly about mobile manipulation, I found that the current interactive control seems unable to collect such data, I would like to ask what methods can collect mobile manipulation data, there is a successful rl method? I saw demos write no, does this mean that no training has been successful?

[StoneT2000]

You can solve that task with PPO, there's an example in the baselines.sh file of the ppo folder.

After solving it you can try to generate demos from there