[Question] Unable to achieve 100% Success Rate on Lift Task Using BCC_GMM

[AryanSarswat]

I have collected approximately 140 samples of teleoperated data for the Lift Task. I verified the collected data using the replay_demo.py provided in repo.

Using this dataset, I trained a BCC_GMM model utilizing the provided train.py script provided in the repo. Here is also attached the config and logs from train.py.

config.txt
log_train.txt

Despite these efforts, my trained model achieves only around a 50% success rate (from 10 rollouts using play.py). I am reaching out to inquire if anyone has successfully trained a model to consistently achieve a 100% success rate on the Lift Task within IsaacLab.

Questions:

Has anyone been able to achieve or approach a 100% success rate for the Lift Task?

If so, could you share insights or adjustments made to your training approach or configuration that contributed significantly to improved performance?

Thank you for your assistance!

[RandomOakForest]

Thank you for posting this. I will move this post to our Discussions sections for the team to follow up.

[RandomOakForest]

Following up, to increase your success rate you could just add more examples to your training. 140 is a low number. Consider using Isaac Lab Mimic.

[AryanSarswat]

As a general rule of thumb how many samples do you think is needed?