search

rail-berkeley / serl

SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning
https://serl-robot.github.io/
MIT License
330 stars 32 forks source link

Gripper operation in bc #46

Closed elishafer closed 5 months ago

elishafer commented 5 months ago

I wanted to ask if you had issues with running bc in the bin example, specifically with operating the gripper; I'm running an experiment similar to yours and the model doesn't close/open the gripper. What happens is that when the model runs, the arm moves to the object that we want to pick up and stays there. If I intervene with the spacemouse by clicking the close gripper button, the robot closes the gripper and continues on the trajectory. The same happens when the gripper has to release an object: the robot stays in place until I press the gripper open button and then it continues with the trajectory. I suspected the issue may be that there weren't enough data points for gripper close, so during data collection I pressed the buttons longer, but it still didn't work. Have you encountered similar problems when using bc?

I'm using a Franka Panda with the panda hand

charlesxu0124 commented 5 months ago

Yes is a common problem with BC, and your intuition is correct: there are not enough transitions to close the gripper when it looks open. One thing you could try is to sample those close gripper steps more frequently during training, although I've never tried that myself.

elishafer commented 5 months ago

In the paper, you mention that you have a 58% success rate with BC for object relocation. Did you encounter any problems with the gripper closing then?