Updates robomimic config for Franka-Lift environment

[yangcyself]

Description

1. Enlarge the hard-coded success range from 0.002 to 0.02 so that it's possible for a human operator to succeed.
2. Update from object_positions to object_relative_tool_positions in lift_bc.json and lift_bcq.json, to match with the configuration in lift_cfg.py

Fixes #68

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist

• I have run the pre-commit checks with ./orbit.sh --format
• I have made corresponding changes to the documentation: Not applicable
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works: Not applicable
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file: Not applicable

[Mayankm96]

Are you able to train a policy through imitation learning with these changes?

I think in the earlier version, object_positions were the object pose in the environment frame, not relative to tool frame.

[yangcyself]

I actually did not succeed in training a policy to grab the tube with behavior cloning. But I thought it was because of the bad demonstrations I provided.

And yes, in the earlier version object_positions was used in lift_cfg.py but you changed it to object_relative_tool_positions in #33 (link), so I thought we need to update and use object_relative_tool_positions in imitation learning.

So maybe it's better to stick to the previous setting and use object_positions? So what I should do it's to config the observation to collect object_positions in the data collection step, by changing the file collect_demonstrations.py?

[Mayankm96]

Training with human data can indeed be challenging. I recommend checking this paper on tips for imitation learning

https://arxiv.org/abs/2108.03298

Maybe to check if the observations make sense, you can collect data with the state machine and see how it performs. It would be easier to collect data that way.

I will test your PR later this week or early next week.

[Mayankm96]

Looks good. Would need to attempt re-training again but will do that when the time comes.