danieltakeshi / gctn_physical

Physical GCTN

Physical GCTN Experiments.

Physical experiments for our GCTN paper: https://berkeleyautomation.github.io/bags/

For the simulation code, see https://github.com/DanielTakeshi/deformable-ravens

Disclaimer: a lot of this code depends on our physical robot and some modifications to the GCTN training code. We recommend using the code as a reference, but not actually running it.

Usage

• Make sure we have a model of GCTN available.
• Make sure takeshi is running with a script.
• Make sure we're in our Python 3.6 conda env (I name it franka).

Run the following commands:

python main.py [args]

See that script for more, and daniel_utils.py for a bunch of utility methods.

Run this to get demonstrations:

python get_demo_data.py

See the scripts/ directory (and the associated README) for testing a bunch of things.

Directories:

• data/ is meant for output from different methods.
• demos/ is meant for saving demonstration data.
• images/ is meant for debugging data_collector.py's image querying.
• goals_real_cable_line_notarget/ has the goal images for test time rollouts.
• real_input/ and real_output/ are for handling GCTN synchronization.

NOTE: we technically have a termination time step but we are planning to just use CTRL+C for exiting. This should still save the right amount of data.

The main script will save at:

data/gctn/trial_002__2023-02-16_17-57-33:
total 104M
-rw-rw-r-- 1 testseita testseita  246 Feb 16 17:57 args.json
-rw-rw-r-- 1 testseita testseita 817K Feb 16 17:58 c_img_pre_action_time_00.png
-rw-rw-r-- 1 testseita testseita 816K Feb 16 18:00 c_img_pre_action_time_01.png
-rw-rw-r-- 1 testseita testseita 821K Feb 16 18:01 c_img_pre_action_time_02.png
-rw-rw-r-- 1 testseita testseita 816K Feb 16 18:03 c_img_pre_action_time_03.png
-rw-rw-r-- 1 testseita testseita 814K Feb 16 18:05 c_img_pre_action_time_04.png
-rw-rw-r-- 1 testseita testseita 817K Feb 16 18:06 c_img_pre_action_time_05.png
-rw-rw-r-- 1 testseita testseita  99M Feb 16 18:08 trial_info.pkl

Then later we might inspect things and populate those directories with more images.

My Setup

To get images:

• I use the robot e-stop to move its gripper to a reasonable spot, then python basic_franka_usage.py to fine-tune it with code. I want a good top-down image.
• Then test with python data_collect.py to check bounding boxes and cropping convention and segmentation masks.
• Then check that the robot can actually reach the extremes of the workspace.

Only once those are done, then I can actually collect 'real' demonstration data. I do this by hand since it's a LOT faster than having the robot move there.

Some random tips from Oliver's students:

From Kevin:

• The issue is that you mounted a kinect on the end-effector which affects the weight of the robot. With our impedance controller, we need to set the accurate weight of the robot. When setting use_impedance=False, you would be using the Franka's internal ik and stiffness which are better at counteracting small differences in weight

• Yeah so use_impedance is set to true by default because in certain joint configurations the inbuilt ik doesn't work, but usually the inbuilt ik results in smoother motions.