search

Learning-and-Intelligent-Systems / kitchen-worlds

A library of long-horizon Task-and-Motion-Planning (TAMP) problems in kitchen and household scenes, as well as planners to solve them
https://learning-and-intelligent-systems.github.io/kitchen-worlds/
MIT License
69 stars 11 forks source link

Report new results #12

Open CeHao-NUS opened 2 weeks ago

CeHao-NUS commented 2 weeks ago

Hi, Zhutian,

I tested the latest program and here I report some of the results.

B-2 can basically work.

python examples/test_data_generation.py --config_name kitchen_full_pr2.yaml

This config has more frequent world creation failure, as cant ensure_cfree.

python examples/test_data_generation.py --config_name kitchen_full_feg.yaml
python examples/test_data_generation_pigi.py

Two other configs are good. Their world and planning are mostly successful.

C-1 and C-2 can work.

The planning and images are saved correctly.

But C-3 seems strange.

The robot becomes transparent, and it only has a few frames. It might be because the episode is too short. I will double-check this problem.

image

Do you have the program for PIGNet

https://piginet.github.io/

In the RSS paper, you used a transformer to evaluate the planned trajectories. So have you also open-source this part? I would also like to test the transformer to plan the TAMP problem.

I may adjust the state and action space a bit. So for comparison, I would like to align the setting and compare it with your results if possible. Then I can also share my implementation.

Great thanks for your great kitchen-world.

Best, Ce Hao

zt-yang commented 2 weeks ago

Re B-2: can you give me some example seeds that result in ensure_cfree? I ran 5+ times and didn't see any.

Re C-3: that was actually a feature that makes the robot and cabinet doors transparent so that you can see all the movables. I added it as a default-false flag make_robot_and_links_transparent in your_project_folder/configs/config_replay.yaml. Sometimes the sampled goal is just to pick and object so the replay would be short. I also added the --timestep to slow down replay, I suggest setting it to 0.05 - 0.1

Re piginet code: let me test that repo again and make it public. I'll update you on that in a week.