Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

errors with amp #37

Open
BRoln7 opened this issue Dec 4, 2024 · 11 comments
Open

errors with amp #37

BRoln7 opened this issue Dec 4, 2024 · 11 comments

Comments

@BRoln7
Copy link

BRoln7 commented Dec 4, 2024

I used the SMPL humanoid to train an amp task, and I found in inference stage, the humanoid just stand, what is the problem? The motion file contains several actions including walking and running

@tesslerc
Copy link
Collaborator

tesslerc commented Dec 4, 2024

What command did you run for training and what for test?

phys_anim/train_agent.py +exp=amp +robot=smpl +backbone=isaacgym motion_file=phys_anim/data/motions/smpl_humanoid_walk.npy should result in a simple walking motion with less than an hour.

@BRoln7
Copy link
Author

BRoln7 commented Dec 4, 2024 via email

@BRoln7
Copy link
Author

BRoln7 commented Dec 10, 2024

How should I prepare a dataset when training for a path-following task? Specifically, how should I select motions from the AMASS dataset?

@tesslerc
Copy link
Collaborator

Convert the dataset to Isaac format, see the data section in the README.
Then, create a .yaml file containing the motion styles you want as a prior for your policy. See examples in the data/yaml_files/ folder for how these files should be structured.
Then, as explained in the README, you can pre-package the data for faster loading.

The result is a .pt file that you provide to the flag motion_file=<path to .pt file>.

@BRoln7
Copy link
Author

BRoln7 commented Dec 10, 2024

Thank you for your reply! But actually, I would like to know how to determine which motions from the AMASS dataset are what I need, since the dataset is quite large.

@tesslerc
Copy link
Collaborator

In this file you have all the motions and their textual labels.
For path following you can focus on walk/run/crouch-walk/crawl/stand as a good representation of the types of motions.

For reproducing the Trace and Pace paper, I suggest looking into the motions they used here.

@BRoln7
Copy link
Author

BRoln7 commented Dec 10, 2024

Actually, I want to reproduce the path-follower in Isaac Lab. When I try to use the purple ball shown in the video to guide the humanoid forward, the motion is very stiff. What could be the issue? The dataset I prepared contains some forward-walking motions from KIT

3_1733737051.mp4

@tesslerc
Copy link
Collaborator

The video doesn't load.
How many motions did you include? Did you turn off the height conditioning in the path follower config?

@BRoln7
Copy link
Author

BRoln7 commented Dec 11, 2024

3_1733737051.mp4

Yes, I turned it off. I included five walking motions from KIT in amass dataset.

@tesslerc
Copy link
Collaborator

Can you try adding some more motions and see if that's still the case?

@BRoln7
Copy link
Author

BRoln7 commented Dec 20, 2024

When I load motions, it ocurrs such a problem:

/home/bingyi/miniconda3/envs/isaaclab/lib/python3.10/site-packages/smplx/body_models.py:137: DeprecationWarning: Please use `csc_matrix` from the `scipy.sparse` namespace, the `scipy.sparse.csc` namespace is deprecated.
  data_struct = Struct(**pickle.load(smpl_file,
/home/bingyi/miniconda3/envs/isaaclab/lib/python3.10/site-packages/chumpy/ch.py:20: DeprecationWarning: Please use `LinearOperator` from the `scipy.sparse.linalg` namespace, the `scipy.sparse.linalg.interface` namespace is deprecated.
  from scipy.sparse.linalg.interface import LinearOperator
/home/bingyi/miniconda3/envs/isaaclab/lib/python3.10/site-packages/chumpy/ch.py:1203: DeprecationWarning: inspect.getargspec() is deprecated since Python 3.0, use inspect.signature() or inspect.getfullargspec()
  want_out = 'out' in inspect.getargspec(func).args
Sampling motion: tensor([0, 0, 0, 0])
/home/bingyi/socnav_sim/poselib/poselib/skeleton/skeleton3d.py:1240: DeprecationWarning: Please use `gaussian_filter1d` from the `scipy.ndimage` namespace, the `scipy.ndimage.filters` namespace is deprecated.
  filters.gaussian_filter1d(velocity, 2, axis=-3, mode="nearest")
  0%|                                                                                    | 0/3 [00:00<?, ?it/s]

Then the progress of the load action is stuck.I wonder what the problem is?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants