search

haosulab / ManiSkill

SAPIEN Manipulation Skill Framework, a GPU parallelized robotics simulator and benchmark
https://maniskill.ai/
Apache License 2.0
864 stars 154 forks source link

Coordinate Transformation for pd_ee_delta_pose in PushCube-v1 Dataset #642

Open zqy-an opened 1 week ago

zqy-an commented 1 week ago

Hello, We are using the dataset available on Hugging Face, specifically examining the first trajectory of PushCube-v1. We have noticed that when using pd_ee_delta_pose, the action does not equate to delta_tcp_pose.

We found some explanations at this link, but we are still unclear about how to specifically transform the coordinates.

Could you please provide more detailed guidance on how to perform this transformation?

fig1

StoneT2000 commented 1 week ago

Could you share a script to reproduce this?

I will also note that we run by default 5 physics steps for every 1 control step with a sim freq of 100. It is possible that the robot has not reached the target (current_ee_pose + delta_pose) in the one control step. Although a difference of 0.2 in the X direction seems concerning. Difference in Z could be due to hitting the table and so you are prevented from going under it (hence larger z difference values).

StoneT2000 commented 3 days ago

Hi @zqy-an is this still an issue?

zqy-an commented 2 days ago

Thanks!

However, I am still very interested in how the action is converted into delta_pose. Could you please tell me where I can view the code responsible for this process?

StoneT2000 commented 17 hours ago

See the replay trajectory tool under mani_skill/trajectory/replay_trajectory.py for how we convert action spaces.