search

stepjam / PyRep

A toolkit for robot learning research.
MIT License
692 stars 165 forks source link

all zero action fed in get_linear_path, output a non-stationary path #370

Open zwbx opened 7 months ago

zwbx commented 7 months ago

Hi,

I am using RLBenchmark to test a manipulation policy network and have encountered a bug in PyRep:

When a zero action (position: 0, 0, 0; Euler angles: 0, 0, 0) is given with relative_to set to the gripper tip, the get_linear_path function generates a trajectory. All items in this trajectory are expected to be identical because the gripper should not move at all. However, I observed slight variations in the last decimal places of these items. And I do observe slightly movement of the gripper. I am wondering if this issue is due to numerical inaccuracies, or is it an intentional design aspect?

zwbx commented 7 months ago

similar problem refers to https://github.com/stepjam/RLBench/issues/154