Touch Sensors not always registering contact #2483
Unanswered
kuds
asked this question in
Asking for Help
Replies: 2 comments
-
It's a bit hard to see in your video, but I think the touch sensor is not properly aligned, it seems to be above the box rather than enclosing it. If you have a deep penetration then the position of the contact could easily be quite deep in the paddle, which would not be caught by the sensor. |
Beta Was this translation helpful? Give feedback.
0 replies
-
PS, you might want to consider using a fromto sensor, and registering a collision when the fromto vector flips direction |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Intro
Hi!
I am an independent researcher interested in using reinforcement learning to solve various problems and to build experiments in MujoCo.
My setup
Google Colab T4
Python Version: 3.11.11
Torch Version: 2.5.1+cu124
Is Cuda Available: True
Cuda Version: 12.4
Gymnasium Version: 1.0.0
Numpy Version: 1.26.4
Mujoco Version: 3.3.0
Stable-Baselines3 Version: 2.5.0
Matplotlib Version: 3.10.0
My question
I am trying to build a reinforcement learning system that learns to bounce a ball using Gymnasium, Stable Baselines3, and MuJoCo. In reviewing the simulation recording and contact data, there appears to be a misalignment between what the video shows and how many times the touch sensor registers a force. The embedded video shows at least six (6) bounces, but the sensor data only lists three (3) (see attached CSV).
ball_bounce_200000.csv
Relevant Issues:
The above issues stem from the touch sensor area being too small. So, I increased the site size to include the entire paddle, but the touch sensor does not always appear to register contact.
Relevant Documentation:
What is the reason why MuJoCo is not registering all contact instances via the touch sensor? Is it because the timestep is too big, or how the solref is defined for the paddle and ball geom?
Any help would be greatly appreciated!
Minimal model and/or code that explain my question
Link to Google Colab Notebook
[Ball Bounce] Soft Actor-Critic (SAC)
MuJoCo World Model
Confirmations
Beta Was this translation helpful? Give feedback.
All reactions