Jetbot isaac sim sample

I have 2 questions:
1- I tried to change the number of epochs in jetbot_train.py file. but the model still running more than the number that I specified for the n_epochs.
2- when I run the model in a regular mode, the “value_loss” goes down but when I run it in headless mode, the “value_loss” is increasing. Why?


“–total_steps”,
help=“the total number of steps before exiting and saving a final checkpoint”,

It’s the total_steps that controls how long you run the training.

How long was your training? You can plot with the tensorboard step in this and watch the reward Reinforcement Learning — Omniverse Robotics documentation

The reward/loss goes up and down as the training goes, as you can see in the tensorboard plots. Usually it takes more than 50k steps to start getting good results.

1 Like

Thank you for your reply!
I am running the sample in this device: Alienware Aurora R9
Using the default value for “–total_steps” It took 2 days to reach 30 iterations and after that crashed.
I run the code again by changing the “–total_steps” to 50000. It took almost 1 hour to run and after reaching to 50000 steps it terminated by below terminal.


I have no idea if the training has finished normally or not.
If it is normal, now how can I run the sample?
I mean if its trained now, how can I see the output? how can I use this trained model?

It saves out models frequently in the params folder.

checkpoint_callback = CheckpointCallback(save_freq=args.save_freq, save_path="./params/", name_prefix=“rl_model”)

It’s mentioned here.

https://docs.omniverse.nvidia.com/app_isaacsim/app_isaacsim/sample_training_rl.html#evaluate-trained-models

There are lots of useful information on that page, including how to evaluate a trained model, how to continue training, etc. You can let it train overnight once it sta

1 Like

Can you please check the provided link? it doesnt work for me!

Sorry, i edited the link Reinforcement Learning — Omniverse Robotics documentation

I usually let it train overnight, once verifying everything works to your liking.

1 Like

Thank you for reply.
I have 2 questions:
1-How can I save synthetic data?
2-is it possible instead of CNN for training I use another deep learning model? for example Faster RCNN?

  1. Synthetic Data — Omniverse Robotics documentation
  2. DofBot Sample Application — Omniverse Robotics documentation

You can search for rcnn in the code base. The online_generation is using torchvision.models.detection.maskrcnn_resnet50_fpn

The dofbot is using torchvision.models.detection.fasterrcnn_mobilenet_v3_large_fpn.

1 Like

Great!
Now I have a trained DofBot, How can I test it?
how can I have an output like this?

Are the steps in the doc not clear? Did you try running those?
https://isaac.gitlab-master-pages.nvidia.com/omni_isaac_sim/app_isaacsim/app_isaacsim/sample_dofbot.html#training-the-cube-detection-using-isaac-sim-s-synthetic-data-pipeline

1 Like

Thank you.
I have a question about robot navigation. It is not related to training a model.
How can I find a robot navigation guide from scratch in isaac sim? I found some tutorials on isaac stack navigation but I dont know how to apply that on isaac sim.

If you search “navigation” in the docs, you’ll see all these ROS/ROS2 navigation guides as well.

https://docs.omniverse.nvidia.com/app_isaacsim/app_isaacsim/sample_ros_nav.html
https://docs.omniverse.nvidia.com/app_isaacsim/app_isaacsim/sample_ros2_nav.html?highlight=navigation

I got the same error. Have you found a solution?