IsaacDynamics

This project is the first implementation of a co-design framework for the discovery of mechanisms and controllers through isaac lab.

Table of Contents

• Installation
• Usage
• Params
• Information
• Example Commands
• Useful Commands
• Known Issues

Installation

Instructions for installation.

TODO: Add the instructions from Isaac lab here

Example:

sudo apt install cmake build-essential
./isaaclab.sh --install # or "./isaaclab.sh -i"

If you use the jax framework make sure to run the following to get all the dependencies:

pip install --upgrade "jax[cuda12]"
pip install skrl["all"]

Usage

To train an environments:

    ./isaaclab.sh -p scripts/run.py \
        --task Isaac-Pendulum-Simple-Direct-v0 \
        --num_envs 100 \
        --ml_framework jax \
        --mode train \
        --algorithm *Controller \
        --duration 3.0 \
        --max_rollouts 100 \
        --headless

To visualize some test data:

    ./isaaclab.sh -p scripts/run.py \
        --task Isaac-Pendulum-Simple-Direct-v0 \
        --num_envs 1 \
        --ml_framework jax \
        --mode test \
        --algorithm *Controller \
        --duration 3.0 \
        --max_rollouts 10 \
        --record

Params TODO:

To see the flags you can use for the run.py run ./isaaclab.sh -p scripts/run.py --help

Information (TDOD: Add Pedro's Notes)

• To create an environment, a .usd file must be created from a .urdf
• The .usd files from this project are to be found at source/isaaclab_assets/isaaclab_assets/custom. The .py files set up the simulation.
• The task definitions (RL-problem) are defined in source/isaaclab_tasks/isaaclab_tasks. Inside there are the managed tasks and the direct tasks. Managed tasks use higher level abstractions that reduce the difficulty while direct tasks allow granular control.
• For now, all the environments are being run as managed tasks

Example commands

Run to train pendulum environment:

./isaaclab.sh -p scripts/effort_calcs/run.py --task Isaac-Pendulum-v0 --num_envs 64 --ml_framework jax
--controller 'rl' --headless --mode train

Run to record some data on the logs:

./isaaclab.sh -p scripts/effort_calcs/run.py --task Isaac-Pendulum-v0 --num_envs 1 --ml_framework jax --mode test
--controller pid --record --duration 1.0 --max_runtime_iterations 1000 --headless

Run to visualize the firs episode of the latest test:

./isaaclab.sh -p ./scripts/effort_calcs/visualize.py --task Isaac-Pendulum-v0  --ml_framework jax

Useful commands

If you need to setup wandb again:

export WANDB_API_KEY=3a8c037b3fd5bda9fb344d61686b81afc661b0cc
pip install wandb
wandb login

If you want to debug in pycharm, create a conditional in the script like this and also create a run configuration of "Python Debugger" with the code mappings pointing at the root folder:

# Handle debugger connection
if args_cli.debugger:
    import pydevd_pycharm

    pydevd_pycharm.settrace(
        host="localhost",
        port=5678,
        stdoutToServer=True,
        stderrToServer=True,  # Optional: waits for debugger to attach before running
    )

Known Issues

If the isaac sim warning that is stuck persists, run:

gsettings set org.gnome.mutter check-alive-timeout 10000

If you run out of GPU memory while using both isaac AND jax on it, then allow dynamic memory allocation to the start of your script:

# this is required to be able to run both sim and learning in cpu
os.environ["XLA_PYTHON_CLIENT_PREALLOCATE"] = "false"
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"