Reinforcement Learning Workflows

Reinforcement Learning Workflows#

We provide wrappers to different reinforcement libraries. These wrappers convert the data from the environments into the respective libraries function argument and return types.

Preset Selectors#

All training and play commands accept physics=NAME, renderer=NAME, and presets=NAME[,NAME,...] tokens appended directly to the command (no leading dashes). See Hydra Configuration System for all available names and how the selectors work.

# Switch physics backend
./isaaclab.sh train --rl_library <library> \
    --task <task-name> --headless physics=newton_mjwarp

# Switch renderer (camera environments)
./isaaclab.sh train --rl_library rsl_rl \
    --task Isaac-Cartpole-Camera-Direct --headless \
    --enable_cameras renderer=newton_renderer

# Combine selectors freely
./isaaclab.sh train --rl_library rsl_rl \
    --task Isaac-Cartpole-Camera-Direct --headless \
    --enable_cameras physics=newton_mjwarp renderer=newton_renderer presets=rgb

Note

Not all environments support the Newton backend yet. Using physics=newton_mjwarp with an environment that has not been configured for Newton will raise an error at launch. See Newton Backend for more details, and the Migrating to Isaac Lab 3.0 guide for how to add Newton support to your own environments.

Newton does not require Isaac Sim (kit-less mode). See Kit-less Installation for setup.

Observation-mode Presets#

Some environments support multiple observation modes selectable via presets=. Unlike physics or renderer presets, observation-mode presets affect the checkpoint structure: you must pass the same preset to both the training and play commands. Using a different preset (or none) at play time will cause a model-architecture mismatch when loading the checkpoint.

For example, Isaac-Repose-Cube-Shadow-Vision-Direct-v0 defaults to RGB + depth + segmentation inputs but can be switched to RGB-only with presets=rgb:

# Train with RGB-only observations
./isaaclab.sh train --rl_library rsl_rl \
    --task Isaac-Repose-Cube-Shadow-Vision-Direct-v0 --headless \
    --enable_cameras presets=rgb

# Play — must use the same preset to load the matching checkpoint
./isaaclab.sh play --rl_library rsl_rl \
    --task Isaac-Repose-Cube-Shadow-Vision-Direct-Play-v0 \
    --enable_cameras presets=rgb

Other available presets for this environment: albedo, simple_shading_constant_diffuse, simple_shading_diffuse_mdl, simple_shading_full_mdl. The depth preset is intended for benchmarking only (see the environment’s config for details).

RL-Games#

Attention

When using RL-Games with the Ray workflow for distributed training or hyperparameter tuning, please be aware that due to security risks associated with Ray, this workflow is not intended for use outside of a strictly controlled network environment.

  • Training an agent with RL-Games on Isaac-Ant:

    # install python module (for rl-games)
./isaaclab.sh -i rl_games
# run command for training
./isaaclab.sh train --rl_library rl_games --task Isaac-Ant --headless
# run command for training with Newton backend
./isaaclab.sh train --rl_library rl_games --task Isaac-Ant --headless physics=newton_mjwarp
# run command for playing with 32 environments
./isaaclab.sh play --rl_library rl_games --task Isaac-Ant --num_envs 32 --checkpoint /PATH/TO/model.pth
# run command for recording video of a trained agent (requires installing `ffmpeg`)
./isaaclab.sh play --rl_library rl_games --task Isaac-Ant --headless --video --video_length 200

    :: install python module (for rl-games)
isaaclab.bat -i rl_games
:: run command for training
isaaclab.bat train --rl_library rl_games --task Isaac-Ant --headless
:: run command for training with Newton backend
isaaclab.bat train --rl_library rl_games --task Isaac-Ant --headless physics=newton_mjwarp
:: run command for playing with 32 environments
isaaclab.bat play --rl_library rl_games --task Isaac-Ant --num_envs 32 --checkpoint /PATH/TO/model.pth
:: run command for recording video of a trained agent (requires installing `ffmpeg`)
isaaclab.bat play --rl_library rl_games --task Isaac-Ant --headless --video --video_length 200

RSL-RL#

  • Training an agent with RSL-RL on Isaac-Reach-Franka:

    # install python module (for rsl-rl)
./isaaclab.sh -i rsl_rl
# run command for training
./isaaclab.sh train --rl_library rsl_rl --task Isaac-Reach-Franka --headless
# run command for training with Newton backend
./isaaclab.sh train --rl_library rsl_rl --task Isaac-Reach-Franka --headless physics=newton_mjwarp
# run command for playing with 32 environments
./isaaclab.sh play --rl_library rsl_rl --task Isaac-Reach-Franka --num_envs 32 --load_run run_folder_name --checkpoint /PATH/TO/model.pt
# run command for recording video of a trained agent (requires installing `ffmpeg`)
./isaaclab.sh play --rl_library rsl_rl --task Isaac-Reach-Franka --headless --video --video_length 200

    :: install python module (for rsl-rl)
isaaclab.bat -i rsl_rl
:: run command for training
isaaclab.bat train --rl_library rsl_rl --task Isaac-Reach-Franka --headless
:: run command for training with Newton backend
isaaclab.bat train --rl_library rsl_rl --task Isaac-Reach-Franka --headless physics=newton_mjwarp
:: run command for playing with 32 environments
isaaclab.bat play --rl_library rsl_rl --task Isaac-Reach-Franka --num_envs 32 --load_run run_folder_name --checkpoint /PATH/TO/model.pt
:: run command for recording video of a trained agent (requires installing `ffmpeg`)
isaaclab.bat play --rl_library rsl_rl --task Isaac-Reach-Franka --headless --video --video_length 200

  • Training and distilling an agent with RSL-RL on Isaac-Velocity-Flat-Anymal-D-v0:

    # install python module (for rsl-rl)
./isaaclab.sh -i rsl_rl
# run command for rl training of the teacher agent
./isaaclab.sh train --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --headless
# run command for rl training of the teacher agent with Newton backend
./isaaclab.sh train --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --headless physics=newton_mjwarp
# run command for distilling the teacher agent into a student agent
./isaaclab.sh train --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --headless --agent rsl_rl_distillation_cfg_entry_point --load_run teacher_run_folder_name
# run command for playing the student with 64 environments
./isaaclab.sh play --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --num_envs 64 --agent rsl_rl_distillation_cfg_entry_point

    :: install python module (for rsl-rl)
isaaclab.bat -i rsl_rl
:: run command for rl training of the teacher agent
isaaclab.bat train --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --headless
:: run command for rl training of the teacher agent with Newton backend
isaaclab.bat train --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --headless physics=newton_mjwarp
:: run command for distilling the teacher agent into a student agent
isaaclab.bat train --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --headless --agent rsl_rl_distillation_cfg_entry_point --load_run teacher_run_folder_name
:: run command for playing the student with 64 environments
isaaclab.bat play --rl_library rsl_rl --task Isaac-Velocity-Flat-Anymal-D-v0 --num_envs 64 --agent rsl_rl_distillation_cfg_entry_point

SKRL#

  • Training an agent with SKRL on Isaac-Reach-Franka:

    # install python module (for skrl)
./isaaclab.sh -i skrl
# run command for training
./isaaclab.sh train --rl_library skrl --task Isaac-Reach-Franka --headless
# run command for training with Newton backend
./isaaclab.sh train --rl_library skrl --task Isaac-Reach-Franka --headless physics=newton_mjwarp
# run command for playing with 32 environments
./isaaclab.sh play --rl_library skrl --task Isaac-Reach-Franka --num_envs 32 --checkpoint /PATH/TO/model.pt
# run command for recording video of a trained agent (requires installing `ffmpeg`)
./isaaclab.sh play --rl_library skrl --task Isaac-Reach-Franka --headless --video --video_length 200

    :: install python module (for skrl)
isaaclab.bat -i skrl
:: run command for training
isaaclab.bat train --rl_library skrl --task Isaac-Reach-Franka --headless
:: run command for training with Newton backend
isaaclab.bat train --rl_library skrl --task Isaac-Reach-Franka --headless physics=newton_mjwarp
:: run command for playing with 32 environments
isaaclab.bat play --rl_library skrl --task Isaac-Reach-Franka --num_envs 32 --checkpoint /PATH/TO/model.pt
:: run command for recording video of a trained agent (requires installing `ffmpeg`)
isaaclab.bat play --rl_library skrl --task Isaac-Reach-Franka --headless --video --video_length 200

    Warning

    It is recommended to install JAX manually before proceeding to install skrl and its dependencies, as JAX installs its CPU version by default. Visit the skrl installation page for more details. Note that JAX GPU support is only available on Linux x86_64 and Linux aarch64. Use the CUDA 12 wheel on Linux x86_64 and the CUDA 13 wheel on Linux aarch64 systems such as DGX Spark.

    Hint

    When using JAX its default behavior is to pre-allocate 75% of the GPU memory for its own computations. If you run into memory issues, you can set the XLA_PYTHON_CLIENT_PREALLOCATE=false environment variable to disable this behavior, or reduce the amount of pre-allocated memory by setting export XLA_PYTHON_CLIENT_MEM_FRACTION=0.5 which will allocate 50% of the GPU memory for JAX. Any value between 0 and 1 can be set, where 0 will allocate no memory for JAX and 1 will allocate 100% of the GPU memory for JAX.

    # install python module (for skrl)
./isaaclab.sh -i skrl
# install JAX for CUDA 12
./isaaclab.sh -p -m pip install -U "jax[cuda12]"
# install skrl dependencies for JAX
./isaaclab.sh -p -m pip install "skrl[jax]"

    # install python module (for skrl)
./isaaclab.sh -i skrl
# install JAX for CUDA 13
./isaaclab.sh -p -m pip install -U "jax[cuda13]"
# install skrl dependencies for JAX
./isaaclab.sh -p -m pip install "skrl[jax]"

    # run command for training
./isaaclab.sh train --rl_library skrl --task Isaac-Reach-Franka --headless --ml_framework jax
# run command for training with Newton backend
./isaaclab.sh train --rl_library skrl --task Isaac-Reach-Franka --headless --ml_framework jax presets=newton_mjwarp
# run command for playing with 32 environments
./isaaclab.sh play --rl_library skrl --task Isaac-Reach-Franka --num_envs 32  --ml_framework jax --checkpoint /PATH/TO/model.pt
# run command for recording video of a trained agent (requires installing `ffmpeg`)
./isaaclab.sh play --rl_library skrl --task Isaac-Reach-Franka --headless --ml_framework jax --video --video_length 200

    • Training the multi-agent environment Isaac-Shadow-Hand-Over-Direct-v0 with skrl:

    # install python module (for skrl)
./isaaclab.sh -i skrl
# run command for training with the MAPPO algorithm (IPPO is also supported)
./isaaclab.sh train --rl_library skrl --task Isaac-Shadow-Hand-Over-Direct-v0 --headless --algorithm MAPPO
# run command for playing with 32 environments with the MAPPO algorithm (IPPO is also supported)
./isaaclab.sh play --rl_library skrl --task Isaac-Shadow-Hand-Over-Direct-v0 --num_envs 32 --algorithm MAPPO --checkpoint /PATH/TO/model.pt

    :: install python module (for skrl)
isaaclab.bat -i skrl
:: run command for training with the MAPPO algorithm (IPPO is also supported)
isaaclab.bat train --rl_library skrl --task Isaac-Shadow-Hand-Over-Direct-v0 --headless --algorithm MAPPO
:: run command for playing with 32 environments with the MAPPO algorithm (IPPO is also supported)
isaaclab.bat play --rl_library skrl --task Isaac-Shadow-Hand-Over-Direct-v0 --num_envs 32 --algorithm MAPPO --checkpoint /PATH/TO/model.pt

Stable-Baselines3#

  • Training an agent with Stable-Baselines3 on Isaac-Velocity-Flat-Unitree-A1-v0:

    # install python module (for stable-baselines3)
./isaaclab.sh -i sb3
# run command for training
./isaaclab.sh train --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --headless
# run command for training with Newton backend
./isaaclab.sh train --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --headless physics=newton_mjwarp
# run command for playing with 32 environments
./isaaclab.sh play --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --num_envs 32 --checkpoint /PATH/TO/model.zip
# run command for recording video of a trained agent (requires installing `ffmpeg`)
./isaaclab.sh play --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --headless --video --video_length 200

    :: install python module (for stable-baselines3)
isaaclab.bat -i sb3
:: run command for training
isaaclab.bat train --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --headless
:: run command for training with Newton backend
isaaclab.bat train --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --headless physics=newton_mjwarp
:: run command for playing with 32 environments
isaaclab.bat play --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --num_envs 32 --checkpoint /PATH/TO/model.zip
:: run command for recording video of a trained agent (requires installing `ffmpeg`)
isaaclab.bat play --rl_library sb3 --task Isaac-Velocity-Flat-Unitree-A1-v0 --headless --video --video_length 200

RLinf#

RLinf is a distributed RL infrastructure for fine-tuning Vision-Language-Action (VLA) models such as GR00T. It uses Ray for distributed computing and FSDP for model parallelism, enabling RL training of large VLA models that don’t fit on a single GPU.

For installation instructions, see RL Post-Training for VLA Models.

  • Training a VLA agent with RLinf:

    # Train with a specific config
./isaaclab.sh train --rl_library rlinf \
    --config_name isaaclab_ppo_gr00t_assemble_trocar \
    --model_path /path/to/checkpoint

  • Evaluating a trained VLA agent:

    # Evaluate with video recording
./isaaclab.sh play --rl_library rlinf \
    --config_name isaaclab_ppo_gr00t_assemble_trocar \
    --model_path /path/to/checkpoint --video

All the commands above log the training progress to Tensorboard in the logs directory in the root of the repository. The logs directory follows the pattern logs/<library>/<task>/<date-time>, where <library> is the name of the learning framework, <task> is the task name, and <date-time> is the timestamp at which the training command was executed.

To view the logs, run:

# execute from the root directory of the repository
./isaaclab.sh -p -m tensorboard.main --logdir=logs

:: execute from the root directory of the repository
isaaclab.bat -p -m tensorboard.main --logdir=logs