Skip to content

skrl-v2.0.0

Choose a tag to compare

View all tags
@Toni-SM Toni-SM released this 08 Apr 20:45
· 8 commits to main since this release
2.0.0
c29ced6
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

[2.0.0] - 2026-04-08

Summary of the most relevant features:

  • RL algorithm implementations in NVIDIA Warp
  • Differentiate between environment observations and states (also known as privileged observation)
  • Support for MuJoCo Playground and ManiSkill environments

Added

  • Implement RL algorithms in NVIDIA Warp
  • Add loader and wrapper for MuJoCo Playground environments
  • Add wrapper for ManiSkill environments
  • Add Tabular model instantiator (epsilon-greedy variant)
  • Add clip_mean_actions parameter to Gaussian and Multivariate Gaussian models
  • Add render_interval option to trainers to specify the rendering interval for the environments
  • Add compute_space_limits space utility to get Gymnasium spaces' limits
  • Add ScopedTimer utils to measure code execution time
  • Add SummaryWriter implementation to log data to TensorBoard without relying on third-party libraries
  • Log agent inference and algorithm update, and environment steeping time to TensorBoard

Changed

  • Update minimum supported Python version to 3.10
  • Drop support for PyTorch versions prior to 1.11 (the previous supported version was 1.10)
  • Call observation/state preprocessors once when computing the actions during training

Changed (breaking changes)

  • Refactor the library to differentiate between environment observations and states (also known as privileged observation)
  • Implement agent/multi-agent and trainer configurations using Python Data Classes
    • Unify the different learning rate settings under the learning_rate configuration
    • Rename lambda to gae_lambda
    • Remove the clip_predicted_values redundant configuration by checking for value_clip > 0
    • Remove specific exploration noise settings (initial_scale, final_scale and timesteps)
      in favor of generic scheduling functions
  • Update tabular model definition to operate in any number of parallel environments
  • Refactor multi-agent environment wrappers to support homogeneous and heterogeneous states spaces

Fixed

  • Add entropy loss to the policy loss for on-policy agents/mulit-agents in JAX
  • Fix time limits handling for termination and truncation signals
  • Fix the randomness of the environments by seeding right after initialization (on the first reset)

Removed

  • Remove NumPy backend for JAX implementation
  • Remove checkpoints/models migration support from other RL libraries
  • Remove support for Isaac Gym and Omniverse Isaac Gym environments (deprecated in favor of Isaac Lab)
  • Remove support for Brax and DeepMind environments (in favor of MuJoCo Playground environments)
  • Remove support for Bi-DexHands and robosuite environments
  • Remove Isaac Gym (web viewer, inverse kinematic) and Omniverse Isaac Gym (local environment instance, inverse kinematic) utils
Assets 2
Loading