Release Notes#

Arena 0.3 / Isaac Lab 3.0 Upgrade#

This checkout pins Arena 0.3.1 at 481f5ae5f19df7bc5d24a17240fc0d96ac81eaf7 and its nested Isaac Lab commit 28a386f063e41c04c07f50e63eefb83fd8408fbe (v3.0.0-EA plus 3 commits), with Isaac Sim 6.1.0 Docker images. See the support matrix for the complete version set.

  • Workflow commands, environment profiles, and source dataset metadata use Lab’s IsaacContrib-* task names directly.

  • Generation and annotation default to state observations. Use AutoData’s --enable_cameras option to include a task’s configured image observations.

  • Annotation signal lengths now match action lengths. Generation and annotation preserve nonzero exit status when they fail.

  • GR1 generation shares the USD-to-URDF conversion across its independent IK controllers.

  • Docker constrains runtime dependencies, checks package consistency, and supports the pinned cuRobo release with Warp 1.16.

  • The conda installer uses Arena’s locked Sim 6.1.0.0/Lab/PyTorch packages, supports named environments with -n, and includes AutoData’s test tools. It checks prerequisites before creating an environment and refuses to replace an existing environment with -c.

Rebuild the Docker image when upgrading, or rerun ./conda_installer.sh -i -n <environment-name> for a Python 3.12 conda environment. See Installation for both setup paths.

v0.1.0#

This initial release of AutoData provides a standalone framework for transforming a small set of annotated demonstrations into larger robot-learning datasets. It brings data-generation algorithm parity with Isaac Lab Mimic without depending on the Isaac Lab Mimic package and covers the workflow from demonstration annotation through parallel generation, validation, and replay.

Key features of this release include:

  • Generation algorithms: MimicGen for single-arm tasks, DexMimicGen for coordinated bimanual tasks, and SkillGen for collision-aware transit planning with cuRobo.

  • Declarative task integration: Reusable YAML task descriptors and embodiment configurations define subtask signals, source-selection strategies, generation policies, observations, and task-space action mappings without requiring an Isaac Lab Mimic environment configuration.

  • Isaac Lab and Arena interoperability: A unified datastream and embodiment-adapter layer works with standard Isaac Lab and Isaac Lab-Arena environments. Environment profiles create scene, reset-randomization, and planner variants without registering duplicate environments.

  • Dataset tooling: Tools for manual and automatic demonstration annotation, parallel data generation, structural HDF5 validation, and replay using Isaac Lab-compatible datasets.

  • Example workflows: Complete tutorials and pre-annotated source datasets for Franka cube stacking with MimicGen, Fourier GR-1 and Unitree G1 pick-and-place with DexMimicGen, and Franka cube and bin stacking with SkillGen.

  • Reproducible installation: A Docker development environment for MimicGen and DexMimicGen, an opt-in cuRobo image for SkillGen, and an optional conda installation with pinned Isaac Sim, Isaac Lab, and Isaac Lab-Arena revisions.

  • Migration and developer documentation: A migration guide for moving Isaac Lab Mimic tasks to AutoData, a published platform support matrix, and unit, end-to-end, and data-generation performance tests.

Known limitations:

  • Platform support: AutoData currently supports Linux x86_64 systems with an NVIDIA RTX GPU. See the support matrix for the complete hardware and software requirements.

  • SkillGen: SkillGen is currently single-arm only and requires the optional cuRobo installation. Bimanual collision-aware generation is not supported.

  • Dataset validation: The validation tool checks HDF5 structure and metadata only. Semantic correctness, task success, and action reproducibility must be confirmed by replaying the dataset in simulation.