Release Notes#

v0.3.0#

This release introduces:

  • Prompt-first environment generation.

  • Controlled randomization through variations.

  • Sensitivity analysis.

  • Multi-node evaluations.

  • Expanded policy integrations.

  • Task libraries, including RoboLab-style tasks and our own Kitchen Benchmark for tasks in realistic kitchens.

Details below:

Features and improvements

  • Agentic environment generation: Added EnvironmentGenerationAgent for creating environments from natural-language prompts, with intent-to-graph conversion, SimReady asset search, composite-task support, configurable inference endpoints, and an interactive review GUI with simulation snapshots and relation-solver previews (#718, #770, #803, #804, #805, #868, #982, #1050).

  • Typed environment definitions: Added Pydantic-based ArenaEnvGraphSpec YAML parsing and CLI bring-up, typed environment and policy configurations, a reusable environment factory, and a unified path for loading registered Python environments or graph YAML files (#690, #749, #757, #808, #858, #860, #862, #863).

  • Isaac Lab 3.0 workflow interoperation: Updated the Isaac Lab submodule and moved demonstration recording, replay, and teleoperation to Isaac Lab’s scripts through Arena’s external environment registration callback (#960, #994, #1022).

  • Experiments and evaluation reports: Replaced legacy evaluation jobs with typed YAML Experiments and Runs, including environment rebuilds, chunked and multi-node dispatch, combined JSON results, an experiment viewer, grouped success-rate plots, and generated evaluation reports (#751, #772, #802, #867, #873, #896, #1042, #1056, #1070).

  • Variations and sensitivity analysis: Added Hydra-configurable variations with in-memory recording for lighting, camera intrinsics, and object mass, plus an MNPE/NPE sensitivity-analysis workflow, sample dataset, and plotting tools (#729, #746, #755, #775, #789, #913, #916, #1055).

  • Object placement and validation: Added heterogeneous and deterministic placement, mesh-based non-collision constraints, NotNextTo and FaceTo relations, unified placement APIs, build-time physics settling, and pluggable validation checks (#676, #679, #732, #769, #771, #809, #888, #927).

  • Robot reachability validation: Added a build-time, simulation-free cuRobo IK gate for task-relevant object layouts, including Droid and Franka embodiment placement, collision checking, and optional debug visualization (#914, #947, #954, #984, #1014).

  • Policy integrations: Added OpenPI support for DROID and Pi0.5, GR00T remote closed-loop policies with pluggable action schedulers, Cosmos and DreamZero policies, and OSMO workflows for parallel and multi-policy evaluation (#595, #655, #658, #663, #688, #823, #843, #963, #991, #1005).

  • Task libraries: Added the RoboLab catalog and more than 15 pick-and-place environments, plus Lightwheel and Replicator kitchen environments for pick-and-place, opening doors and microwaves, pressing buttons, and turning knobs (#845, #857, #962, #965, #966, #1048, #1149, #1155).

  • Progress tracking: Added fine-grained subtask state, episode recording, order-independent composite tasks, unified success checks, and normalized progress scoring (#677, #758, #810, #822, #840, #1137).

Documentation

  • Agentic generation workflows: Added concept, validation, tabletop, kitchen, GUI, reachability, and collision-handling documentation for prompt-first environment creation (#844, #1025, #1028, #1033, #1046, #1064, #1076, #1085, #1135).

  • Evaluation workflows: Added documentation for Experiments and Runs, generated reports, multi-node execution, variations, sensitivity analysis, environment seeds, and RigidObjectSet (#1030, #1045, #1051, #1053, #1063, #1068, #1072).

  • Website and catalogs: Added the Kitchen Benchmark and RoboLab catalogs, redesigned the overview and motivation pages, and updated the release documentation for Arena 0.3 (#1041, #1048, #1089, #1126, #1217, #1222).

Infrastructure and CI

  • Native installation: Added a fully source-based uv installation path with lightweight OpenPI and GR00T clients, runtime package configuration, and dedicated native-uv CI validation (#908, #934, #938, #944, #1084, #1092, #1101).

  • Docker workflows: Added per-clone container naming, host-user command execution, optional container suffixes, and improved Docker support for agent workflows (#643, #760, #762, #763).

  • CI coverage: Added all-environment smoke testing, a GR00T closed-loop end-to-end job, camera test coverage, and improved Isaac Sim shader caching and subprocess isolation (#617, #618, #644, #675, #950).

Assets and tests

  • Kitchen assets: Added Replicator kitchen examples, publicly hosted USD-randomizer kitchens, Lightwheel kitchen registration, mesh-placement examples, and kitchen task configurations (#962, #965, #1088, #1090).

  • RoboLab assets: Added the RoboLab catalog, missing object assets, agent-generated pick-and-place environments, and OpenPI-versus-Cosmos experiment configurations (#815, #817, #820, #845, #857, #1020).

Bug fixes

  • Placement correctness: Fixed relation-solver bounding-box scaling, rotation-around placement failures, IK reachability for relation-placed robots, and geometric pick-and-place success checks (#816, #921, #1081, #1132).

  • Rendering and cameras: Fixed corrupt kitchen rendering, Droid camera quaternion ordering, review-GUI viewport setup, and viewpoint camera poses in parallel environments (#673, #941, #1043, #1223).

  • Evaluation correctness: Fixed normalized progress scores, missing-video report handling, timeout termination semantics, and external environment registration for Isaac Lab workflows (#865, #994, #1017, #1137).

  • Background physics: Fixed kitchen placement collisions, relation anchors in background collision meshes, and physics reset behavior for kitchen and other backgrounds (#988, #1003, #1102, #1180).

Limitations

  • Installation from a published Python package is not yet supported; use the native uv source workflow or Docker.

What’s Next

Future releases will focus on:

  • Enhanced Newton support for benchmarks involving contact-rich insertion, cables, and deformables.

  • Prompt-to-scene, robot, and task workflows.

  • Improved performance for multi-node evaluation.

v0.2.0#

This release introduces major new capabilities including Isaac Lab 3.0 (Newton) support, a sequential task chaining framework, semantic object placement, teleoperation, GR00T N1.6 and DROID integration, RL workflows, and a large set of new tasks and embodiments.

Features and improvements

  • Isaac Lab 3.0 (Newton) upgrade: Updated the framework to Isaac Lab 3.0 (Newton), including an updated interop layer (#464, #533).

  • Sequential task framework: Added SequentialTaskBase class for chaining atomic skills into long-horizon tasks, with mimic and metrics support, user-specifiable final subtask success states, and an example that puts an object into a microwave and closes the door (#289, #323, #337, #365).

  • Semantic object placement: Added a differentiable relation-based object placement solver supporting On, NextTo, AtPosition, and PositionLimits relations, multiple anchors, RotateAroundSolution, and full integration with ArenaEnvBuilder and ObjectPlacer (#328, #354, #358, #362, #387, #574).

  • Teleoperation: Added teleoperation support for G1 loco-manipulation and GR1 using Quest XR hand-tracking and CloudXR; updated to IsaacTeleop v1.1 (#286, #350, #577, #605).

  • GR00T N1.6 and DROID integration: Upgraded GR00T to N1.6 and added DROID dataset support, local inference pipeline, and language instruction support for closed-loop evaluation (#334, #416, #418, #420, #519).

  • New tasks, embodiments, and evaluation capabilities: Added Sorting, FactoryAssembly, TurnKnob, CloseDoor, and AdjustPose atomic tasks; Galbot and Agibot A2D embodiments; G1 WBC-AGILE end-to-end velocity policy; RSL-RL policy evaluation; distributed multi-GPU policy runner; multi-task evaluation job runner (#371, #285, #315, #295, #305, #391, #292, #489, #333, #411, #277, #445, #394).

Documentation

  • README and Getting Started overhaul: Comprehensive README rewrite and full reorganization of the Getting Started section and navigation structure (#480, #505).

  • Concepts and placement docs: Added a humanized concepts page and a dedicated object placement documentation page (#526, #511).

  • RL and evaluation workflows: Added RL workflow docs, policy evaluation section, Newton evaluation example from IsaacLab DexSuite, and GTC DLI workflow docs (#363, #451, #536, #390).

  • External repository and advanced usage: Added a dedicated page for using Arena from an external repository, an advanced custom task example, GR00T closed-loop docs, and DROID usage instructions (#518, #550, #519).

Infrastructure and CI

  • Docker and dependency improvements: Moved Python dependencies from the Dockerfile to setup.py; pinned Newton mujoco version; fixed docker pipx stall; added missing Arena package to NGC docker (#535, #629, #640, #578).

  • Repository governance: Added CODEOWNERS, issue templates (brought over from IsaacLab), SECURITY.md, AGENTS.md, and CLAUDE.md (#583, #584, #585, #586, #492, #454).

Assets and tests

  • RoboLab objects and HDR library: Added the RoboLab asset library and HDR lighting support for use in robolab scenes (#429, #428, #431).

  • New example environments: Added GR1 DLI environment, G1 AGILE tabletop environment, and a Rubik’s cube pick-and-place environment (#385, #562, #421).

  • USD asset paths: Updated object library paths to use the ISAAC_NUCLEUS_DIR prefix and updated USDs to reference the Isaac Sim 6.0 staging bucket (#291, #621).

Bug fixes

  • Sequential task metrics: Fixed subtask success rate metric and desired_subtask_success_state check in sequential task evaluation (#405, #410).

  • RSL-RL evaluation metrics: Fixed an extra episode appearing in policy evaluation metrics when using RSL-RL (#530).

  • Object placement correctness: Fixed bounding box rotation in world frame, rejection of overlapping placements by ObjectPlacer, and initialization of On-relation objects within their parent’s footprint (#400, #439, #538).

  • Recorder dataset filename collision: Fixed filename collisions when multiple recorders write concurrently to a shared /tmp directory (#469).

  • Metrics serialization: Sanitized NumPy types in metrics output to prevent serialization errors (#602).

v0.1.1#

This release includes bug fixes, documentation improvements, CI and infrastructure updates, and several API and workflow enhancements over v0.1.0.

Features and improvements

  • Object configuration: Object configuration is now created as soon as an asset is called, so users can edit object properties before a scene is created (#239).

  • Scene export: Added support for saving a scene to a flattened USD file (#237). Scene export now correctly handles double-precision poses and adds contact reporters when exporting rigid objects (#242).

  • Parallel environment evaluation: Enabled parallel environment evaluation for GR00T policy runner, with documentation for closed-loop GR00T workflows (#231, #236).

  • Episode length: Increased episode length for loco-manipulation to support rollout through box drop (#235).

  • Microwave example: Increased reset openness for the microwave example (#311).

Bug fixes

  • Reference object poses: Fixed reference object poses so they correctly account for the parent object’s initial pose; poses are now relative and composed at compile time (#232).

  • IsaacLab-to-stage path conversion: Fixed a bug when the asset name appeared twice in the prim path (replaced both instances instead of one) (#241).

  • qpsolvers: Patched breakage with Isaac Lab 2.3 due to qpsolvers upgrade by pinning to 4.8.1 (#252).

  • Parallel eval: Removed comments that were breaking the parallel eval run commands (#262).

Documentation

  • Multi-versioned docs: Documentation is now versioned so users can read docs that match their release (#272, #300).

  • Links and structure: Updated README docs link to the public location (#270), corrected doc pointers (#301), and added release warnings (#303).

  • Installation: Private Omniverse/Nucleus access is described on a separate page to clarify it is not required for normal installation (#261).

Infrastructure and CI

  • Runners: Release 0.1.1 CI runners moved from local (Zurich) to AWS (#433).

  • CI workflow: Added YAML anchors to reduce repetition in the CI workflow (#245).

  • Contribution guide: Added signoff requirements for external contributions (#238).

  • Docker: Fixed Dockerfile pip usage and added SSL certificate support for Lightwheel SDK (#449).

  • Tests: Finetuned GR00T locomanip model is now generated on the fly in tests instead of mounting a pre-finetuned models directory, improving public CI compatibility and testing the fine-tuning pipeline (#247).

Assets and tests

  • G1 WBC: Updated G1 WBC embodiment file paths to use S3 (#251).

  • Test assets: Removed internal or custom-only assets from tests: custom cracker box (#234), custom USD in ObjectReference test (#240), internal asset from USD utils test (#244). ObjectReference test now composes USD on the fly via scene export (#240).

v0.1.0#

This initial release of Isaac Lab Arena delivers the first version of the composable task definition API. Also included are example workflows for static manipulation tasks and loco-manipulation tasks including GR00T GN1.5 finetuning and evaluation.

Key features of this release include:

  • Composable Task Definition: Base-class definition for Task, Embodiment, and Scene that can be subclassed to create new tasks, embodiments, and scenes. ArenaEnvBuilder for converting Scene, Embodiment, and Task into an Isaac Lab runnable environment.

  • Metrics: Mechanism for adding task-specific metrics which are reported during evaluation.

  • Isaac Lab Mimic Integration: Integration with Isaac Lab Mimic to automatically generate Mimic definitions for available tasks.

  • Example Workflows: Two example workflows for static manipulation tasks and loco-manipulation tasks.

  • GR00T GN1.5 Integration: Integration with GR00T GN1.5 including a example workflows for finetuning and evaluating the model on the static and loco-manipulation workflows.

Known limitations:

  • Number of Environments/Tasks: This initial is intended to validation the composable task definition API, and comes with a limited set of tasks and workflows.

  • Loco-manipulation GR00T GN1.5 finetuning: GR00T GN1.5 finetuning for loco-manipulation requires a large amount of GPU resources. (Note that static manipulation finetuning can be performed on a single GPU.)

On this page