---

name: roboverse description: "Routes RoboVerse research and engineering tasks across simulation, task authoring, learning, benchmark integrations, and cross-simulator parity." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

RoboVerse

Use this skill when a task involves RoboVerse, roboverse-py, roboverse_pack, RoboVerse tasks/robots/scenes, MetaSim package discovery, robot learning pipelines, LIBERO/ManiSkill/MJLab/robosuite/RobotWin/SimplerEnv integrations, or simulator parity.

RoboVerse is the downstream content, learning, dataset, benchmark, and example layer. MetaSim is the upstream owner of core scenario/config types, task registry, simulator handlers/backends, and environment abstractions. Keep that boundary explicit before editing code.

Start here

1. Read references/installation-and-boundaries.md for package identity, install variants, and the MetaSim/RoboVerse ownership boundary. Run the smallest import check before backend work.
2. Route by intent:
   • simulation-workflows for ScenarioCfg, robots, scenes, grounds, assets, cameras, queries, randomization, teleop, rendering, and basic execution.
   • task-development for new/changed tasks, registration, observations, rewards, resets, callbacks, and tests.
   • learning-pipelines for RL, IL, VLA, fusion, datasets, checkpoints, runners, and evaluation.
   • benchmark-integrations for benchmark metadata, data/demo conversion, replay, and external stacks.
   • parity-and-tooling for measured cross-simulator comparisons, registration audits, and diagnostic scripts.
3. Read references/troubleshooting.md whenever imports, optional dependencies, assets, data/config validation, or backend behavior fails.
4. Check references/repo-provenance.md before treating this skill as current for a changed checkout.

Minimal checks

python -m pip install -e ".[mujoco]"
python -c "import roboverse_pack, metasim; print('RoboVerse imports OK')"
python -c "import torch; print(torch.cuda.is_available())"  # report, do not assume

Choose dev for focused tests and examples for tutorial helpers. Add learn, vla, or a simulator-specific extra only for the selected workflow. Do not claim every advertised simulator or external benchmark is verified from a CPU import or from one MuJoCo run.

Operational rules

• Validate and normalize at boundaries; fail clearly for unsupported task, robot, backend, data, or config values.
• Run end-to-end before claiming numerical parity. Report exact backends and measured deltas; closed-loop policy transfer is a separate claim.
• Prefer additive task-family extensions and existing config composition.
• Keep GPU, display, external-data, credential, real-robot, and long-training actions explicit and bounded. The skill does not provide deployment advice.