The data-sourcing umbrella for robium. Before any policy gets trained, something\nhas to decide where the training data comes from — an existing hub dataset, data\ngenerated in simulation, or demonstrations collected via teleop on a real robot —\nand how it will be stored and versioned once it exists. This skill owns that\nselection and the cross-cutting sourcing rules; it does not own hub mechanics\n(huggingface), the LeRobotDataset format (lerobot), or the mechanics of\ngenerating synthetic data inside a simulator (isaac-sim, gazebo). It also\ndoes not own training itself — that is lerobot and isaac-lab's territory.
environments is a required early step for reproducibility.huggingface.\nThis skill decides which dataset or source strategy to use; it does not\nown hub auth or transfer mechanics.lerobot. This skill decides whether to record real\ndemonstrations at all; lerobot owns how a recording actually happens.isaac-sim or gazebo. This skill\ndecides whether sim-generated data is the right call for a task.lerobot (or isaac-lab for\nthe NVIDIA RL stack).architect (routes here).1. Define the task and embodiment precisely — robot morphology, action\nspace, camera views, task description. This is the search key for step 2 and\nthe compatibility check for step 3.
\n2. Search for an existing dataset first. Check the Hub's robotics/LeRobot\ntags and Open X-Embodiment for a dataset matching the task and embodiment\n(mechanics: huggingface). If one exists and the embodiment matches, use it\ndirectly — skip to step 5.
3. If no match, decide sim-generation vs teleop using the trade-off table\nin Decision guidance. Route to isaac-sim or gazebo for sim-generation\nmechanics, or lerobot for teleop-based recording mechanics.
4. Plan storage and versioning before collecting anything. Decide the\ntarget dataset format (LeRobotDataset, mechanics in lerobot) and where it\nwill be versioned (a Hub repo with explicit revisions, mechanics in\nhuggingface) so episodes land in their final shape from the first one\nrecorded, not migrated after the fact.
5. Record the chosen source strategy in the project's architecture brief\n(the section architect maintains) so later phases don't re-litigate it.
Offline-first funnel:
\nSearch Hub + Open X-Embodiment for the task/embodiment\n│\n├─ Match found, embodiment matches → use it directly (huggingface + lerobot)\n├─ Partial match (task ✓, embodiment ✗) → candidate for co-training/fine-tune,\n│ not a drop-in; still need new data\n└─ No match → choose sim-generation or teleop below\nSim-generation vs teleop/real-robot collection trade-offs:
\n| Factor | \nSim-generated | \nTeleop / real-robot | \n
|---|---|---|
| Cost per episode | \nLow — scales to thousands of episodes with compute, not human time | \nHigh — a human operator per episode, hardware wear | \n
| Scale | \nEasy to get large volumes via domain randomization | \nBounded by operator time; large datasets are expensive | \n
| Realism / sim-to-real gap | \nReal risk — visual and physics gaps unless deliberately closed (domain randomization, matched sensor noise) | \nGround truth by construction — no sim-to-real gap | \n
| When to prefer | \nEarly iteration, pretraining, cases where large scale matters more than perfect fidelity | \nFinal validation, tasks with contact-rich or hard-to-simulate dynamics, or when the sim-to-real gap can't be closed cheaply | \n
| GPU requirement | \nisaac-sim route needs the NVIDIA RTX GPU floor; gazebo route does not | \nNone beyond the target robot and a recording workstation | \n
A common effective pattern is both: bulk sim-generated episodes for scale and\ncoverage, plus a smaller teleop set for real-world validation and to measure\n(and later close) the sim-to-real gap. Decide the mix explicitly and record it\nrather than defaulting to only one source.
\nisaac-sim for data generation means meeting its NVIDIA RTX GPU floor first\n(see that skill's Key directives); choosing gazebo does not require a GPU.\nDon't plan a sim-generation-heavy data strategy around Isaac Sim before the\nGPU question is confirmed — fall back to gazebo or a teleop-heavy plan\notherwise.lerobot's own gotchas) a working\nkeyboard/input teleop path that doesn't fully work over a headless/Wayland\nsession — plan collection sessions on a machine with a real display and\ninput device attached.gazebo-generated data, since teleop collection for\nnavigation is comparatively cheap (no arm/gripper precision required).lerobot).huggingface (hub operations), lerobot (LeRobotDataset\nformat and recording mechanics), isaac-sim and gazebo (synthetic-data\ngeneration mechanics), isaac-lab (RL training that consumes this data),\narchitect (routes here, records the sourcing decision in the brief).