# Collection quality

Use this when episode count is growing but the learned behavior is not.

## Coverage before volume

- Define the smallest workspace and variation needed by the application, then
  sample it densely. Add range only after the policy is reliable inside that
  envelope.
- In Robium's VLA trial, 50 episodes spread over roughly 30 cm failed to locate
  grasps reliably. Seventy-five episodes over roughly 10 cm reached 80%
  success. These are observations from that task and embodiment, not universal
  collection targets.
- Track coverage across object pose, robot pose, camera visibility, action
  range, and failure-relevant conditions. Raw episode count hides holes in all
  of them.

## Demonstration integrity

- Store an explicit success result for every episode.
- In a successful-expert dataset, when a scripted oracle misses, retry or
  discard the episode. Do not label a failed terminal state as a successful
  expert demonstration. Corrective, recovery, DAgger, and failure-learning
  collections may retain it under an explicit, different label and sampling
  plan.
- Put a bounded-attempt or minimum-success-rate guard around automated
  generation. A broken oracle should fail the run rather than loop until it
  eventually reaches a requested count.
- Review teleoperation for accidental pauses, resets, camera occlusion, and
  inconsistent task completion before publishing or training. LeRobot owns
  device, keyboard, desktop-permission, and headless teleoperation mechanics.

## Validate before scaling

- Load a small slice through the intended training stack and verify feature
  names, shapes, rates, temporal alignment, and episode boundaries.
- Train a cheap smoke run only to expose schema and pipeline failures. Do not
  interpret smoke-run quality as evidence that the dataset is sufficient.
- Evaluate by condition, not only aggregate success, so missing coverage is
  visible.

---
name: data
description: Choose and structure training data for robot-learning projects.
---

# Data

Begin with coverage: decide what behavior, embodiment, and conditions the
policy must see before deciding how to collect them.

## Choose the source

- Search existing datasets first. Confirm task, action space, degrees of
  freedom, gripper, cameras, state features, timing, license, and provenance.
- Use an exact embodiment match directly. Treat a near match as pretraining or
  co-training material, not a drop-in dataset.
- When schemas and task labels cannot distinguish two environments, compare a
  deterministic scene observation from the dataset with the pinned target
  environment. Prefer a stable reference camera over a randomized wrist view.
- Generate in simulation when scale, controlled variation, or labels matter
  more than perfect realism.
- If a documented search finds no dataset for the exact scene and control
  contract, generate demonstrations in the pinned application environment and
  retain only episodes that satisfy its success condition.
- Collect on the real robot when contact, appearance, or hardware behavior is
  difficult to reproduce faithfully.
- Mix sources deliberately: simulation can provide coverage; a smaller real
  set can expose the remaining sim-to-real gap.

## Protect the useful signal

- Constrain the task and workspace before adding more episodes. Dense coverage
  of the behavior matters more than a large headline episode count.
- For a successful-expert imitation dataset, keep only demonstrations that meet
  the task's success definition. Retry or discard oracle failures, and stop
  loudly if the success rate collapses. Do not apply this rule to DAgger,
  corrective, recovery, or failure-learning datasets that intentionally retain
  non-expert transitions.
- Define the episode boundary, observations, actions, rates, success label,
  splits, and target storage format before collection starts.
- Record the source revision and collection conditions. Dataset facts and
  licenses must come from the current card or repository, not memory.

## Go deeper only when needed

- For the Robium evidence behind workspace density and demonstration quality,
  read [COLLECTION-QUALITY.md](COLLECTION-QUALITY.md).
- Use Hugging Face guidance only when the decision reaches Hub discovery,
  inspection, transfer, or publication.
- Use LeRobot guidance when the decision reaches LeRobotDataset recording,
  editing, training, evaluation, or platform-specific teleoperation controls.
- Use simulator guidance only after choosing simulation as a source; Isaac Sim
  and Gazebo own their generation mechanics.
- Test fixtures belong to test-assets, not this training-data decision.

## Done

- The chosen sources cover the target embodiment and task, the gaps are named,
  and the first small collection or dataset slice can validate the plan before
  scale or paid compute.