---

name: data-and-replay-buffers description: "Inspect Diffusion Policy dataset interfaces, ReplayBuffer stores, SequenceSampler horizons and padding, and normalizer contracts." metadata: disco-role: operating disable-model-invocation: true license: MIT

data-and-replay-buffers

Use this sub-skill when you need to understand or validate Diffusion Policy data stores, dataset adapters, sequence sampling, and normalization without touching training loops or policy internals.

Use for

• zarr or zip ReplayBuffer inspection
• episode counts, episode_ends, shapes, dtypes, chunking, and length mismatches
• low-dimensional versus image dataset samples and shape_meta routing
• SequenceSampler horizon, padding, episode_mask, and key_first_k behavior
• normalizer selection, key lookup, and dataset conversion routes

Route elsewhere

• training, evaluation, Ray multiruns, and metrics aggregation -> training-and-evaluation
• policy inference, checkpoint logic, model families, and losses -> policies-and-models
• live robot capture, RealSense, UR5, SpaceMouse, and safety-gated execution -> real-robot-operations

Fast validation

1. Run python scripts/inspect_replay_buffer.py --path <replay-buffer> --max-keys 12 from this sub-skill directory to inspect a zarr directory or zip store without modifying it.
2. Add --json when another tool should consume the summary.
3. Use the checklist in references/troubleshooting.md when episode lengths, keys, or shapes are inconsistent.

What to read next

• references/data-and-replay-buffers.md
• references/api-reference.md
• references/troubleshooting.md

Bundled helper

• scripts/inspect_replay_buffer.py — read-only inspector for ReplayBuffer zarr directories and zip stores.

First checks when data looks wrong

• Compare meta/episode_ends[-1] to every data/* leading dimension.
• Confirm the sample shape matches the dataset family: flat low-dim tensor versus dict of image and state keys.
• Confirm the normalizer keys match the returned sample keys exactly.
• Treat repeated boundary frames as expected padding, not corruption.