yypeng666/metamdp-isaac-conveyor-expert

Meta-MDP Isaac Conveyor Expert

Expert conveyor pick-and-place recordings from Isaac Sim / Isaac Lab with Newton physics and RTX rendering. Access requests require manual approval.

Delivery status

Normal archives and all 447 original staged files passed complete pinned-commit Hub readback, including every original archive member. Portable reader recovery has passed; the storage acceptance receipt is published separately from deferred model-input admission. Collection, archive integrity, Hub readback, portable recovery and model-input qualification are separate gates. The actual VLA adapter audit is deferred by the owner; these files do not claim complete SFT admission or trained policy performance.

Contents and splits

• releases/normal/: 200 training episodes (seeds 10000..10199), five development episodes (20000..20004); ten balls each, all-success collection gates.
• releases/dense/ and releases/sparse/: supplemental collection was cancelled by the owner. These folders contain only status and frozen plan provenance, no accepted episode archives. Dense stopped during its first 64-world group; Sparse never started. Future collection requires an explicit new run.
• views/normal-only-sft.json: explicit Normal selection and admission status.

Final-test seeds 30000..30024 are reserved and are not included. Same-seed pairing across loads stays within a split. Do not mix reserve loads into SFT by globbing.

Each archive contains one original episode directory, dual-camera H.264 video, full physics traces, actions/states/events/frame indices, configuration bindings and the original collection gate. Adjacent backup receipts bind the archive and all uncompressed members. RGB is native 640x640 at real 25 FPS, CRF20/yuv420p; lossy video decoding is not byte-identical to online raw RGB. Physics and control are 500 Hz and 50 Hz. Recording rate does not grant free actor observations.

Recovery and provenance

Source code is maintained only in the private Git repository latency_meta_control_isaac, branch codex/c0-environment-assets. Dataset approval does not grant Git access. See code-location.json for the collection commit and code/data boundaries. The implementation is published at Git commit 6e129a9524bb873d8d856943c9a3ab2ce904bc92. Download the fixed data revision c05275130b3dd6b47ee4127cb134c961d5746761, unpack once with scripts/prepare_teacher_data.py --data-root <download>/releases/normal, and use TeacherDataset(data_root). See the Git checkout's docs/CORPUS_RECOVERY.md for exact commands. No separate restoration directory, index regeneration or normalization rebinding is required. The loader resolves paths locally and consumes the original index/statistics unchanged.

The collection plan retains the original Git commit and per-file SHA256 identities. No source archive is needed or distributed. Preserve original data and receipts.

At the owner's request, the former source archives and repository history were removed. Old Hub commits and source archive hashes in immutable historical receipts identify the original checks; they are not current download targets. Original data, indexes, normalization and acceptance receipts remain unchanged.

No simulator installation, upstream model weights, credentials or model training is included. Legacy RoboSuite/LIBERO task checkpoints are not initializations for this Isaac corpus. No new redistribution license for third-party assets is granted.

Portable recovery acceptance

Normal storage and alternate-root reader recovery passed for all 205 episodes and all 235643 sample rows. See releases/normal/ACCEPTANCE.json. Original archives and numeric normalization remain unchanged. Formal OpenPI input admission, positive forecasts, new-host environment setup, training and learned-policy demos remain separate open gates. dataset_admitted remains false. Dense/Sparse collection remains cancelled.