ctr-pi05-pick-dual-bottles-sequential-idlemask-e307-step10000

<!-- ctr-nonctr-idlemask-deprecated-20260919 -->

DEPRECATED / 已废弃 — E307

Do not use this checkpoint or any evaluation results derived from it.

废弃原因：它们错误地开启了 idle mask，而且 idle mask 的定义也不正确。

This checkpoint was trained on a non-CTR dataset with IdleMask incorrectly enabled, and the IdleMask definition itself was incorrect. All checkpoints from this training run, including intermediate checkpoints and published copies, and all associated evaluation metrics are deprecated. Historical success rates, verification PASS receipts, and archive approvals are retained for provenance only and do not establish scientific validity or eligibility for reuse.

Changing inference settings or the card does not repair the trained checkpoint. Use a newly trained checkpoint with the correct non-CTR training recipe and IdleMask disabled, followed by a separate evaluation.

Deprecation authorized by the repository owner on 2026-09-19. This notice supersedes any earlier eligibility or reuse statements below. Original weights, configuration, and historical results are retained.

---

E307 — Pick Dual Bottles / sequential

Inference checkpoint at 10000 optimizer updates. Training Run: E307-R001.

Closed-loop evaluation has not been reported for this checkpoint.

Training identity

• Dataset: CTR-local/q123-sequential-20260919
• Revision: aa3e76fc5dc959231311465c6478c82c77620b0f
• Camera preset: centered_fovy90 calibrated wrist cameras.
• IdleMask: enabled; inactive arm loss weight zero.
• Exact configuration and source identities: ctr-provenance.json and resolved configuration.

Loading and limitations

OpenPI JAX full parameter tree at pinned source e9ba7b7732a3e66e4bd87d6d3429f7cef6352ead. Use create_trained_policy with the included params, assets, matching resolved configuration and dataset asset ID. Joint targets are delta-encoded during training and decoded once; grippers are absolute commands. CTR loader and inference adapter are required for these RoboTwin conventions.

No optimizer state is included. SHA256SUMS identifies every public payload file. Source checkpoint and training logs remain preserved on the training machine. Training completion and raw evaluation result archival do not establish deployment safety or scientific causal conclusions.