Shiki42/ctr-dp-scan-object-ctr-noidle-e621-step64749

CTR DP Scan Object — CTR without idle mask (uniform50)

This public inference checkpoint is the verified final checkpoint from CTR training Experiment E621, which is linked from the core experiment data table.

• Architecture: LeRobot Diffusion Policy
• Task: RoboTwin Scan Object
• Training variant: CTR without idle mask (uniform50)
• Final optimizer update: 64749
• Dataset: Shiki42/ctr-scan-object-uniform50-20260917
• Dataset revision: a34f2a32d3738e3607f226a777b41d42d0522908
• Source Experiment: E621

IdleMask disabled. The repository contains inference parameters, processors/assets, resolved configuration, and a generated SHA256SUMS manifest. Optimizer state, data-loader state, logs, credentials, and training caches are intentionally excluded.

Results remain governed by the source Experiment and its separate evaluation Experiments. Publication does not upgrade a reported or audit-pending result to archived/confirmed.

Archival identity and evaluation version mapping (2026-10-05)

Scientific weights/configuration/actual normalization unchanged. E607 original central Experiment record unchanged. The historical inference revision below remains the weights identity. Only evaluation bindings for this exact step/revision are listed; unresolved receipt references stay explicit.

{
  "experiment": "E621",
  "training_run": "E621-R001",
  "step": 64749,
  "task": "scan-object",
  "algorithm": "dp",
  "idle_mask": false,
  "data_condition": "noidle",
  "revision": "4937169e1c5e62cf1342a856cb35e4baedfbcc16",
  "package_manifest_sha256": "c09a24bfc180dc16eeff92af6ee5222ab047f812acb7776e9f0f59e0669a3346",
  "training_configuration": {
    "ablation": "IdleMask ablation (user-approved 2026-09-21 for Q4-A): same corrected CTR Uniform50 revision as the with-mask CTR column, training with IdleMask DISABLED; observation.arm_active_mask must not be loaded or consumed; only ordinary temporal padding mask is applied. Purpose: isolate the contribution of IdleMask in the Q4-A generality table.",
    "budget_basis": "E354 DP 600-epoch paper-budget recipe (batch128 as 4x32)",
    "budget_unit": "eligible frame-window exposure; exact batches carry epoch tails",
    "dataset": {
      "dataset_status": "formally_usable_ctr_dataset_mask_not_consumed",
      "episodes": 50,
      "fps": 25,
      "frames": 14163,
      "idle_mask": false,
      "publication_experiment": "E519",
      "repo_id": "Shiki42/ctr-scan-object-uniform50-20260917",
      "revision": "a34f2a32d3738e3607f226a777b41d42d0522908",
      "sample_range": "all 50 episodes",
      "schedule": "uniform50",
      "manifest_sha256": "e321f63b608403ac36c657cca1aaee7584452d0f8ff47ae3898f4ab18fa60e70",
      "manifest_source": "E629 host plan.json (same pinned revision a34f2a32, verified live on 39051)"
    },
    "drop_n_last_frames": 7,
    "effective_batch_size": 128,
    "epochs": 600,
    "estimated_gpu_hours": 5.02,
    "gradient_accumulation": 1,
    "idle_mask": false,
    "micro_batch_size": 128,
    "normalization": "Official LeRobot STATE/ACTION MIN_MAX and ImageNet MEAN_STD; global bounds verified in CPU prepare; quantiles not consumed (as E354)",
    "policy": "DP",
    "seed": 87431,
    "task": "Scan Object",
    "trainer": "./scripts/train dp (official LeRobot Diffusion Policy, lerobot044 runtime)",
    "training_samples": 13813,
    "training_variant": "noidle",
    "updates": 64749,
    "recipe_source": "E354 single_task_budget DP recipe: micro_batch=128 direct (no accumulation), qualify_budget_and_statistics computes/verifies updates in CPU prepare"
  },
  "dataset": null,
  "actual_normalization": null,
  "eligibility": null,
  "evaluation_bindings": [
    {
      "experiment": "E622",
      "title": "Q4-A Scan Object / DP / CTR without idle mask：默认评测",
      "checkpoint_step": null,
      "protocol": {
        "checkpoint_selection": "final checkpoint selected a priori from the source training Experiment",
        "condition": "default",
        "metrics": "success rate and Tcap",
        "optimizer_updates": 0,
        "policy": "DP",
        "protocol_status": "adapted_to_frozen_uniform_suite_ctr.uniform_lerobot_eval",
        "scope": "result",
        "section": "Q4-A",
        "task": "Scan Object",
        "training_variant": "noidle",
        "protocol": {
          "episodes": 100,
          "manifest": "eval/robotwin/suites/paired-uniform-native100-20260917/scan_object.json",
          "native_action_limit": "task_config/_eval_step_limit.yml scan_object (not re-verified at registration)",
          "prompt": "Scan the object.",
          "reference_experiments": [
            "E238-R006",
            "E257"
          ],
          "sha": "056d214f65912164583b4bde806e46d9f2c97fe99d9eef17788dfad2fff6874c",
          "suite": "paired-uniform-native100-20260917/scan_object",
          "suite_state": "frozen 100 scenes; expert_checked=true"
        },
        "protocol_reference": "Fixed native full-task New FOV evaluation, 100 frozen scenes, native action limit, no scene replacement or retry (see E592/E540/E534)",
        "adaptation_note": "2026-09-21 user authorised adapting the ACT/DP evaluator to the E238/E257 frozen uniform suite (same scenes, RoboTwin 9125005d); run spec experiments/run_specs/E622/plan.json."
      },
      "runs": [
        {
          "id": "E622-R001",
          "status": "completed",
          "code_commit": null,
          "result": null,
          "result_uri": null
        }
      ],
      "step_binding": "verify raw evaluation receipt"
    }
  ],
  "preservation": "Keep every historical version. Deprecation flags remain visible and do not constitute restored qualification. Missing exact evaluation step must be resolved from raw receipt, not inferred."
}