| { |
| "evidence_tier": "development; one training seed; final checkpoint only", |
| "episodes": 512, |
| "sustained_successes": 332, |
| "success_rate": 0.6484375, |
| "by_family": { |
| "Lift": { |
| "episodes": 64, |
| "sustained_successes": 64, |
| "success_rate": 1.0, |
| "wilson95": [ |
| 0.9433759402071946, |
| 1.0 |
| ] |
| }, |
| "PickPlace": { |
| "episodes": 64, |
| "sustained_successes": 40, |
| "success_rate": 0.625, |
| "wilson95": [ |
| 0.502503786840045, |
| 0.7333401982117536 |
| ] |
| }, |
| "Stack": { |
| "episodes": 64, |
| "sustained_successes": 43, |
| "success_rate": 0.671875, |
| "wilson95": [ |
| 0.549991058490915, |
| 0.7742944209553081 |
| ] |
| }, |
| "Door": { |
| "episodes": 64, |
| "sustained_successes": 34, |
| "success_rate": 0.53125, |
| "wilson95": [ |
| 0.41072087051471007, |
| 0.6482401257482395 |
| ] |
| }, |
| "NutAssembly": { |
| "episodes": 64, |
| "sustained_successes": 32, |
| "success_rate": 0.5, |
| "wilson95": [ |
| 0.3810209474394701, |
| 0.6189790525605299 |
| ] |
| }, |
| "ToolHang": { |
| "episodes": 64, |
| "sustained_successes": 9, |
| "success_rate": 0.140625, |
| "wilson95": [ |
| 0.0757857255603086, |
| 0.24616281741577023 |
| ] |
| }, |
| "Push": { |
| "episodes": 64, |
| "sustained_successes": 61, |
| "success_rate": 0.953125, |
| "wilson95": [ |
| 0.8710034625990631, |
| 0.983930983213707 |
| ] |
| }, |
| "Drawer": { |
| "episodes": 64, |
| "sustained_successes": 49, |
| "success_rate": 0.765625, |
| "wilson95": [ |
| 0.6486674241743136, |
| 0.8525010440607586 |
| ] |
| } |
| }, |
| "success_rule": "10 consecutive physical steps satisfying original task predicate", |
| "seeds": [ |
| 30000, |
| 30063 |
| ], |
| "test_access": false, |
| "trained_calls_audited": 45254, |
| "trained_steps_audited": 365450, |
| "base_pickplace": { |
| "episodes": 64, |
| "sustained_successes": 0, |
| "success_rate": 0.0, |
| "wilson95": [ |
| 0.0, |
| 0.05662405979280534 |
| ] |
| }, |
| "paired_pickplace": { |
| "base_successes": 0, |
| "trained_successes": 40, |
| "both_success": 0, |
| "trained_only": 40, |
| "base_only": 0 |
| }, |
| "interface_difference": "Original: one RGB, chunk1, discrete diffusion4, bridge_orig. Adapted: two RGB+proprio8, chunk8, Flow Matching4, panda_multitask_v1." |
| } |
|
|
