yangfy0627's picture
Visual BCs: use epoch 1400 for all nine locally trained jobs (user decision 2026-10-27); add ToolHang diverse100; archive prior releases
cc52a28 verified
|
Raw History Blame Contribute Delete
7.6 kB

Visual-only BC: reviewed selections

2026-10-26: 12 finalized selections in this repository (8 local, 4 assigned to the other host). DrawerCleanup mixed50 and Transport focused50 now have frozen choices after a completed 2000-epoch phase and an additional common100 candidate-validation panel, followed by reporting-only final200. Eight jobs remain assigned to the other host; 4 now have finalized records in this index. The present two-checkpoint review did not re-evaluate those other-host models.

2026-10-26 update: ToolHang focused100 was reselected to e800 (final200 17.5%), using candidate validation over e250鈥揺1600. The prior e250 release (final200 15.5%) is archived under its releases/ folder. Review.

2026-10-27 update: by user decision all nine locally trained BCs (Threading, DrawerCleanup, LiftTray diverse/mixed; Transport focused50; ToolHang focused100 and diverse100) now use epoch 1400. ToolHang diverse100 is newly published. Their final200 reused the panel of the prior selections and was seen before the decision, so for these rows it is decision evidence, not an independent test. Prior releases are archived under each job's releases/before_20261027_e1400 folder. Paired comparison. Other-host records are unchanged.

Task / cohort Selected or historical epoch Final200 Status Checkpoint
Can / diverse20 1400 57.5% Historical candidate; NOT finalized weights
Can / mixed20 1400 53.0% Historical candidate; NOT finalized weights
DrawerCleanup / diverse50 1400 57.0% User decision: uniform e1400 weights
DrawerCleanup / mixed50 1400 70.5% User decision: uniform e1400 weights
LiftTray / diverse50 1400 39.0% User decision: uniform e1400 weights
LiftTray / mixed50 1400 43.0% User decision: uniform e1400 weights
Threading / diverse100 1400 27.0% User decision: uniform e1400 weights
Threading / mixed100 1400 28.0% User decision: uniform e1400 weights
ToolHang / diverse100 1400 23.5% User decision: uniform e1400 weights
ToolHang / diverse50 1400 12.5% Historical candidate; NOT finalized weights
ToolHang / focused100 1400 24.0% User decision: uniform e1400 weights
ToolHang / focused50 1400 15.5% Historical candidate; NOT finalized weights
ToolHang / mixed100 1400 23.0% Historical candidate; NOT finalized weights
ToolHang / mixed50 1400 13.0% Historical candidate; NOT finalized weights
Transport / diverse50 1400 57.5% Historical candidate; NOT finalized weights
Transport / focused50 1400 41.0% User decision: uniform e1400 weights
Transport / mixed50 1400 44.0% Historical candidate; NOT finalized weights

Selection combines training maturity, validation behavior and success. These practical choices do not establish zero overfitting or a global success maximum. Final200 never drives reselection; the new candidate-validation100 is explicitly selection data. Periodic neighborhood scores average different weights on reused seeds.

The previous LiftTray mixed50 e250 release and its final200 (41%) remain archived; selected e1250 scored39.5%, which is not evidence of a success improvement. See its reporting-only comparison.

Latest two-task curves 路 Candidate comparison 路 Prior phase review. All new weights use collision_geoms=false. Original sealed configs, recorded runtime schedules and prior releases are preserved. Earlier collision-rendered root trees remain unchanged.

2026-10-27: three accepted epoch1250 BCs

Training was stopped by the user; these weights were fixed before final200 and accepted after review. Transport mixed50 is excluded pending reselection.

Task / cohort Epoch Final200 Checkpoint
Transport / diverse50 1250 50.0% (100/200) weights
ToolHang / mixed100 1250 19.0% (38/200) weights
ToolHang / mixed50 1250 15.5% (31/200) weights

2026-10-27: Transport Mixed50 finalized

The previously pending Transport mixed50 selection is now epoch2000, accepted by the user after comparing the shared200-episode panel. Actual original terminal evaluation: 40.5% (81/200). This panel was used for selection.

Weights 路 Selection and comparison.

2026-10-27: current selections updated to epoch1400 for all eight local BCs

This update supersedes the eight earlier selections below; historical weights remain available. All scores are actual200-episode results on the reused comparison panel. The user chose a uniform epoch1400, including cohorts with lower or equal scores.

Task / cohort Current epoch SR Successes Weights
Transport / diverse50 1400 57.5% 115/200 weights
Transport / mixed50 1400 44.0% 88/200 weights
ToolHang / mixed100 1400 23.0% 46/200 weights
ToolHang / focused50 1400 15.5% 31/200 weights
ToolHang / mixed50 1400 13.0% 26/200 weights
ToolHang / diverse50 1400 12.5% 25/200 weights
Can / diverse20 1400 57.5% 115/200 weights
Can / mixed20 1400 53.0% 106/200 weights