Add post-hoc best saved epoch for ToolHang mixed100 (training stopped by the user at epoch 379/1000): epoch 300, mean of 5 periodic 50-episode evaluations at epoch-20..epoch+20 = 29.2%. Same rule as focused100/diverse100. best_bc.json and ToolHang/README.md updated; original checkpoint bytes, sidecars and evaluation records retained.
25 days ago
Publish best saved ToolHang BC50/BC100 EMA checkpoints
Add post-hoc best saved epochs (max mean of 5 periodic 50-episode evaluations around each saved epoch): focused50 ep2500, diverse50 ep200, mixed50 ep1500, focused100 ep300, diverse100 ep300 (training in progress). Final epoch3000 BC50 weights unchanged. Original checkpoint bytes, sidecars and evaluation records retained; SHA256 manifests updated.
27 days ago
Publish Transport diverse50 final BC EMA checkpoint (epoch 3000)
Add Transport diverse50 epoch3000; seed 1, H8, 45 train + 5 validation; final 30/50 success. Preserve the five previously published models. Original checkpoint bytes, metadata, evaluation records and SHA256 manifests retained.
28 days ago
Publish ToolHang diverse50 and Transport mixed50 final BC EMA checkpoints
Add ToolHang diverse50 and Transport mixed50 epoch3000; seed 1, H8, 45 train + 5 validation. Preserve the three previously published models. Original checkpoint bytes, metadata, evaluation records and SHA256 manifests retained.
29 days ago
Publish three completed distribution BC EMA checkpoints (epoch 3000)
ToolHang focused50/mixed50 and Transport focused50; seed 1, H8, 45 train + 5 validation. Original checkpoint bytes, metadata, evaluation records and SHA256 manifests retained.
Initialize Robomimic study1 and study2 directory structure