pi05-spline-field-masked-putcab-e044-r002-2k

Pi0.5 masked Spline Field - RoboTwin put_object_cabinet (E044-R002)

Public full-state checkpoints from protocol-native run E044-R002.

• Task: RoboTwin 2.0 put_object_cabinet
• Training dataset: Shiki42/robotwin_put_obj_cabinet_50_dynFcam_nFov_dynamicMain_lerobot
• Dataset revision: 6280b6f02751e55dc902299ddf1464a4e7534707
• Producer code commit: 8771d8f1605d87bb95b07ec67204e4329f7619ed
• Train config: pi05_putcab_spline_field_masked_anchor_adapt_matched_lora
• Seed: 87431
• Batch size: 16
• Training: step zero to 2,000 optimizer updates
• Checkpoints: 500, 1000, 1500, 2000

The training repair propagates LeRobot action_is_pad, performs valid-timestep weighted Spline fitting, weights coefficient supervision by B-spline basis energy, and normalizes loss by valid per-sample weight.

Status

Training completed with finite loss/gradient metrics and successful full-state save/reload checks. The frozen offline action-quality audit is still pending at publication time because the shared GPU is occupied. No Spline Field rollout success rate is claimed yet.

Integrity

checkpoint_inventory.json records every file size and SHA-256 plus a tree SHA-256 for each checkpoint. These are full-state checkpoints containing params, optimizer/train state, and normalization assets.

Expected tree SHA-256:

• step 500: 3ffdad8bf7816e3ef704f628b3dc508df391b717dd433914117330f7baf73416
• step 1000: 9b7091bd4aeb0b19c511f0c09404d25d8e5e8505563313d41fdc5cb948839dbb
• step 1500: db1530a0393a6e9730abce9a08ab67de35b0090221cd0bcdde3dad6d4de90c30
• step 2000: 398a489bc43dc1b57fbf97be86a95246cc7fbc27719c1e763dcb1bbefbdafbd1