gulucaptain/RoboKata-1111

RoboKata 11.11 · A+B1+Foresight Indexing

This release is prepared for training from Zero-WAM pretrain on 25,000 directed pairs, comprising 13,000 selected old RoboICL pairs and 12,000 RoboKata v2.1 pairs. There are 19,000 Franka-demo pairs and 6,000 ALOHA-demo pairs, with Franka queries.

Demonstrations use external-camera RGB only. Queries use external and left-wrist RGB. B1 and frame-key regression supervise both query views. Foresight matching uses external-camera features. Frozen evaluation comprises 1,200 pairs.

Training and publishing

Four GPUs, batch 16 per GPU, effective batch 64, 20 epochs, learning rate 5e-5, warmup 0, demo dropout 0.3, B1 weight 0.5 and key weight 0.5. The inherited exact-length sampler retains 23,552 presentations per epoch, yielding 368 updates per epoch and 7,360 updates over 20 epochs.

Each completed epoch is exported to epoch-NNN/adapter.safetensors with an adapter.json receipt. The publisher verifies remote file size and SHA-256 before recording a successful upload. Training weights appear as epochs finish.

See adapter_config.json for the base revision, architecture and cache identities, and training_plan.json for the experiment settings. Inference uses the matching RoboKata mixed25k integration and verified demo caches. Semantic future targets are training-only.