README.md

Decoupled reading: model weights (backup 2026-09-30)

• reader_weights/: decoupled reader (12 obs_encoder.ptr_* tensors) + normalizer + cfg for every sequence-training run (INDEX.json); the BPP policy is frozen, so released BPP checkpoint + these = full model.
• full/: full checkpoints, learned-count reader (LEVT L1/L2) and joint-training ablations.
• full_policy_trained/: runs that also trained the policy (sync/content/nocopy/lnc/dag/factorslot).
• pi05_lora/: pi0.5 LoRA on LIBERO-Gen goal chains (openpi), steps 5k-50k; evaluated: 10000, 20000, 49999.
• baselines/: RoboSSM (20,596 steps) and UniSkill (epoch 500) retrained on LIBERO-Gen Chain.
• earlier_lines/: models from earlier project lines (subgoal/goal image, demo generators, keyframe, pose, trajectory, anchor, phase, boundary reader).