maniskill-myws-pld-libero-checkpoints

PLD LIBERO checkpoint backup

This repository is the durable backup for the PLD LIBERO reproduction in State-Estimation/maniskill_myws.

The backup contains the exact source-only aligned base and residual specialist used by the reported evaluation, plus the original pretrained pi0 weights, offline base-policy rollouts, configs, experiment design, results, and a SHA-256 manifest.

Restore map

• model.safetensors, config.json, pretrained_provenance.json: official pretrained pi0 (/workspace/checkpoints/pi0_base_pytorch_v2). The installed HF CLI placed this directory's contents at repository root.
• weights/aligned_base_model.safetensors: selected 3000-update source-only LIBERO SFT base policy.
• weights/aligned_base_metadata.pt: aligned checkpoint metadata.
• weights/residual_step_50000.pt and .json: final 50,000-step visual residual specialist.
• data/offline_base_rollouts.npz: successful frozen-base rollouts used for Cal-QL initialization.
• configs/anchor_bowl_otf.json, docs/EXPERIMENT_DESIGN.md, docs/RESULTS.md: configuration and provenance.
• manifests/weights_manifest.json: local source paths, byte sizes, SHA-256 hashes, and remote paths.

The original machine-local outputs include many intermediate/failed optimizer snapshots; they are not required to restore the published experiment. This HF repository is the durable copy of every weight required to reproduce the selected run.