casm_hard_gate_put_object_cabinet_parallel_50_20000_bs1

ABORTED research artifact — retained for reproducibility, not deleted.

This artifact is retained as historical evidence only. Its cached segment-start main-camera observations are not eligible for dynamic-main-view claims. See ABORTED.yaml for the machine-readable archival record.

Archival registry mapping:

experiment_id: E004

---

CASM hard-gate — put_object_cabinet

This repository contains the params-only Orbax checkpoint used for the unified RoboTwin 2.0 CASM hard-gate evaluation.

• Model: CASM hard-gate on Pi0.5
• Task: put_object_cabinet
• Training data: 50 parallel dual-arm episodes from Shiki42/parallelvla_putcab_official_clean50_native_path_retimed_paired_v2
• Optimizer steps: 20,000
• Global batch size: 1
• Train config: pi05_putcab_casm_hard_gate_lora
• Checkpoint directory SHA-256: 13796cb0be3f008a7ccd5a2481615da67b909d05d327b9b2c992d3844a87ed3b

This is the learned CASM hard-gate checkpoint. It is not either of the earlier CASM-lite checkpoints. It contains inference parameters and assets, not optimizer state, so it must not be treated as a resumable full-state training checkpoint.

See checkpoint_manifest.json for the exact relative file list, byte sizes, and SHA-256 digests.