CTR PutCab π0.5 E609, step 10000
This public repository contains the inference checkpoint from CTR Experiment E609-R001, trained on the PutCab simulation task using the formally usable no-IdleMask revision Shiki42/PutCab-CTR-Train50-V4@575634eaaaed193458001b41050d0f9f55ecf545.
- Policy: OpenPI π0.5, single action expert, LoRA fine-tuning.
- Source training run: E609-R001, 10,000 optimizer updates.
- Source checkpoint verification: fresh-process reload PASS; local receipt SHA-256
73dd13b537b764dc1fec038f0aaab5f1a8ce69dfb3b1e45a98688f6d7da92978. - OpenPI source commit:
e9ba7b7732a3e66e4bd87d6d3429f7cef6352ead. - Inference payload: parameters, resolved assets and normalization, and checkpoint metadata. Training optimizer and data-loader state are excluded.
- Published for reproducible CTR simulation evaluation, including E802 on the frozen E549 C00 suite. E610 reported 55/100 success on its separate default suite; this does not predict E802's result.
The published SHA256SUMS and immutable repository revision identify the exact inference files. CTR Experiment and Run receipts remain the authority for training provenance and result lifecycle. This checkpoint is for simulation research, not real-robot use.
Archival clarification: E609 is WITHOUT IdleMask
E609-R001 step10000 is the PutCab (simulation) PI0.5 CTR without IdleMask ablation. The registered training_plan.idle_mask is false; observation.arm_active_mask is not consumed. Ordinary temporal padding masking remains enabled. This annotation changes no weights, original normalization, training configuration, repository name or historical evaluation result. E607 is a different Scan Object no-mask experiment and its original record remains unchanged. Evaluation E610 depends on E609; use each evaluation Run receipt to verify the exact protocol and checkpoint identity.