Drop the README commands whose configs no longer exist

[Vertax42]

Follow-up to #12. I checked every --config-name / --policy.config= / train.py <name> in the README against all_configs() — the real resolver, so the five Python-generated RoboArena baselines count as valid, not just the YAML files under configs/_examples/.

Seven names did not resolve. Those commands fail at startup for anyone who copies them.

Deleted — the config behind them is gone

What	Config
a training run inside the assemble-box block	..._assemble_box_with_phone_stand_lora_0403
a training run inside the shoe-insole block	..._shoe_insole_retrieval_and_packing_0607_h100
whole subsection "BiFlexiv — earbuds case sequential insertion task" (training)	..._earbuds_case_sequential_insertion_teleop_rtc_0513_h100
whole subsection "BiFlexiv — earbuds case assembly with lid operation inference"	same config
whole subsection "BiARX5 — training-time RTC inference"	pi05_base_arx5_lora_training_time_rtc
a serve command inside the assemble-box inference block	..._assemble_box_with_phone_stand_lora_0410_merged_fixed

Where a block had both working and broken commands, only the broken command went — the assemble-box training block keeps its 0422 and 0430 runs, the shoe-insole block keeps its two 0515 runs, and the assemble-box inference block keeps three serve commands.

Corrected instead of deleted — these two were not actually dead

..._newbalance_shoe_insole_... is a misspelling. The config ships as ..._newbalacne_... — the typo is in the filename, and the typo is the one that works. Six call sites, including step ① of the three-machine dewu_video_switch runbook, which is otherwise live. Deleting a working procedure over a one-letter transposition would have been the wrong trade.

The getting-started walkthrough used pi05_base_arx5_lora as its stand-in name across compute-norm-stats / train / serve. Deleting a tutorial's commands leaves the tutorial empty, so it now names pi05_base_xtac_umi_pick_up_cube_0807_h200 — ships, and is the config under active work.

Result

Every config name in README.md now resolves. Net −52 / +7.

Left alone, deliberately

examples/droid/README_train.md still names pi05_droid_finetune and pi05_full_droid_finetune, which this fork does not ship. Deleting those would empty that document's entire Run section, and whether the fork keeps DROID fine-tuning docs at all is a separate call — pi0_droid, pi05_droid and pi0_fast_droid all still ship, so DROID is not simply gone. Happy to take it either way in a follow-up.

Verification

• re-ran the audit against all_configs(): no stale names left in README.md
• code fences balanced (62 markers), no orphaned subsection headings
• ruff check / ruff format --check clean, pytest src/openpi/training/ 54 passed

🤖 Generated with Claude Code