roadmap23-leader-corrective-plus300

Roadmap23 corrective +300

Research checkpoint: 300 additional supervised updates from parent revision e7d65c5babf0212f8a2c95eec585868e4c4f0059, not training from scratch. Architecture: LingBot-VLA v2. Vision encoder frozen; language/action components trained. The mixture combines standard LIBERO replay, prior swap examples, and corrective demonstrations for bowl, milk, and salad-dressing tasks. Original 600-update schedule was paused at 300. No reinforcement learning was used.

Evidence and limits (2026-09-06)

Targeted matched bowl/milk tests: parent18/48 versus this checkpoint37/48; new-layout subset11/30 versus22/30. This was an interrupted larger test, with unresolved exact camera-equivalence/numerical-noise caveats, not proof of broad benefit. Separate 112-attempt LIBERO-PRO panel: this checkpoint73/112, folded six-component estimate0.601190; Apex comparator84/112,0.714286. Fresh actual-parent112 control is pending at publication. One initial-state trial per selected task/variant; not an official subnet result or a full benchmark. No championship claim.

Inference weights and configs only. Dataset, private scene banks, optimizer, and distributed training state are NOT included. See MODEL_MANIFEST.json for source hashes and parent provenance. Uses upstream LingBot-VLA v2 inference with official LIBERO normalization (training/norm_stats.json); our evaluation predicted50 actions and executed5 before replanning, with10 denoising steps. Tokenizer/processor assets come unchanged from the public parent. Parent did not publish a license/model card at the pinned revision; this card does not grant additional rights over upstream components. Check upstream terms.