ml-developers/pi05-eUiSKJr4zg3Q

LingBot-VLA 2.0 — expert-only fine-tune (SN80 competition 2)

Canonical parent (per the OpenRoboto Model Originality policy): openroboto-ai/champion@d54728aa7347c2996fe1545cc75d2055909458b0 (official champion snapshot, mirror of ApexUltron/pi05-KsYerY3Q8siH@4ac5fea918107a26ba525e3d1b6bdbe615d4ee72).

Training method: continued training with train_expert_only=true — the entire Qwen3-VL backbone (vision tower, all 36 LLM layers, embeddings) frozen and byte-identical to the parent; gradients applied only to the MoE action expert, action/state/time projections, and align heads. 2,000 steps, muon, lr 3e-6 -> 3e-7 cosine, global batch 64, fp32.

Data: LIBERO demonstrations (lerobot/libero) plus successes-only policy rollouts collected on randomized layouts of the four TRAINING suites (no evaluation-suite BDDL files or init states were used).