Ablation V6 of the Stage3 Qwen2.5-VL framework: two-phase training in which
Phase 1 pretrains BaseVLA on a 20-domain mix (16 stage1 robosuite/mimicgen +
4 stage2 maniskill/OXE domains, ~8.4M frames, state proj OFF), and Phase 2
finetunes the resulting checkpoint on the LIBERO island A+B paper-200
protocol with plain V1-style BC (state proj re-enabled, no aug, no InfoNCE,
no SIGReg).
Evaluated on the… See the full description on the dataset page:
https://huggingface.co/datasets/disentangled-vla/libero-island-ablation-v6-phase2-stage2-pretrain.
