LingBot-VLA 2.0 — Stack-the-cubes ablations
Finetunes of LingBot-VLA 2.0 (6B, MoE action expert) on Unitree G1 Dex1 dual-arm cube stacking. All runs freeze the vision encoder; distillation teachers stay frozen, current+future depth/DINO query heads train. Muon, lr 1e-4, L1_fm, bounds_99_woclip, absolute actions.
| Subfolder | Data | Action space | Steps | Note |
|---|---|---|---|---|
joint100/ |
100ep | joint (14 arm + 2 grip) | 20000 (0.42 epoch) | vision frozen, 2xA100 |
eef/ |
150ep (v1+v2) | EEF pose (14 + 2 grip) | 20000 (0.29 epoch) | vision frozen, 2xA100 |
eef100/ |
100ep | EEF pose (14 + 2 grip) | 20000 (0.42 epoch) | vision frozen, 2xA100 |
eef100_10ep/ |
100ep | EEF pose (14 + 2 grip) | 59979 (10 full epochs) | vision frozen, GBS=16, 2xB200 |
Ablation axes: joint100 vs eef100 = joint vs EEF (same data); eef vs eef100
= mixed 150ep vs clean 100ep; eef100 vs eef100_10ep = training exposure
(0.42 epoch vs 10 full epochs of the same 100ep EEF data — the original runs never
completed even one full pass over their training set). Each */global_step_*/
holds deployable HF weights (model-0000x-of-00006.safetensors + tokenizer/config).
EEF models output end-effector poses — the client must IK them back to joints
(see g1-client main_eef.py).