g1_fdmV2_500

g1_fdmV2_500 — LingBot-VA G1 post-trained transformer (step 500, FDM v2)

Fine-tuned transformer for LingBot-VA on Unitree G1 (Dex1) data (bobchenyx/g1_dex1_pick_pink_lerobot033, task: "pick up the pink object and place it on the blue cross mark").

• Base: robbyant/lingbot-va-base
• Post-training: 50 demos, lr 1e-5, FDM v2 recipe — the author's mutually-exclusive per-step regime: each optimizer micro-step flips a rank-synchronized coin (fdm_prob=0.5) and computes either the FDM video-only loss (Eq. 13, lambda_fdm=1.0) or the standard IDM loss (L_dyn + L_inv) — one forward, one backward. This is not the earlier g1_fdm_500 two-forward recipe and is not directly comparable to it.
• Checkpoint at optimizer step 500 (of 2000).
• This repo contains only the transformer/ — vae/, text_encoder/, and tokenizer/ are unchanged from robbyant/lingbot-va-base.

Assemble an eval-ready checkpoint

hf download robbyant/lingbot-va-base   --local-dir lingbot-va-base
hf download armanakbari4/g1_fdmV2_500   --local-dir g1_fdmV2_500_dl

mkdir -p g1_fdmV2_500
ln -sf $(realpath g1_fdmV2_500_dl/transformer)  g1_fdmV2_500/transformer
ln -sf $(realpath lingbot-va-base/vae)           g1_fdmV2_500/vae
ln -sf $(realpath lingbot-va-base/text_encoder)  g1_fdmV2_500/text_encoder
ln -sf $(realpath lingbot-va-base/tokenizer)     g1_fdmV2_500/tokenizer

Serve with CONFIG_NAME=g1 MODEL_PATH=g1_fdmV2_500.

transformer/config.json has attn_mode: torch (inference-ready; flex attention is training-only).