UnitreeG1_putawaytoolsV2_rndchnk_500step — LingBot-VA G1 post-trained transformer

Fine-tuned transformer for LingBot-VA on Unitree G1 (Dex1), task XiaoweiLinXL/pi05-unitree-g1-put-away-tools-v2.1 (cleaner recollection of the original put_away_tools task): "Put the battery on the shelf labeled 'battery' and put the screwdriver on the shelf labeled 'Philips'."

• Base: robbyant/lingbot-va-base
• Post-training: 70 demos (43,851 frames) — cleaner than v1 (48 demos). Single-task, lr 1e-5, FDM v2 recipe — mutually-exclusive per-microstep regime (rank-synced coin fdm_prob=0.5: FDM video-only L_fdm Eq.13 lambda_fdm=1.0 OR standard IDM L_dyn+L_inv; one forward, one backward). Per-step randomized chunk_size ∈ {1,2,3,4} and window_size ∈ {4..64} (the "rndchnk" in the repo name).
• 4 GPUs × grad_accum=4 = effective batch 16, optimizer step 500 of a 5000-step schedule (very early — uploaded specifically to A/B test the overfitting hypothesis: later checkpoints converged to suspiciously low loss [vid=0.0072 at step 5000, vs cup_broccoli baseline 0.052]; this 500 ckpt is still at vid=0.130, comparable to the working v1 step-500).
• Action range under quantile normalization: ±1.26 (vs v1's ±2.0 — v21 has a much tighter quantile distribution).
• This repo contains only transformer/ — vae/, text_encoder/, tokenizer/ are unchanged from robbyant/lingbot-va-base.

A/B testing context

The v21 5000-step checkpoint (EmbodyX/UnitreeG1_putawaytoolsV2_rndchnk_5000step) underperforms the matching-recipe v1 checkpoint (EmbodyX/g1_putawaytools_rndchnk_4000step) at deployment, despite v21 having more demos + cleaner data + lower training loss. This 500-step checkpoint exists to test whether late-stage overfitting on v21's narrower action distribution is the cause.

Assemble an eval-ready checkpoint

hf download robbyant/lingbot-va-base                              --local-dir lingbot-va-base
hf download EmbodyX/UnitreeG1_putawaytoolsV2_rndchnk_500step       --local-dir g1_pat_v2_500_dl

mkdir -p g1_pat_v2_500
ln -sf $(realpath g1_pat_v2_500_dl/transformer)  g1_pat_v2_500/transformer
ln -sf $(realpath lingbot-va-base/vae)           g1_pat_v2_500/vae
ln -sf $(realpath lingbot-va-base/text_encoder)  g1_pat_v2_500/text_encoder
ln -sf $(realpath lingbot-va-base/tokenizer)     g1_pat_v2_500/tokenizer

Serve with CONFIG_NAME=g1_putawaytools_v21 MODEL_PATH=g1_pat_v2_500. transformer/config.json has attn_mode: torch (inference-ready).