README.md

---
license: apache-2.0
tags:
- robotics
- lingbot-va
- unitree-g1
- world-model
---

# UnitreeG1_putawaytoolsV2_minmax_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`:
*"Put the battery on the shelf labeled 'battery' and put the screwdriver
on the shelf labeled 'Philips'."*

**Same data, same recipe as the `rndchnk` series — only difference: action
normalization is MIN/MAX (not q01/q99 quantile).** See "Why min/max" below.

- Base: `robbyant/lingbot-va-base`
- Post-training: 70 demos (43,851 frames), lr 1e-5, **FDM v2 recipe** —
  mutually-exclusive per-microstep regime (`fdm_prob=0.5`, `lambda_fdm=1.0`).
  Per-step randomized chunk_size ∈ {1..4} and window_size ∈ {4..64}.
- 4 GPUs × `grad_accum=4` = effective batch 16, optimizer **step 500** of a
  5000-step schedule (very early — uploaded specifically for A/B testing
  against the matching `rndchnk_500step` quantile-normalized ckpt).
- **Action normalization: dataset min/max** — every training target bounded
  strictly to [-1, +1]. (Codebase variable names are still `q01`/`q99`
  because that's all the loader supports; the values stored there are
  min/max — drop-in replacement.)
- This repo contains **only `transformer/`** — `vae/`, `text_encoder/`,
  `tokenizer/` are unchanged from `robbyant/lingbot-va-base`.

## Why min/max (the v21 quantile series underperformed)

The earlier v21 5k training under quantile normalization had its right-arm
joints overflow: R-wrist-roll absmax was **4.11**, R-shoulder-roll 3.55,
R-wrist-yaw 3.55. The model's bounded prediction range
(`[~-1.5, ~+1.5]`) cannot match those targets → during deployment the model
under-predicts the precise reach-extension moments → arm under-extends →
misses the shelves. Min/max normalization bounds every target to ±1
(verified absmax = 1.0000 over all 43,851 training rows), eliminating
out-of-range targets and restoring deployment quality.

Loss curves under min/max are higher than the quantile run by design — the
quantile run's suspiciously-low video loss (0.0072 at step 5000) was the
signature of fitting a compressed bulk distribution while ignoring
unreachable extremes. The min/max run's loss (0.0347 video at step 5000)
reflects the model now learning a wider, fully-reachable target range.

## Assemble an eval-ready checkpoint

```bash
hf download robbyant/lingbot-va-base                              --local-dir lingbot-va-base
hf download EmbodyX/UnitreeG1_putawaytoolsV2_minmax_500step        --local-dir g1_pat_v2_mm_500_dl

mkdir -p g1_pat_v2_mm_500
ln -sf $(realpath g1_pat_v2_mm_500_dl/transformer)  g1_pat_v2_mm_500/transformer
ln -sf $(realpath lingbot-va-base/vae)              g1_pat_v2_mm_500/vae
ln -sf $(realpath lingbot-va-base/text_encoder)     g1_pat_v2_mm_500/text_encoder
ln -sf $(realpath lingbot-va-base/tokenizer)        g1_pat_v2_mm_500/tokenizer
```

Serve with `CONFIG_NAME=g1_putawaytools_v21 MODEL_PATH=g1_pat_v2_mm_500`.
`transformer/config.json` has `attn_mode: torch` (inference-ready).

**IMPORTANT — config must match training**: the inference config's
`norm_stat` must contain the same MIN/MAX values used during training
(NOT the original quantile values). The `va_g1_putawaytools_v21_cfg.py`
in the lingbot-va repo has been updated in lockstep — using the original
quantile config at inference with this checkpoint would denormalize wrong.