BaseVLA — LIBERO 4-suite (Qwen2.5-VL-3B) with domain augmentation
Plain behaviour cloning baseline: no V-JEPA, no latent alignment, no
decorrelation. Just L1(action) on the LIBERO 4-suite mix, with domain
augmentation switched on. Trained for 100,000 steps.
This is the control condition for a study on viewpoint↔task spurious
correlation — the "what does augmentation alone buy you?" cell.
Data
The four standard LIBERO suites, sampled with equal weight per suite
(weight / n_frames, so suite size does not bias the mix):
suite
episodes
frames
libero-spatial
432
52,970
libero-object
454
66,984
libero-goal
428
52,042
libero-10
379
101,469
total
1,693
273,465
Front camera (observation.images.image) only; the wrist camera is dropped.
10 fps, 8-D state, 7-D action.
Architecture
Qwen2.5-VL-3B + LoRA ─ forward_full_hidden ─┬─ Head A → z_a ─┐
└─ Head B → z_b ─┴─ concat → ResNetActionHead → 10×7
L = L1(action) ← that is the whole objective
Augmentation detail — applied to the context frame only:
lighting gain 0.3–2.0, sensor noise ISO 1–4, random crop, warping
(corner shift 0.08–0.20), each drawn independently per sample.
Training curve
step
loss
10,000
0.0823
20,000
0.0731
30,000
0.0655
50,000
0.0537
70,000
0.0492
100,000
0.0479
17 h 28 min on one A100 80GB.
Not evaluated on a robot or in simulation. The number above is a
training-set L1 on normalised actions. No LIBERO rollout success rate is
reported here, and rollout is what actually matters for this benchmark.
Contents
checkpoint.pt holds policy (full VLM incl. LoRA), latent_head,
free_latent_head, action_head, optimizer, agg_stats (the normalisation
statistics used), and args (the exact CLI configuration).
Load with:
python
1import torch
2ck = torch.load("checkpoint.pt", map_location="cpu", weights_only=False)3ck["step"]# 1000004ck["args"]# full training config5ck["agg_stats"]# action/state normalisation used at train time
