LingBot-VA — Galaxea A1 Fruit Placement EEF — Step 1000
LingBot-VA fine-tuned for fruit placement into a bowl or plate with a Galaxea A1 arm. The model predicts robot actions and video latents.
Data
| Field | Value |
|---|---|
| Episodes | 130 |
| Frames | 44,824 at 30 FPS |
| Tasks | 5 |
| Cameras | front 480×480 RGB; wrist 640×480 RGB |
| Action | Episode-relative EEF pose + continuous normalized gripper |
Training epochs
Approximately 123.08 equivalent fine-tuning epochs, calculated as
1,000 optimizer updates × 16 effective global batch size / 130 training segments.
Each selected episode contributes one latent training segment; the denominator
is the segment count, not the raw video-frame count. This estimate assumes
all selected segments have valid latent caches and excludes base-model
pretraining. See training_summary.json.
Evaluation
The following qualitative rollouts evaluate fruit placement into a bowl or a plate. The lemon-to-bowl task is out-of-distribution (OOD); all other shown fruit–receptacle combinations are in-distribution (ID). Videos are trimmed after task completion and shown at 6× speed.
These evaluation rollouts were run with
pengyue-polaron/galaxea-a1-runtime at commit 423c53a.
| Fruit | Bowl | Plate |
|---|---|---|
| Banana | ID | ID |
| Lemon | OOD | ID |
| Mango | ID | ID |
These GIFs are qualitative examples. See the full real-robot evaluation for success rates and failure analysis.
Files and configuration
transformer/ contains the step-1000 weights. The base tokenizer, text encoder,
and VAE are included. Use configs/va_a1_cfg.py for the
EEF action mapping and normalization.
Training details · Training summary
License
Apache License 2.0. See LICENSE.txt.