LingBot-VA — Galaxea A1 Plug Insertion EEF — Step 500
LingBot-VA fine-tuned for picking up a charger and inserting it into the leftmost socket of a power strip with a Galaxea A1 arm. The model predicts robot actions and video latents.
Data
| Field | Value |
|---|---|
| Available demonstrations | 31 episodes / 24,638 frames at 30 FPS |
| Training split | 26 episodes / 20,109 frames |
| Reserved split | Episodes 0, 7, 15, 23, 30 / 4,529 frames |
| Cameras | front 480×480 RGB; wrist 640×480 RGB |
| Action | Episode-relative EEF position + quaternion + normalized gripper |
Training epochs
Approximately 307.69 equivalent fine-tuning epochs, calculated as
500 optimizer updates × 16 effective global batch size / 26 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.
Only the 26 training episodes are counted; the five reserved episodes are excluded. Random crops of at most 64 latent frames mean this is a segment-sampling epoch count, not a count of complete passes over every raw frame.
Evaluation
Real-robot evaluation on August 8, 2026: 1/12 successful insertions (8.3%). Five additional runs stopped during the initial grasp and were excluded. See the rollouts and failure analysis.
Files and configuration
transformer/ contains the step-500 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
The saved training code is in training/; this is an inference checkpoint without optimizer state.
License
Apache License 2.0. See LICENSE.txt.