# Training

| Setting | Value |
| --- | --- |
| Optimizer steps | 100 |
| Hardware | 2 × NVIDIA A100 |
| Distributed strategy | Full-parameter FSDP |
| Precision | bfloat16 |
| Effective global batch size | 16 |
| Optimizer | Fused AdamW |
| Learning rate | 1e-5 with 10-step warmup, then constant |
| Adam betas / weight decay | (0.9, 0.95) / 0.1 |
| Objective | Video latent loss + action loss |

Across the 100 optimizer steps, the mean video-latent loss was `0.161962` and
the mean action loss was `0.020129`.

The eight source action values are mapped to LingBot-VA action channels
`[0, 1, 2, 3, 4, 5, 6, 28]`; channel 28 carries the gripper command. The exact
normalization statistics and model settings are included in
`configs/va_a1_cfg.py`.