| # 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`. | |