DA3-XVLA — RoboReal/RoboPRO Ablation Checkpoint Collection
Checkpoints for the DA3-XVLA (Depth-Anything-3 + X-VLA) ablation study on the
RoboReal/RoboPRO closed-loop manipulation benchmark. Each subfolder is one
(ablation, checkpoint) in standard HF format (XVLA.from_pretrained-compatible:
config.json + pytorch_model.bin + tokenizer).
Baseline axes (the original DA3-XVLA): DA3-BASE backbone, K=160 geometry tokens, last-layer latent, 504-upscale resolution, frozen DA3. Each ablation changes exactly ONE axis vs this baseline.
All runs: frozen DA3, trained posed + true-RGB (XVLA_POSED_DA3=1,
XVLA_RGB_INPUT=1), finetuned from X-VLA-Pt, fused AdamW, bf16.
Checkpoints in this collection
| folder | DA3 | K | layer | resolution | bs | iters | train loss | notes |
|---|---|---|---|---|---|---|---|---|
native-bs16-ckpt90000 |
BASE | 160 | last | native 224 | 16 | 90k / 125k | ~0.10 | resolution ablation (vs 504-upscale baseline) |
multi-native-ckpt125000 |
BASE | 160 | multi | native 224 | 16 | 125k / 125k | ~0.068 | feature-layer ablation (4-layer DPT fusion vs last-layer) |
large-native-bs32-ckpt100000 |
LARGE | 160 | last | native 224 | 32 | 100k | ~0.073 | DA3-LARGE + native ablation (4×H200) |
vggt-native-bs32-ckpt100000 |
VGGT-Omega-1B | 160 | last | native 224 | 32 | 100k | ~0.107 | VGGT backbone vs DA3 (4×H200) |
More ablations appended as they finish: multi-layer fusion, K=320, DA3-LARGE, VGGT-vs-DA3, ...