Upload franka_multitask_v1/0020000/README.md with huggingface_hub

franka_multitask_v1/0020000/README.md

@@ -0,0 +1,35 @@
+# Franka real-robot multi-task finetune — onground/ckpt
+
+Resumed from `SeonghuJeon/oxe-mimicgen-all-h1to8-teacherdepth-300k`.
+Architecture: DA3-Giant + Shallow12-AR predictor + ActionHeadV2 
+(`mlp_resnet`, `chunk_size=8`, `chunk_position_encoding=none`, 
+`pool_mode=mean`); T5-base language, concat conditioning; 2 RGB cameras.
+
+## What's in this upload
+
+- `franka_multitask_v1/0020000/0020000.pt` — finetune checkpoint
+- `franka_multitask_v1/0020000/config_resolved.yaml` — YAML the trainer logged
+- `franka_multitask_v1/0020000/ds_config.json` — ZeRO-2 / mb=64 DeepSpeed config
+- `franka_multitask_v1/0020000/franka_realrobot.json` — shared action stats
+- `franka_multitask_v1/0020000/franka_realrobot_proprio.json` — shared proprio stats
+- `franka_multitask_v1/0020000/metadata.json` — eval history + run metadata
+
+## Action / proprio / image convention
+
+See `docs/realrobot-franka-deploy-handoff.md` in
+[ONground-Korea/3DA](https://github.com/ONground-Korea/3DA) — it is the
+single-file deploy entry point (literal task prompts, inverse mapping
+pseudo-code, gripper polarity invert, gotchas). Companion training spec:
+`docs/realrobot-finetune-franka.md`.
+
+## Eval history (held-out 41 episodes)
+
+| step | l1_norm | l1_raw (m) | r2_norm | rel@5% | rel@10% |
+|---|---|---|---|---|---|
+| 12000 | 0.3096 | 0.0020 | 0.193 | 0.386 | 0.550 |
+| 14000 | 0.3176 | 0.0022 | 0.185 | 0.371 | 0.542 |
+| 16000 | 0.3158 | 0.0023 | 0.186 | 0.374 | 0.552 |
+| 18000 | 0.3129 | 0.0020 | 0.186 | 0.374 | 0.552 |
+| 20000 | 0.3118 | 0.0019 | 0.180 | 0.386 | 0.557 |
+
+**Best by `l1_raw`: step 20000** (`0.0019 m`, `r2_norm=0.180`).