README: task1 demo+recovery checkpoints
Browse files
README.md
CHANGED
|
@@ -11,6 +11,34 @@ base_model: IliaLarchenko/behavior_50t_checkpoint
|
|
| 11 |
Params-only checkpoints for offline evaluation. `train_state` is **not** included, so these
|
| 12 |
cannot be resumed from — inference/eval only.
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
## `task5_human_only_21k`
|
| 15 |
|
| 16 |
Single-task fine-tune of the 50-task meta checkpoint on **task 5 = `setting_mousetraps`**,
|
|
|
|
| 11 |
Params-only checkpoints for offline evaluation. `train_state` is **not** included, so these
|
| 12 |
cannot be resumed from — inference/eval only.
|
| 13 |
|
| 14 |
+
<!-- task1-begin -->
|
| 15 |
+
## `task1/` — picking_up_trash, human demos + recovery data
|
| 16 |
+
|
| 17 |
+
Single-task fine-tune of the 2025 winner's **`checkpoint_2`** (the official checkpoint covering
|
| 18 |
+
task 1) on **task 1 = `picking_up_trash`**, using the 200 human demos plus 788 scripted recovery
|
| 19 |
+
episodes from `fastwalker1118/B1k_Recovery` (`task1/Demo+Recovery`). Planned 5 epochs; these are
|
| 20 |
+
the checkpoints nearest the epoch boundaries (one epoch = 15,089 steps).
|
| 21 |
+
|
| 22 |
+
| checkpoint | step | epochs | `action_loss` at that step (train set) |
|
| 23 |
+
|---|---|---|---|
|
| 24 |
+
| [`task1/task1_demo_recovery_epoch1_15000`](https://huggingface.co/fastwalker1118/b1kcheckpoints/tree/main/task1/task1_demo_recovery_epoch1_15000) | 15,000 | 0.99 | 0.0635 |
|
| 25 |
+
|
| 26 |
+
| | |
|
| 27 |
+
|---|---|
|
| 28 |
+
| init | `IliaLarchenko/behavior_submission` `checkpoint_2` (params) |
|
| 29 |
+
| config | `pi_behavior_b1k_fast` (FSDP 4, EMA, FAST aux loss), `B1K_TASK_SPACE=100` |
|
| 30 |
+
| data | 988 episodes (200 demos + 788 recoveries), 1,931,447 samples per epoch; source videos (head 720², wrists 480²) re-encoded to 224×224 with lanczos + the upstream x265 option string (crf 28, keyint 8), frame alignment verified |
|
| 31 |
+
| batch | 128 on 4× H200 |
|
| 32 |
+
| LR | 2.5e-7 → 2.5e-6 at step 500 → cosine to 2.5e-7 at step 75,445 (5 epochs) |
|
| 33 |
+
| norm stats | 2025 stats + robot-frame `base_qvel` fix (state dims 0–2) **+ `trunk_3` un-frozen**: action dim 6 / state trunk_3 mean 0, std 0.1315, measured on the recoveries (the 2025 stats have std ≈ 0 because the demos never rotate the waist; the recoveries do) |
|
| 34 |
+
| targets | normalized flow-matching action targets clipped to ±10σ (scripted-planner recoveries have rare far-out chunk starts) |
|
| 35 |
+
|
| 36 |
+
> **Eval: load the norm stats shipped in each checkpoint's `assets/`**, not the 2025
|
| 37 |
+
> `checkpoint_2` stats. With the 2025 stats the waist (`trunk_3`) input and output are mis-scaled.
|
| 38 |
+
|
| 39 |
+
`action_loss` is training-set loss only (no validation split, no rollouts yet).
|
| 40 |
+
<!-- task1-end -->
|
| 41 |
+
|
| 42 |
## `task5_human_only_21k`
|
| 43 |
|
| 44 |
Single-task fine-tune of the 50-task meta checkpoint on **task 5 = `setting_mousetraps`**,
|