README: task2 demo+recovery checkpoints
Browse files
README.md
CHANGED
|
@@ -41,6 +41,35 @@ the checkpoints nearest the epoch boundaries (one epoch = 15,089 steps).
|
|
| 41 |
`action_loss` is training-set loss only (no validation split, no rollouts yet).
|
| 42 |
<!-- task1-end -->
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
## `task5_human_only_21k`
|
| 45 |
|
| 46 |
Single-task fine-tune of the 50-task meta checkpoint on **task 5 = `setting_mousetraps`**,
|
|
|
|
| 41 |
`action_loss` is training-set loss only (no validation split, no rollouts yet).
|
| 42 |
<!-- task1-end -->
|
| 43 |
|
| 44 |
+
<!-- task2-begin -->
|
| 45 |
+
## `task2/` — putting_away_Halloween_decorations, human demos + recovery data (round 0)
|
| 46 |
+
|
| 47 |
+
Single-task fine-tune of the 2025 winner's **`checkpoint_1`** (the official checkpoint covering task 2) on
|
| 48 |
+
**task 2 = `putting_away_Halloween_decorations`**, using the 200 human demos plus 543 scripted recovery episodes
|
| 49 |
+
from `fastwalker1118/B1k_Recovery` (`task2/Demo+Recovery_video_r0a`). The recoveries start from failure frames
|
| 50 |
+
that a video-only inspection found in `checkpoint_1` rollouts, and from edited variants of those frames
|
| 51 |
+
(no public-test instances). 2 epochs; one epoch =
|
| 52 |
+
31,262 steps.
|
| 53 |
+
|
| 54 |
+
| checkpoint | step | epochs | `action_loss` near that step (train set) |
|
| 55 |
+
|---|---|---|---|
|
| 56 |
+
| [`task2/task2_demo_recovery_video_r0a_epoch1_30000`](https://huggingface.co/fastwalker1118/b1kcheckpoints/tree/main/task2/task2_demo_recovery_video_r0a_epoch1_30000) | 30,000 | 0.96 | 0.0476 |
|
| 57 |
+
|
| 58 |
+
| | |
|
| 59 |
+
|---|---|
|
| 60 |
+
| init | `IliaLarchenko/behavior_submission` `checkpoint_1` (params) |
|
| 61 |
+
| config | `pi_behavior_b1k_fast` (FSDP 4, EMA, FAST aux loss), `B1K_TASK_SPACE=100` |
|
| 62 |
+
| data | 743 episodes (200 demos + 543 recoveries), 4,001,658 samples per epoch; all videos re-encoded to 224×224 with lanczos + the full upstream x265 option string (crf 28, keyint 8) |
|
| 63 |
+
| batch | 128 on 4× H200 |
|
| 64 |
+
| LR | 2.5e-7 → 2.5e-6 at step 500 → cosine to 2.5e-7 at step 62,524 (2 epochs) |
|
| 65 |
+
| norm stats | 2025 `checkpoint_1` stats + robot-frame `base_qvel` fix (state dims 0–2); `trunk_3` stays frozen (the task-2 recoveries never rotate the waist) |
|
| 66 |
+
| targets | normalized flow-matching action targets clipped to ±10σ |
|
| 67 |
+
|
| 68 |
+
> **Eval: robot-frame base velocity in the state, and the norm stats shipped in each checkpoint's `assets/`.**
|
| 69 |
+
|
| 70 |
+
`action_loss` is training-set loss only (no validation split).
|
| 71 |
+
<!-- task2-end -->
|
| 72 |
+
|
| 73 |
## `task5_human_only_21k`
|
| 74 |
|
| 75 |
Single-task fine-tune of the 50-task meta checkpoint on **task 5 = `setting_mousetraps`**,
|