← Back to main
Exp14 Step 3: Full Dataset + WINDOW=8 + 32×32
Step 2 대비: 데이터 45→150 ep, WINDOW 3→8, image 16²→32² grayscale.
입력 dim: 268 → 1056 floats. MLP: 512→256→128→8.
Step 2 (baseline)
75.9%
Step 2 reference (45 ep, W=3)
Step 3 PM
77.0%
(403/525)
데이터
2626
windows (150 ep)
PM per Path Type (test split)
| Path Type | Correct/Total | PM |
|---|
| center_straight | 47/56 | 83.9% |
| center_left | 33/54 | 61.1% |
| center_right | 25/54 | 46.3% |
| left_straight | 68/72 | 94.4% |
| left_left | 39/55 | 70.9% |
| left_right | 42/57 | 73.7% |
| right_straight | 71/72 | 98.6% |
| right_left | 40/57 | 70.2% |
| right_right | 39/48 | 81.2% |
End-to-End Gate Check (Exp17 / Exp18)
| Model | PM | Closed-loop | FPE | Interpretation |
|---|
| Exp14 Step2 | 75.9% | 66.7% | 0.55m | Strongest practical baseline |
| Exp17 | 76.95% | 11.1% | 1.04m | PM only improved, rollout failed |
| Exp18 | 27.62% | 11.1% | 1.04m | Text-fusion gate failed |
설정
입력: WINDOW=8 × (cx, cy, area, has_bbox) + 32×32 grayscale image
출력: 8-class discrete action | epochs=300 | AdamW lr=2e-3
Exp18 실패 패턴 요약
best checkpoint는 val_loss 1.325였지만, 실제 PM은 27.62%, closed-loop는 11.1%에 그쳤습니다.
PM confusion 기준으로는 FORWARD 95개가 전부 FWD+R로 오분류됐고, closed-loop에서는 center_right만 성공했습니다.
즉 text embedding fusion은 현재 end-to-end instability를 해결하지 못했고, strongest baseline은 여전히 Exp14 Step2입니다.