Week 63 - Test offline, MoE (teleoperado) vs Moe (teleoperado + DAgger)
| Métrica | Dagger (Teleoperado + DAgger) | MC (Solo teleoperado) | Δ (MC − Dagger) | Δ% vs MC |
|---|---|---|---|---|
| RMSE Steer | 0.275 | 0.521 | 0.246 | 47.24% |
| MAE Steer | 0.163 | 0.477 | 0.314 | 65.77% |
| RMSE Throttle | 0.575 | 0.492 | −0.084 | −17.06% |
| MAE Throttle | 0.510 | 0.481 | −0.029 | −6.10% |
Nota (fórmulas)
Col. 4 — Δ (MC − Dagger): Δ = MC − Dagger
Col. 5 — Δ% vs MC: Δ% = ((MC − Dagger) / MC) × 100
Interpretación: Δ > 0 y Δ% > 0 ⇒ Dagger reduce el error; Δ < 0 y Δ% < 0 ⇒ Dagger empeora. Si MC = 0, Δ% no está definida.
En la evaluación offline (mezcla de expertos curva/recta), Dagger (teleoperado + DAgger) reduce de forma marcada los errores de dirección frente a MC (solo teleoperado), especialmente en tramos rectos, mientras que en aceleración empeora globalmente (con ligera mejora en curvas). En síntesis: DAgger fortalece el control lateral, pero el control longitudinal requiere ajuste (normalización, balance de datos o calibración de la salida).
| Segmento | Métrica | Dagger | MC | Δ (MC − Dagger) | Δ% vs MC |
|---|---|---|---|---|---|
| GLOBAL | RMSE Steer | 0.274848 | 0.520959 | 0.246111 | 47.241955% |
| GLOBAL | MAE Steer | 0.163238 | 0.476945 | 0.313708 | 65.774348% |
| GLOBAL | RMSE Throttle | 0.575458 | 0.491596 | −0.083862 | −17.059209% |
| GLOBAL | MAE Throttle | 0.509839 | 0.480524 | −0.029315 | −6.100561% |
| RECTA | RMSE Steer | 0.176157 | 0.560013 | 0.383856 | 68.544117% |
| RECTA | MAE Steer | 0.087130 | 0.538385 | 0.451256 | 83.816492% |
| RECTA | RMSE Throttle | 0.625331 | 0.483878 | −0.141453 | −29.233242% |
| RECTA | MAE Throttle | 0.540846 | 0.474337 | −0.066509 | −14.021432% |
| CURVA | RMSE Steer | 0.386433 | 0.448789 | 0.062356 | 13.894328% |
| CURVA | MAE Steer | 0.289410 | 0.375090 | 0.085680 | 22.842493% |
| CURVA | RMSE Throttle | 0.481530 | 0.504131 | 0.022600 | 4.482999% |
| CURVA | MAE Throttle | 0.458433 | 0.490780 | 0.032346 | 6.590778% |