ML Scaling Laws in Autonomous Driving

Figure 2. Fit model capacity (number of parameters) and training steps scaling laws using training loss envelope.

Figure 3. Behavior eval improvements with scaling. Negative/positive numbers indicate relative improvements/regressions.

Figure 4. Fitted Model Performance Scaling Function L(N, D)

Figure 5. Use the mAP envelope to find the optimal models at different budgets. “Behavior” vertical lines are for the comparison of complexity difference to Perception.

Figure 6. The evaluation result of the perception model at different scales.