Training-Data Axes in Imitation Learning Cascade into Hybrid Reinforcement-Learning Fine-Tuning: A Leave-One-Out Bridge Study Under Matched and Cascade Evaluation

Abstract

Hybrid imitation-learning-to-reinforcement-learning (IL→RL) driving stacks are typically evaluated against a single fixed IL prior, leaving open whether IL training-data quality determines downstream RL outcomes and whether hybrid actuator decoupling (IL steers, RL controls only speed) isolates the speed controller from IL deficiencies. We evaluate five trajectory-IL checkpoints from a fixed conditional imitation learning (CIL) architecture (one matched baseline corpus plus four equal-budget leave-one-axis-out (LOO) ablations: drop_map, drop_weather, drop_traffic, drop_perturbation) under a 2 × 2 grid: matched proximal-policy optimisation (PPO) retraining versus frozen-PPO cascade evaluation, crossed with 0NPC versus 20NPC traffic (non-player characters; n = 3 PPO seeds per arm, wired v2-progressive reward). Matched co-training yields 442.6-unit cross-seed mean-reward spread at 0NPC (drop_map high, drop_perturbation low): the IL training corpus alone changes how well the same PPO recipe can learn speed control. Kendall τ against prior LOO pure-tier distance-importance is mild at matched 0NPC (τ=+0.333), stronger under matched 20NPC (τ=+0.667), mildly negative under cascade at 0NPC (τ=−0.333), and null under cascade at 20NPC (τ=0.00), so absolute gaps cascade more reliably than axis order. Under 20NPC traffic, cascade buffers the weak arms: the largest cascade-matched gap is on drop_perturbation (Δ=+428.3), with drop_weather still large (Δ=+205.5), meaning a baseline-trained speed head is more tolerant of a deficient IL prior at inference than a speed head co-trained against that prior. Frozen-head decoupling is therefore more stable at inference than per-axis co-training once multi-agent stress activates failure modes absent from zero-traffic training.