HybridLoss : an adaptive planning-oriented loss function for end-to-end autonomous vehicle

Abstract

Autonomous driving often suffer from a decoupled feedback loop between prediction and planning. While prediction losses focus on the accuracy of surrounding agents, planning losses typically imitate recorded Autonomous Vehicles’ (AVs) trajectories that may contain suboptimal or aggressive behaviors, leading to unstable interactions in mixed traffic. This paper presents HybridLoss, an adaptive planning-oriented objective that unifies prediction and planning through planner-in-the-loop supervision and interaction-aware consistency. HybridLoss integrates an adaptive motion-planning module which replaces ground-truth targets with optimized reference trajectories, and a multi-term loss combining prediction, adaptive planning, safety potential, and social force objectives. Evaluations on the INTERACTION dataset indicate that HybridLoss significantly outperforms strong baselines. Beyond standard metric improvements—reducing ADE/FDE from 1.36/1.64 m to 1.11/1.36 m and collision rates from 0.19% to 0.11%—extensive stress-testing reveals superior system maturity. First, HybridLoss exhibits the highest robustness under input perturbations, maintaining the lowest planning deviation and endpoint standard deviation. Second, it demonstrates strong generalization, maintaining stable success rates (87.7%) in unseen scenarios with high computational efficiency (64.3 Hz). Third, multi-objective analysis confirms that HybridLoss achieves the optimal Pareto trade-off between efficiency, safety, and comfort, avoiding the speed-safety collapse seen in baseline methods. Finally, social force evaluations highlight that HybridLoss fosters implicit cooperation, achieving higher yield rates and reduced conflict indices while maintaining safe interaction buffers. These results validate HybridLoss as a robust, socially compliant, and adaptive solution for end-to-end driving. Note to Practitioners— HybridLoss addresses the challenge of balancing individual safety and global efficiency in AV planning within mixed-traffic environments. HybridLoss dynamically integrates adaptive motion-planning with a multi-component loss function to create closed-loop feedback between prediction accuracy and planning quality. This adaptability allows AVs to rapidly optimize trajectories for diverse scenarios, such as prioritizing collision avoidance for merging vehicles while maintaining platoon efficiency on highways. Practitioners can deploy HybridLoss in urban or highway settings to enhance robustness and E2E performance, though environment-specific tuning of loss weights is recommended. While introducing complexity in unpredictable edge cases, HybridLoss offers a scalable, computation-efficient solution that outperforms conventional approaches by harmonizing individual and collective AV objectives.