Shift & Drift: A Zero-Shot Benchmark for Generalizable and Robust Autonomous Driving Motion Planning

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

"Shift & Drift" is a novel dual-track benchmark designed to rigorously stress-test autonomous driving motion planners for generalization and robustness. The Semantic Shift Track converts the aerial DeepScenario Open 3D dataset into the nuPlan simulation framework, enabling zero-shot evaluation of planners trained on North American and Singaporean data against 1,182 scenarios from four German cities and San Francisco, featuring dense pedestrian-cyclist interactions. The State-Distribution Drift Track injects stochastic perturbations, including Additive White Gaussian Noise and Ornstein-Uhlenbeck processes, into the ego vehicle's dynamics to quantify robustness against compounding execution errors. Evaluation of diverse planning paradigms, including imitation learning (IL), reinforcement learning (RL), and rule-based methods, revealed that IL methods exhibit significant failures under semantic shift, with performance drops up to 76%, and suffer from persistent drift. In contrast, the RL-based CaRL planner demonstrated superior resilience, maintaining high safety and progress metrics with at most an 8% performance decay even under high-intensity state-distribution drift.

Key takeaway

For Machine Learning Engineers developing autonomous driving systems, you should prioritize evaluation benchmarks that rigorously test generalization and robustness beyond in-distribution performance. Relying solely on imitation learning methods risks significant performance degradation (up to 76%) in novel urban topologies or under execution perturbations. Instead, consider reinforcement learning approaches, which demonstrate superior resilience and graceful degradation, maintaining safety and progress even under high-intensity state-distribution drift, crucial for real-world deployment.

Key insights

The benchmark reveals that current AD planners, especially IL-based, lack generalization and robustness to semantic and state-distribution shifts.

Principles

• Imitation learning struggles with novel environments.
• RL training fosters recovery from off-policy states.
• Rule-based planners prioritize safety over progress.

Method

The Shift & Drift benchmark uses a dual-track approach: converting aerial 3D datasets for semantic shift evaluation and injecting stochastic noise (AWGN, OU process) into ego dynamics for robustness testing.

In practice

• Evaluate AD planners with cross-dataset scenarios.
• Test robustness using AWGN and OU process noise.
• Prioritize RL for resilient, generalizable AD systems.

Topics

• Autonomous Driving
• Motion Planning
• Zero-Shot Generalization
• State-Distribution Shift
• Reinforcement Learning
• nuPlan Benchmark

Code references

• alessandro-canevaro/Shift-Drift

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Shift & Drift: A Zero-Shot Benchmark for Generalizable and Robust Autonomous Driving Motion Planning — Autonomous driving planners need rigorous zero-shot benchmarks to assess generalization and robustness beyond in-distribution performance.
• Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections — MAPS uses hierarchical DRL with a proto-plan embedding for robust, generalizable coordination of autonomous vehicles at unsignalized intersections.
• TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale — TerraZero enables scalable, zero-demonstration autonomous driving agent training through procedural simulation and self-play, achieving robust generalization.
• RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies — RoboDojo unifies sim-and-real evaluation to diagnose generalist robot manipulation policies' capabilities and limitations.
• FlowR2A: Learning Reward-to-Action Distribution for Multimodal Driving Planning — FlowR2A unifies dense reward supervision with dynamic action proposal generation by learning a reward-conditioned action distribution.
• World Engine: Towards the Era of Post-Training for Autonomous Driving — World Engine enables safe autonomous driving by post-training models on synthesized, safety-critical scenarios, overcoming real-world data scarcity and risks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.