RoCA: Robust Cross-Domain End-to-End Autonomous Driving
Rajeev Yasarla, Shizhong Han, Hsin-Pai Cheng, Litian Liu, Shweta Mahajan, Apratim Bhattacharyya, Yunxiao Shi, Risheek Garrepalli, Hong Cai, Fatih Porikli
TL;DR
RoCA is a novel probabilistic framework that enhances the robustness and adaptability of end-to-end autonomous driving models across different domains without requiring extensive retraining or extra inference costs.
Contribution
RoCA introduces a Gaussian process-based probabilistic modeling approach for cross-domain E2E autonomous driving, improving generalization and adaptation capabilities.
Findings
Outperforms direct finetuning in target domains
Enhances cross-domain generalization without extra inference cost
Achieves strong domain adaptation performance
Abstract
End-to-end (E2E) autonomous driving has recently emerged as a new paradigm, offering significant potential. However, few studies have looked into the practical challenge of deployment across domains (e.g., cities). Although several works have incorporated Large Language Models (LLMs) to leverage their open-world knowledge, LLMs do not guarantee cross-domain driving performance and may incur prohibitive retraining costs during domain adaptation. In this paper, we propose RoCA, a novel framework for robust cross-domain E2E autonomous driving. RoCA formulates the joint probabilistic distribution over the tokens that encode ego and surrounding vehicle information in the E2E pipeline. Instantiating with a Gaussian process (GP), RoCA learns a set of basis tokens with corresponding trajectories, which span diverse driving scenarios. Then, given any driving scene, it is able to…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The paper tackles the critical and practical problem of cross-domain generalization and adaptation for E2E models, which is a significant barrier to real-world deployment. 2. The approach of using a Gaussian Process over a learned codebook of tokens. This probabilistic formulation allows for principled uncertainty estimation. 3. The framework demonstrates consistent performance improvements across multiple base E2E models, and across various challenging settings, including closed-loop simulat
1. The paper is hard to follow, and the writing needs to be greatly improved. This is a major barrier to understanding the contribution and evaluating the method's soundness. - Key terminologies (e.g., 'codebook', 'basis token', 'ego and agent states') are used extensively in the abstract and introduction before being formally defined in the methodology. The author is expected to define or explain implementations of terminologies when they first appear. - The paper lacks a clear, formal
1. The proposed framework is general and works with several end-to-end models. 2. The motivation of integrating Gaussian process (GP) is clear.
1. The closed-loop evaluation in Table 1 should be on the whole validation set (220 routes) of Bench2Drive rather than Dev10 subset. The Dev10 is proposed for quick development and ablation studies. For main results, it is better to evaluate on the whole validation set for comprehensive evaluation and convenient comparation. 2. Although the Table 1 reports the closed-loop results, the experiments of domain adaptation and robustness are all open-loop. It is recommended to measure these significan
- For the first time, Gaussian processes are introduced as an uncertainty-aware module into the token/feature space of E2E autonomous driving for probabilistic trajectory modeling. This approach provides a theoretically complete Bayesian framework for autonomous driving decision-making, naturally quantifying the uncertainty of predictions. - Utilizing an active learning strategy guided by Gaussian process prediction variance, performance comparable to or even better than randomly selected 10% o
ROCA is an add-on module that relies on the quality of the Ego/Agent Tokens output by the Scene Encoder of the baseline E2E model. If the base model's tokens exhibit significant instability and domain shift across domains, ROCA's performance will be limited.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDomain Adaptation and Few-Shot Learning · Multimodal Machine Learning Applications · Advanced Neural Network Applications
MethodsBalanced Selection · Gaussian Process · Sparse Evolutionary Training