Domain Generalization for Vision-based Driving Trajectory Generation

Yunkai Wang, Dongkun Zhang, Yuxiang Cui, Zexi Chen, Wei Jing, Junbo Chen, Rong Xiong, Yue Wang

Zhejiang University & Alibaba DAMO Academy

[Code], [Arxiv]

Method:

  • We leverage an adversarial learning approach to train a trajectory generator as the decoder.

  • Based on the pre-trained decoder, we infer the latent variables corresponding to the trajectories, and pre-train the encoder by regressing the inferred latent variable.

  • We fix the decoder but fine-tune the encoder with the final trajectory loss.

Contributions:

  • We formulate the domain generalization for driving trajectory generation problem as a non-linear IRM problem. And we propose a set of network training strategies for optimizing the non-linear IRM problem.

  • We implement a trajectory generator with good domain generalization ability. We test our method on both datasets and simulation, showing that our method has a stronger generalization ability than others in both open-loop and closed-loop experiments.

Experiments

In this paper, we use different ablated models to validate the effectiveness of our proposed method on the open-source driving datasets and compare our method with the state-of-the-art trajectory generation method and recent domain generalization methods on the open-source driving datasets and the CARLA simulation.

Dataset: Oxford Radar RobotCar (RobotCar ), KITTI Raw Data (KITTI), and CARLA Dataset.

Metric: average displacement error (ADE).

Generalization performance (average displacement error in meters) on three different datasets. The models are trained on the dataset labeled by "*", and directly generalize to the testing dataset and other two target datasets:

Closed-loop testing of success rate (%, the first one of each item) and average speed (m/s, the second one of each item) in CARLA. Higher metrics have better results. Highest success rates are highlighted in bold font.

Model generalization results of our method under four different weather conditions in CARLA. The model is only trained on RobotCar dataset and directly generalize to CARLA. The discrete red points are the generated trajectory points.

Video of our method: