Xiaosong Jia (贾萧松)
jiaxiaosong1997 [At] gmail [dot] com or jiaxiaosong [At] fudan [dot] edu [dot] cn

I am interestied in Autonomous Driving and Dexterous Hand. Currently, I am focusing on (i) building authentic world simulators with generative models & reconstructive models and then (ii) combining imitation learning & reinforcement learning to train decision-making agent in an end-to-end way.
I am recruiting research assistants, master, and Ph.D. students。 (报考智能机器人与先进制造创新学院，可信具身智能智能研究院。报考创智学院、中关村学院、河套学院并获得offer的可以招收。).
更多招生、合作、研究信息请见 此页

* denotes Co-First Authors, ✉ denotes Correspondence Authors

Mixture of expert π0 for autonomous driving.

Introduce memory retrieval for autonomous driving. New Paradigm.

Tokenizer for trajectory. Champion of Waymo Sim Agents Challenge 2026 ($10000).

Efficiently traverse in the video.

Interleaved inputs for VLA enabling zero-shot visual prompts.

Clsoed-loop autonomous driving benchmark based on generative models which could replay real world data in a reactive way.

End-to-end Reinforcement Learning for Autonomous Driving.

Unifying all tasks with queries and attention. Strong Scaliability.

Transformer based Fusion with 73.7 NDS + 10.1 FPS in nuScenes val set.

First benchmark for multi-ability end-to-end autonomous driving.

GPT-style Motion Prediction. State-of-the-art performance on Waymo Motion.

World model based reinforcement learning for autonomous driving. The first & only learning-based model could solve 39 complex scenearios in CARLA Leaderboard 2.0.

Planning-oriented data selection for end-to-end autonomous driving. Training with 30% data beats full data.

The first survey of LLM for autonomous driving. 800+ stars. Continuous updating.

New paradigm for end-to-end autonomous driving without causal confusion.

BEV-based scalable end-to-end autonomous drving model.

All modules in one Transformer-based end-to-end network for autonomous driving.

Unified heterogeneous graph neural network for driving scene encoding. SOTA methods on INTERACTION and Waymo challenge.

Self-supervised pretraining for policy learning

Trajectory-guided control paradigm for end-to-end autonomous driving. 1st method on Carla Leaderboard, with only a monocular camera, outperforming other methods with multiple cameras and LiDAR by a large margin.

A plug-and-play module for trajectory prediction by enhancing the temporal correlation among the predicted time-steps.

Rethink the invariance property of the coordinate reprentation for trajectory prediction.

Co-organized the competition in ICCV 2021. I was responsible for the design and implementation of the Joint Prediction and Conditional Prediction Tracks.

Unsupervisedly extracting interactive behaviors in a whether, when, and what hierarchy.

This website's source code is from Jon Barron