Selected work

Representative Papers

ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving

Feb 11, 2026

Existing end-to-end autonomous driving world models suffer from redundant modeling of static regions and insufficient interaction between trajectories and scene dynamics, limiting planning performance. To address these issues, this work proposes the Temporal Residual World Model (TR-World), which directly extracts dynamic object information through temporal residuals without relying on explicit detection or tracking, and predicts high-fidelity future bird’s-eye-view (BEV) representations by leveraging current BEV features. Furthermore, a Future-Guided Trajectory Refinement module (FGTR) is introduced to enable bidirectional co-optimization between trajectories and future scene context, while sparse spatiotemporal supervision is employed to prevent training instability. Evaluated on nuScenes and NAVSIM, the proposed approach significantly improves planning accuracy and robustness, achieving state-of-the-art performance.

0 citationsRead paper

S2R-Bench: A Sim-to-Real Evaluation Benchmark for Autonomous Driving

May 24, 2025

Existing autonomous driving perception benchmarks rely heavily on simulation, failing to capture real-world degradation scenarios—such as extreme weather and sensor failures—leading to a severe sim-to-real performance gap. Method: We introduce S2R-Bench, the first real-scenario-oriented perception robustness benchmark: (i) it establishes a sim-to-real evaluation framework grounded in real-vehicle-collected camera/LiDAR corruption data; (ii) it systematically models four-dimensional correlated degradations—temporal, weather, illumination, and road conditions; and (iii) it proposes a consistency analysis framework bridging simulation and real-world results, alongside a standardized corruption protocol and open-source evaluation toolkit (GitHub). Results: Experiments reveal that mainstream perception models suffer an average 32.7% mAP drop under real-world corruptions, exposing critical deployment risks. S2R-Bench provides a reproducible, comparable, and operationally relevant safety evaluation standard for perception algorithms.

0 citationsRead paper

Recent publications

Latest Papers

ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving

Feb 11, 2026

Existing end-to-end autonomous driving world models suffer from redundant modeling of static regions and insufficient interaction between trajectories and scene dynamics, limiting planning performance. To address these issues, this work proposes the Temporal Residual World Model (TR-World), which directly extracts dynamic object information through temporal residuals without relying on explicit detection or tracking, and predicts high-fidelity future bird’s-eye-view (BEV) representations by leveraging current BEV features. Furthermore, a Future-Guided Trajectory Refinement module (FGTR) is introduced to enable bidirectional co-optimization between trajectories and future scene context, while sparse spatiotemporal supervision is employed to prevent training instability. Evaluated on nuScenes and NAVSIM, the proposed approach significantly improves planning accuracy and robustness, achieving state-of-the-art performance.

0 citationsRead paper

S2R-Bench: A Sim-to-Real Evaluation Benchmark for Autonomous Driving

May 24, 2025

Existing autonomous driving perception benchmarks rely heavily on simulation, failing to capture real-world degradation scenarios—such as extreme weather and sensor failures—leading to a severe sim-to-real performance gap. Method: We introduce S2R-Bench, the first real-scenario-oriented perception robustness benchmark: (i) it establishes a sim-to-real evaluation framework grounded in real-vehicle-collected camera/LiDAR corruption data; (ii) it systematically models four-dimensional correlated degradations—temporal, weather, illumination, and road conditions; and (iii) it proposes a consistency analysis framework bridging simulation and real-world results, alongside a standardized corruption protocol and open-source evaluation toolkit (GitHub). Results: Experiments reveal that mainstream perception models suffer an average 32.7% mAP drop under real-world corruptions, exposing critical deployment risks. S2R-Bench provides a reproducible, comparable, and operationally relevant safety evaluation standard for perception algorithms.

0 citationsRead paper