Haoyu Wang

Directed Temporal Representations for Offline Visual Control

arXiv · October 2026

Chenyang Yuan, Haoyu Wang, Zhuo Sun, Xiaoyuan Cheng

This work learns a control representation that captures how long reaching a visual goal takes. Starting from frozen LeWorldModel (LeWM) features, DTRC learns a directed temporal quasimetric from offline trajectories, allowing forward and reverse reaching costs to differ. Short-range trajectory offsets calibrate the distance, bootstrapped targets extend it to longer horizons, and action-conditioned consistency ties the representation to local transition dynamics.

A reduction in this distance measures progress toward the goal and supplies a temporal critic for training a goal-conditioned flow policy. Model-assisted refinement is constrained by behavior support and dynamics agreement. Across ten visual tasks spanning navigation, reaching, and manipulation, DTRC performs strongly against planning and direct-policy baselines. Temporal supervision improves the same flow-policy architecture on all four LeWM tasks—PushT, TwoRoom, Reacher, and Cube—while execution requires no iterative trajectory search.

Fisher Decorator: Refining Flow Policy via a Local Transport Map

NeurIPS 2026

Xiaoyuan Cheng^*, Haoyu Wang^*, Wenxuan Yuan, Ziyan Wang, Zonghao Chen, Li Zeng, Zhuo Sun

* Co-first authors (equal contribution).

This work introduces a Fisher-guided local transport map to improve flow policies while preserving multimodal behavior. It achieves state-of-the-art performance on OGBench and D4RL, with applications in robot locomotion and manipulation. We evaluate it in simulation with both state and visual inputs, and show that the learned policy can continue improving through online fine-tuning.

GEM: Guided Expectation-Maximization for Behavior-Normalized Candidate Action Selection in Offline RL

Haoyu Wang, Jingcheng Wang, Shunyu Wu, Xinwei Xiao

This work studies how to choose better actions for robot locomotion and navigation when the data contains several plausible behaviors. GEM combines Gaussian-mixture policies, value estimates, and behavior support to rank candidate actions. We evaluate it on D4RL Locomotion, AntMaze, and Maze2D, with a practical benefit: inference compute can be adjusted without retraining.

Energy-efficient high-precision electric seed metering system: achieving high-speed, high-precision seeding with low energy consumption

Haoyu Wang, Xiaoshuang Zhang, Li Yang, Dongxing Zhang, Tao Cui, Xiantao He, Jinsheng Mu, Lei Bao

This work develops high-precision motor control for high-speed electric seed metering. We combine linear active disturbance rejection control with field-oriented control to improve speed regulation and disturbance rejection while reducing energy use. The resulting controller provides a practical foundation for fast, precise agricultural machinery.

Comparative investigation and evaluation of electric-drive seed-metering systems across diverse speed ranges for enhanced high-precision seeding applications

Haoyu Wang, Li Yang, Dongxing Zhang, Tao Cui, Xiantao He, Tainpu Xiao, Hongsheng Li, Zhaohui Du, Chunji Xie

This work helps guide motor and drive-system selection for intelligent agricultural machinery. We compare five motor configurations under speed changes, disturbances, turn compensation, and section control. The results show where each configuration performs well and where it falls short, providing a reference for future precision-seeding systems and agricultural automation.