Haozhe Du

Bio: Haozhe Du is an academic researcher from Zhejiang University. The author has contributed to research in topics: Filter (signal processing) & Kalman filter. The author has co-authored 3 publications.

Multi-Stage Decision-Making Skill Learning for Soccer Robot

Abstract: The application scenarios of robots are becoming more and more complex, leading to higher and higher demand on skills for robots. This paper proposes a method to learn complex skill by decomposing it into basic skills, then train them respectively as well as combination into a whole, where reinforcement learning are used both in basic skill learning and in the integration. The method proposed is validated on the RoboCup small soccer robot platform via the skill of chasing and shooting ball and tested in both simulation environment and real world. It is verified our method has achieved a higher success rate comparing with traditional methods. Code is available here.

Multi-Agent Trajectory Prediction Based on Graph Neural Network

Abstract: Many tasks have demand on precise predictions of agents or moving objects. Previous prediction methods usually only focus on the kinematic model of moving objects or the environment. However, the target tasks of agents may influence the prediction of agents in great sense, especially in tasks of confrontation. Therefore traditional methods cannot work well in such scenes. In this paper, we propose a heterogeneous graph neural network method to deal with the multi-agent trajectory prediction problem. Our method can aggregate and pass messages representing environment and also agents' tasks due to the elaborate design of graph neural network structure. We validate our method on the Robocup Small Size League simulation platform which focuses on multi-agent coordination and confrontation in the form of soccer games. After making our own ZJUNlictSSL dataset, we predict the position of all robots on the pitch of certain time gaps based on the limited information we get from vision. The results prove that our method is of high prediction accuracy and low prediction error compared with conventional kinematic motion methods. Code is available here.

Fully Differentiable and Interpretable Model for VIO with 4 Trainable Parameters.

Abstract: Monocular visual-inertial odometry (VIO) is a critical problem in robotics and autonomous driving. Traditional methods solve this problem based on filtering or optimization. While being fully interpretable, they rely on manual interference and empirical parameter tuning. On the other hand, learning-based approaches allow for end-to-end training but require a large number of training data to learn millions of parameters. However, the non-interpretable and heavy models hinder the generalization ability. In this paper, we propose a fully differentiable, interpretable, and lightweight monocular VIO model that contains only 4 trainable parameters. Specifically, we first adopt Unscented Kalman Filter as a differentiable layer to predict the pitch and roll, where the covariance matrices of noise are learned to filter out the noise of the IMU raw data. Second, the refined pitch and roll are adopted to retrieve a gravity-aligned BEV image of each frame using differentiable camera projection. Finally, a differentiable pose estimator is utilized to estimate the remaining 4 DoF poses between the BEV frames. Our method allows for learning the covariance matrices end-to-end supervised by the pose estimation loss, demonstrating superior performance to empirical baselines. Experimental results on synthetic and real-world datasets demonstrate that our simple approach is competitive with state-of-the-art methods and generalizes well on unseen scenes.

Cooperation Mode of Soccer Robot Game Based on Improved SARSA Algorithm

Abstract: Football is one of the most popular sports in the world, and its competition has received increasing people’s attention. With the increasing number of robot football competitions, more strategic planning is needed for robot football matches. In a tournament, each player has their own task and must have the skills to complete it. In this paper, we use the RoboCup2D platform to give details of the server and client roles, introduce the agent model in RoboCup2D, and compare the plan design and scheme design presented in the current study using the SARSA algorithm, one of the augmented methods classified as TD learning metrics. In addition, heuristic information was introduced and implemented to enhance learning through the sharing of Q values between participants and reinforcement learning. A comparative analysis of the feasibility of the SARSA algorithm in the context of its application in RoboCup2D was carried out, and the experimental results proved that our algorithm was effective in improving the team’s offensive and defensive capabilities.