Chih-Cheng Liu 劉智誠

This study investigates deep reinforcement learning for bipedal robot gait control using NVIDIA Isaac Gym. A high-degree-of-freedom simulation is developed, and the agent first learns under an unconstrained baseline by directly controlling joint angles. To improve stability and realism, inverse kinematics and Zero Moment Point constraints are progressively introduced, ensuring feasible motion and center of mass balance within the support polygon. Some experiments further employ reference foot trajectories to guide learning and enhance efficiency. The effects of different control conditions—guided versus unguided and with or without constraints—are analyzed to understand how knowledge-based limitations influence the agent’s learning performance and gait behavior.

Chih-Cheng Liu 劉智誠, JAESIK JEONG, CHENG LIN KUO Verified by AVIS CertificateFIRA World Summit 2026Submitted 4 Jun 2026

This paper applies deep reinforcement learning to unmanned vehicles in a simulated environment, aiming to overcome the limitations of traditional path planning methods. The work is divided into two parts: (1) Q-Learning with known environmental information, and (2) Deep Q-Learning for unknown environments. In the Q-Learning approach, the agent’s coordinates in the simulation are used as states to build a Q-table and find the optimal path. For Deep Q-Learning, a neural network is built with PyTorch, and Experience Replay and Fixed Q-targets are used to improve training stability. Experiments verify that, using the weights trained in simulation, the unmanned vehicle can automatically avoid obstacles and reach the destination in a real-world environment.

TSAI CHENG EN, Chih-Cheng Liu 劉智誠, JAESIK JEONGFIRA World Summit 2026Submitted 4 Jun 2026