Biomimetics 2026, 11(8), 570; https://doi.org/10.3390/biomimetics11080570 (registering DOI) - 9 Aug 2026
Abstract
Adaptive locomotion across continuous terrain transitions remains difficult for hexapod robots because contact timing, body attitude, support height, and load distribution change simultaneously along a route. This paper presents a unified proprioception-driven deep reinforcement learning and central pattern generator (DRL-CPG) framework for terrain-transition
[...] Read more.
Adaptive locomotion across continuous terrain transitions remains difficult for hexapod robots because contact timing, body attitude, support height, and load distribution change simultaneously along a route. This paper presents a unified proprioception-driven deep reinforcement learning and central pattern generator (DRL-CPG) framework for terrain-transition locomotion without visual terrain classification, explicit terrain labels, or terrain-specific controller switching. A high-level proximal policy optimization policy maps a 46-dimensional proprioceptive observation to a three-dimensional CPG modulation action comprising oscillation amplitude, swing-phase frequency, and turn modulation. A coupled six-node Hopf oscillator network then expands these modulated parameters into phase-coordinated rhythmic commands, which are mapped to the 18 joint targets of a JetHexa hexapod and executed by a low-level proportional-derivative controller. The observation space contains body linear velocity, body angular velocity, relative joint positions, relative joint velocities, the previous three-dimensional policy action, and inertial measurement unit (IMU)yaw/heading relative to the initial track direction. A continuous route consisting of flat ground, uphill stairs, irregular terrain, downhill stairs, and a recovery segment is defined to evaluate transition-aware locomotion using route completion, velocity-tracking error, lateral deviation, and roll/pitch fluctuation. Compared with the fixed-parameter CPG and end-to-end DRL baselines, the proposed method increased the full-distance success rate at 4.7 m from 9% and 20%, respectively, to 88%, while maintaining smoother velocity, lateral deviation, and roll/pitch responses. The framework preserves the rhythmic prior of CPG control while reducing the exploration burden of reinforcement learning, providing a compact formulation for adaptive hexapod locomotion across terrain transitions.
Full article
(This article belongs to the Section Locomotion and Bioinspired Robotics)
►
Show Figures
Graphical abstract
