Event

New Method Proposed for End-to-End Balanced Navigation of Two-Wheeled Robots

1 reports · 1 sources

Xuhao Huang et al. proposed in an arXiv paper a lightweight state-dependent objective coordination method for end-to-end balance and navigation training of two-wheeled inverted pendulum robots. The method introduces a pitch-angle-conditioned scalar coefficient β(s)∈[0,1], which dynamically adjusts the relative weights of the balance stabilization and navigation objectives during PPO training, without the need for an additional control hierarchy or attitude stabilization controllers such as PID or LQR. The paper claims that, compared with other reward weighting, constraint, and control structure schemes, PPO+β(s) achieved success rates of 98% and 95% on a fixed obstacle map and randomly generated maps with high obstacle density, respectively, with small attitude deviations; zero-shot hardware experiments showed that the robot could maintain bounded attitude and stably avoid obstacles using only onboard observations. Training and evaluation code will be made public after the paper is accepted.