TNavRL: Cross-Modal Transformer for Humanoid Visual Navigation

Abstract

Efficient and safe navigation in cluttered environments remains a fundamental challenge for humanoid robots, particularly due to their complex dynamics and the need to process ego-centric visual observations under significant camera motion. Traditional modular navigation pipelines often suffer from inter-module latency, error accumulation, and scenario-dependent parameter tuning, limiting their robustness and generalization. To address these limitations, we propose TNavRL, a reinforcement learning–based navigation framework for humanoid robots. TNavRL employs a dual-branch architecture that independently encodes proprioceptive and visual observations, and fuses their latent representations via a cross-modal Transformer. This design effectively correlates noisy visual streams with stable proprioceptive states, enabling robust spatial understanding during dynamic locomotion. Crucially, to ensure safety independent of reward shaping, we enforce explicit kinematic constraints on the policy's action space. Extensive simulation and real-world experiments demonstrate that TNavRL achieves robust zero-shot sim-to-real transfer and real-time, collision-free navigation on onboard hardware, significantly outperforming state-of-the-art learning-based navigation baselines. More results are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/FanMecha/TNavRL</uri>.