TRAIN: Terrain-Aware Interactive Navigation
for Quadruped Robots via Reinforcement Learning

1Tohoku University, 2Georgia Institute of Technology


This study introduces TRAIN: TeRrain-Aware Interactive Navigation, a deep reinforcement learning (DRL)-based framework for quadrupedal navigation across unstructured terrain and human-populated environments. Conventional quadrupedal navigation typically separates motion planning from locomotion control, neglecting whole-body constraints and terrain awareness. On the other hand, end-to-end methods are more integrated but require high-frequency sensing, which is often noisy and computationally costly. In addition, most existing approaches assume static environments, limiting their use in human-populated settings. To address these limitations, we propose a two-stage training framework with three DRL pipelines. (1) TRAIN-Loco employs an asymmetric actor–critic (AC) model for quadrupedal locomotion, enabling agile traversal of uneven terrains without explicit terrain or contact observations. (2) TRAIN-Nav applies a symmetric AC framework for navigation with interactive pedestrians, directly mapping transformed LiDAR data to ego-agent actions under differential-drive kinematics. (3) A unified pipeline, TRAIN, integrates TRAIN-Loc and TRAIN-Nav, supporting terrain-aware quadrupedal navigation in uneven and interactive environments. Comprehensive benchmarks against quadrupedal locomotion and robot navigation baselines demonstrate the effectiveness of TRAIN. Hardware experiments further confirm its potential for sim-to-real transfer.


TRAIN framework

Our approach decouples quadrupedal locomotion from crowd navigation by independently training two distinct DRL policies. To ensure reliable locomotion, we first utilize an asymmetric actor–critic (AC) architecture that operates without explicit terrain maps or contact states. By relying strictly on proprioceptive inputs, the policy circumvents the challenges associated with noisy exteroceptive data, thereby streamlining the transition from simulation to real-world deployment. Concurrently, a symmetric AC framework is developed for interactive navigation, employing a differential-drive model governed by kinematic constraints. Subsequently, the navigation policy is fine-tuned and extended for the quadrupedal platform. During this second stage of training, the underlying locomotion policy remains frozen to preserve the learned gait primitives while the higher-level navigation controller is optimized. This integration yields a unified, end-to-end policy capable of robust quadrupedal navigation in uneven and human-populated environments.

TRAIN-Loco comparison

In the first stage, we first train a locomotion policy and evaluate the velocity tracking performance.

TRAIN-Nav comparison

In parallel, we train a crowd navigation policy for a differential-drive model.

TRAIN: unified quadrupedal navigation policy

During the second training phase, the locomotion and navigation policies are harmonized, enabling the quadrupedal robot to navigate uneven and crowded spaces concurrently.

Sim-to-real

We evaluated our policy across a diverse range of environments, including indoor flat and uneven surfaces, outdoor cement and grass slopes, and stairs with a 12 cm step height.

BibTeX

@article{zhu2026train,
  title={TRAIN: Terrain-Aware Interactive Navigation for Quadruped Robots via Reinforcement Learning},
  author={Wei Zhu, Irfan Tito Kurniawan, Ye Zhao, and Mistuhiro Hayashibe},
  journal={arXiv preprint arXiv:2602.12724}
  year={2026},
}