Quadrupedal locomotion in complex environments requires multiple motor skills, stable gait transitions, and perceptive control over a broad range of speeds.
APT-RL (Action Pretrained Transformer-based Reinforcement Learning) is a unified framework for high-speed, multi-skill locomotion. A single policy selects and transitions between gaits and motor skills using only onboard perception and computation.
In real-world experiments, KAIST HOUND traversed stairs, hurdles, stepping stones, gaps, and fallen branches. It reached instantaneous peak speeds of 4.25 m/s on a 60-cm step and 6 m/s during a drop-down transition on a three-step staircase.
APT-RL first learns reusable locomotion representations from trajectory-optimization data and then uses these representations as priors for reinforcement learning on complex terrain.
Trajectory optimization based on single rigid body dynamics generated 180,000 trajectories (15.5 hours of motion) in 8 minutes. The dataset contains both state trajectories and their corresponding control inputs.
The framework consists of three phases:
APT-RL is deployed on KAIST HOUND, an in-house quadruped robot designed for fast and efficient locomotion. The onboard perception system combines an Intel RealSense D435 depth camera with a 2D LiDAR sensor, and all policy and perception computations run onboard.
For high-impact locomotion above 4 m/s, a custom 3D-printed vibration absorber is mounted between the robot head and the LiDAR sensor. This mechanism reduces impact loads that can exceed 10 g and otherwise cause unreliable measurements or sensor malfunctions.
The deployed system operates solely with onboard perception and computation, without external motion-capture systems or offboard state estimation. The depth camera and 2D LiDAR provide terrain observations in real time.
The policy selects trotting or bounding according to terrain geometry, proprioceptive state, and commanded velocity. This enables gait and motor-skill transitions without predefined switching rules.
For geometrically similar descending obstacles, the policy changes its gait according to obstacle height. It selects trotting for the lower obstacle and bounding when descending from the higher platform.
Lower obstacle: trotting selected
Higher platform: bounding selected
With the commanded velocity held constant, the policy selects trotting for the lower obstacle and switches to bounding for the higher obstacle. This comparison isolates the effect of terrain geometry on gait selection.
Same command, lower obstacle: trotting
Same command, higher obstacle: bounding
On the same outdoor staircase, the policy selects different gaits according to commanded velocity. It trots at 1.8 m/s and switches to bounding at 4.3 m/s.
Same staircase, 1.8 m/s: trotting
Same staircase, 4.3 m/s: bounding
Urban environment. The robot completed a 1.1-km campus route containing multi-level stairs, grass, and inclined ramps. It used trotting to descend stairs at a command speed of 1 m/s and bounding at higher speeds, reaching an instantaneous peak speed of 6 m/s on a three-step staircase.
Forest environment. The robot completed a 0.34-km route containing fallen trees, exposed roots, uneven logs, and leaf-covered or slippery ground. The policy selected bounding over elevated logs and trotting in irregular regions, and returned to bounding as the commanded velocity increased.
Pretraining contains two-dimensional trotting and bounding motions on flat terrain. During reinforcement learning, auxiliary actions extend these pretrained torque priors to behaviors required on complex terrain.
2D motion dataset. Pretraining begins with two-dimensional trotting and bounding trajectories generated through trajectory optimization, including the corresponding joint-torque commands.
Log traversal. The pretrained decoder provides most of the joint torque during flat-ground locomotion, whereas the contribution of the auxiliary action increases during jump initiation.
@article{kang2026aptrl,
title={Agile perceptive multi-skill locomotion for quadrupedal robots in the wild},
author={Kang, Jun-Gill and Park, Jaehyun and Song, Tae-Gyu and Kim, Joon-Ha and Hong, Seungwoo and Park, Hae-Won},
journal={Science Robotics},
year={2026},
}