Agile perceptive multi-skill locomotion for quadrupedal robots in the wild

🐕">

Agile perceptive multi-skill locomotion for quadrupedal robots in the wild

Jun-Gill Kang†, Jaehyun Park†, Tae-Gyu Song, Joon-Ha Kim, Seungwoo Hong‡, Hae-Won Park‡

Science Robotics, July 2026 — Cover Article

Paper

arXiv

Video

Quadrupedal locomotion in complex environments requires multiple motor skills, stable gait transitions,<br>and perceptive control over a broad range of speeds.

APT-RL (Action Pretrained Transformer-based Reinforcement Learning) is a unified framework<br>for high-speed, multi-skill locomotion. A single policy selects and transitions between gaits and motor<br>skills using only onboard perception and computation.

In real-world experiments, KAIST HOUND traversed stairs, hurdles, stepping stones, gaps, and fallen<br>branches. It reached an instantaneous peak speed of 4.25 m/s while traversing a 60-cm<br>step and 6 m/s during a drop-down transition on a three-step staircase.

Quadrupedal locomotion in complex environments requires multiple motor skills, stable gait transitions,<br>and perceptive control over a broad range of speeds.

APT-RL (Action Pretrained Transformer-based Reinforcement Learning) is a unified framework<br>for high-speed, multi-skill locomotion. A single policy selects and transitions between gaits and motor<br>skills using only onboard perception and computation.

In real-world experiments, KAIST HOUND traversed stairs, hurdles, stepping stones, gaps, and fallen<br>branches. It reached instantaneous peak speeds of 4.25 m/s on a 60-cm step and<br>6 m/s during a drop-down transition on a three-step staircase.

Summary Video

APT-RL Framework

APT-RL first learns reusable locomotion representations from trajectory-optimization data and then<br>uses these representations as priors for reinforcement learning on complex terrain.

Trajectory optimization based on single rigid body dynamics generated 180,000 trajectories<br>(15.5 hours of motion) in 8 minutes . The dataset contains both state trajectories and their<br>corresponding control inputs.

The framework consists of three phases:

Representation learning. A Transformer-based variational autoencoder learns a<br>structured latent representation and gait-specific torque decoders from trotting and bounding data.

Reinforcement learning. The policy produces latent actions for the pretrained<br>decoders together with auxiliary actions that adapt the learned skills to complex terrain.

Perceptual distillation. A student encoder is trained from a privileged teacher and<br>receives depth images, 2D LiDAR measurements, and proprioceptive observations for deployment.

Deployment Pipeline

APT-RL is deployed on KAIST HOUND , an in-house quadruped robot designed for fast and<br>efficient locomotion. The onboard perception system combines an Intel RealSense D435 depth camera with<br>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<br>robot head and the LiDAR sensor. This mechanism reduces impact loads that can exceed 10 g and otherwise<br>cause unreliable measurements or sensor malfunctions.

The deployed system operates solely with onboard perception and computation, without external motion-capture<br>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<br>velocity. This enables gait and motor-skill transitions without predefined switching rules.

Skill Selection

For geometrically similar descending obstacles, the policy changes its gait according to obstacle height.<br>It selects trotting for the lower obstacle and bounding when descending from the higher platform.

Lower obstacle: trotting selected

Higher platform: bounding selected

Terrain-Conditioned Gait Selection

With the commanded velocity held constant, the policy selects trotting for the lower obstacle and<br>switches to bounding for the higher obstacle. This comparison isolates the effect of terrain geometry<br>on gait selection.

Same command, lower obstacle: trotting

Same command, higher obstacle: bounding

Command-Conditioned Gait Selection Outdoors

On the same outdoor staircase, the policy selects different gaits according to commanded velocity.<br>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

Tested in the Wild

Urban environment. The robot completed a 1.1-km campus route containing multi-level stairs,<br>grass, and inclined ramps. It used trotting to descend stairs at a command speed of 1 m/s and bounding<br>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,<br>uneven logs, and leaf-covered or slippery ground. The policy selected bounding over elevated...