Particle.news
Download on the App Store
3 ARTICLES
·
last wk.

KAIST’s APT-RL Lets One Controller Walk, Run and Jump a Quadruped in the Wild

It pairs fast simulation‑pretrained motor skills with onboard depth‑camera and LiDAR perception so a single policy can plan and switch gaits in real terrain.

The quadrupedal robot performing agile locomotion and high-speed running in real-world environments.
The KAIST HOUND quadrupedal robot navigates a forested terrain
Image

Overview

  • The research was published as the cover paper in Science Robotics on July 15, 2026, and was announced by KAIST via EurekAlert! the next day.
  • The team built APT‑RL (Action Pretrained Transformer‑based Reinforcement Learning) to let a single controller select and switch among walking, running, jumping and other gaits in real time.
  • Developers precomputed 15.5 hours of motion trajectories in simulation in about eight minutes to create reusable motor skills, then used reinforcement learning to teach the controller when to sequence and transition those skills.
  • The system runs on KAIST’s HOUND robot using only an onboard depth camera plus LiDAR for near‑ and far‑range sensing; in indoor and outdoor trials the robot negotiated stairs, logs, gaps, rocky trails and reached peak instantaneous speeds up to 6 m/s.
  • KAIST reports APT‑RL outperformed prior RL and hierarchical control methods on success rate, transition smoothness, energy use and training speed; the project was funded by MOTIR, KEIT and the Agency for Defense Development and is aimed at search‑and‑rescue, inspection and defense use cases.