Agile perceptive multiskill locomotion for quadrupedal robots in the wild

KAIST’s DRCD Lab shows its real HOUND quadruped using APT-RL, a transformer-based controller that autonomously switches gaits and skills from onboard depth and LiDAR. The robot trots, bounds, and jumps over stairs, hurdles, logs, stepping stones, and forest trails, reaching up to 6 m/s with only onboard perception. The work is the July 2026 Science Robotics cover paper on agile multi-skill locomotion in the wild.