In addition to visual information, a new quadrupedal robot control technology has been developed that enables the robot to perceive and judge its surroundings and autonomously adjust its walking strategy. This technology is expected to be scalable to various robot platforms, including wheeled and legged robots as well as humanoid robots.


The walking controller trained with DreamWork++ is undergoing a demonstration process. KAIST

The walking controller trained with DreamWork++ is undergoing a demonstration process. KAIST

View original image

KAIST announced on April 12 that Professor Myung Hwan’s research team from the School of Electrical Engineering, in collaboration with Eurobotics Co., Ltd., a KAIST lab-based startup, has developed a quadrupedal robot control technology called "DreamWaQ++."


DreamWaQ++ achieves "cognition-based locomotion" by fusing proprioceptive sensing with exteroceptive sensing based on cameras and LiDAR. The key innovation is that the robot can detect obstacles in advance and proactively adjust its walking strategy, moving beyond simple reactive control to enable understanding and judgment of its environment.


This technology addresses the shortcomings of "DreamWaQ," the team's previous development. DreamWaQ was based on "blind locomotion" technology, which estimates terrain and walks using only proprioceptive sensors such as joint encoders and inertial sensors. Its key feature was robust movement without visual information.


However, it was limited in that the robot could only adjust its movement when its legs came into direct contact with an obstacle.


To overcome this limitation, the research team designed and applied a multi-sensory reinforcement learning architecture to DreamWaQ++, enabling real-time control based on lightweight computational operations. The team also secured both stability—allowing the robot to automatically switch to walking based on alternative sensory modalities if sensor errors occur—and scalability for application to various robot platforms.


The performance of DreamWaQ++ was also demonstrated experimentally. In tests, a robot equipped with DreamWaQ++ completed a staircase course of 50 steps (30.03 meters horizontal, 7.38 meters vertical) in just 35 seconds, outperforming both the blind locomotion controller and commercial cognition-based controllers.


Professor Myunhyun, KAIST

Professor Myunhyun, KAIST

View original image

Notably, in steep environments, a robot with DreamWaQ++ was able to stably climb a 35-degree slope—3.5 times steeper than its training condition (10 degrees)—and actively adjusted its posture to reduce the rear leg motor torque by 1.5 times compared to previous models.


In various obstacle scenarios, the robot demonstrated learning-based cognitive abilities, such as autonomously selecting efficient paths without separate path planning. In uncertain drop terrains, the robot exhibited "exploration behavior," voluntarily stopping to probe the ground before proceeding.


Agility was also demonstrated: the robot successfully crossed a 41 cm obstacle while carrying a 2.5 kg payload. In simulations, the research team explained that the ANYmal-C (a representative quadrupedal robot developed by ETH Zurich) could overcome obstacles up to 1.0 meters high, while the KAIST Hound (a quadrupedal robot developed by Professor Park Haewon's team at KAIST) equipped with DreamWaQ++ managed obstacles up to 1.5 meters in height.


A noteworthy experimental result is that DreamWaQ++ achieved an 80% success rate in climbing 42 cm stairs, even though it was only trained on relatively low obstacles (27 cm). This indicates that the robot is not simply repeating learned scenarios but is capable of autonomously adapting to new environments.


The research team expects that, in the future, DreamWaQ++ technology can be used for disaster response, industrial facility inspection, and in environments such as forests and agricultural fields where wheeled robots have difficulty accessing.


Professor Myung Hwan stated, "This study confirms that robots have advanced beyond merely moving to understanding and autonomously judging their environment. The research team plans to expand DreamWaQ++ as an 'intelligent mobility technology' so that it can be applied and utilized in a wide range of real-world environments."


This research was led by Dr. I Made Aswin Nahrendra (currently a researcher at Krafton and KAIST Ph.D. graduate) as first author, with Dr. Yoo Byungho (CEO of Eurobotics Co., Ltd.), Dr. Oh Minho (CTO of Eurobotics Co., Ltd.), Dr. Lee Donggyu (CTO of Eurobotics Co., Ltd.), Lee Seunghyun, Lee Hyunwoo (KAIST), and Dr. Lim Hyungtae (postdoctoral researcher at MIT) as co-authors.



The research results were published in the robotics journal "IEEE Transactions on Robotics (T-RO)" in February.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing