Subsequent-generation humanoid robotic can do the moonwalk

KAIST analysis group’s independently developed humanoid robotic boasts world-class driving efficiency, reaching speeds of 12km/h, together with wonderful stability, sustaining stability even with its eyes closed or on tough terrain. Moreover, it could carry out complicated human-specific actions such because the duckwalk and moonwalk, drawing consideration as a next-generation robotic platform that may be utilized in precise industrial settings.

Professor Park Hae-won’s analysis group on the Humanoid Robotic Analysis Middle (HuboLab) of KAIST’s Division of Mechanical Engineering developed the decrease physique platform for a next-generation humanoid robotic. The developed humanoid is characterised by its design tailor-made for human-centric environments, focusing on a top (165cm) and weight (75kg) much like that of a human.

The importance of the newly developed decrease physique platform is immense because the analysis group instantly designed and manufactured all core elements, together with motors, reducers, and motor drivers. By securing key elements that decide the efficiency of humanoid robots with their very own expertise, they’ve achieved technological independence by way of {hardware}.

As well as, the analysis group skilled an AI controller by means of a self-developed reinforcement studying algorithm in a digital setting, efficiently utilized it to real-world environments by overcoming the Sim-to-Actual Hole, thereby securing technological independence by way of algorithms as properly.

Presently, the developed humanoid can run at a most velocity of three.25m/s (roughly 12km/h) on flat floor and has a step-climbing functionality of over 30cm (a efficiency indicator displaying how excessive a curb, stairs, or impediment will be overcome). The group plans to additional improve its efficiency, aiming for a driving velocity of 4.0m/s (roughly 14km/h), ladder climbing, and over 40cm step-climbing functionality.

Professor Hae-Received Park’s group is collaborating with Professor Jae-min Hwangbo’s group (arms) from KAIST’s Division of Mechanical Engineering, Professor Sangbae Kim’s group (palms) from MIT, Professor Hyun Myung’s group (localization and navigation) from KAIST’s Division of Electrical Engineering, and Professor Jae-hwan Lim’s group (vision-based manipulation intelligence) from KAIST’s Kim Jaechul AI Graduate Faculty to implement an entire humanoid {hardware} with an higher physique and AI.

By means of this, they’re growing expertise to allow the robotic to carry out complicated duties akin to carrying heavy objects, working valves, cranks, and door handles, and concurrently strolling and manipulating when pushing carts or climbing ladders. The last word objective is to safe versatile bodily talents to answer the complicated calls for of precise industrial websites.

Throughout this course of, the analysis group additionally developed a single-leg “hopping” robotic. This robotic demonstrated high-level actions, sustaining stability on one leg and repeatedly hopping, and even exhibited excessive athletic talents akin to a 360-degree somersault.

Particularly in a state of affairs the place imitation studying was not possible as a result of absence of a organic reference mannequin, the analysis group achieved important outcomes by implementing an AI controller by means of reinforcement studying that optimizes the middle of mass velocity whereas lowering touchdown impression.

Professor Park Hae-won said, “This achievement is a vital milestone that has achieved independence in each {hardware} and software program features of humanoid analysis by securing core elements and AI controllers with our personal expertise.

“We are going to additional develop it into an entire humanoid, together with an higher physique to resolve the complicated calls for of precise industrial websites and moreover, foster it as a next-generation robotic that may work alongside people.”

The outcomes of this analysis will likely be introduced by JongHun Choe, a Ph.D. candidate in Mechanical Engineering, as the primary writer, on {hardware} improvement at Humanoids 2025, a world humanoid robotic specialised convention held on October 1st.

Moreover, Ph.D. candidates Dongyun Kang, Gijeong Kim, and JongHun Choe from Mechanical Engineering will current the AI algorithm achievements as co-first authors at CoRL 2025, the highest convention in robotic intelligence, held on September twenty ninth.

The presentation papers can be found on the arXiv preprint server.