KAIST’s humanoid robot demonstrates world-class mobility, reaching 12 km/h with excellent stability and performing complex moves like the duckwalk and moonwalk, marking it as a next-generation platform for industrial use.
Development of a Human-Like Lower-Body Platform for Next-Generation Humanoids
Professor Park Hae-won’s team at KAIST’s HuboLab has developed the lower-body platform for a next-generation humanoid robot. Designed for human-oriented environments, the robot stands at 165 cm tall and weighs 75 kg, closely resembling human proportions.
The breakthrough is significant, as the team designed and built all key hardware components, achieving full independence in hardware development.
The team built a reinforcement learning algorithm, training the AI in simulation and successfully applying it to real-world use by overcoming the Sim-to-Real Gap. This also secured autonomy in algorithm development.
The humanoid currently runs at 3.25 m/s (12 km/h) and climbs steps over 30 cm, with goals of reaching 4.0 m/s (14 km/h), ladder climbing, and 40+ cm steps.
Professor Hae-Won Park’s group is collaborating with teams from KAIST and MIT to develop a fully integrated humanoid with upper-body hardware and advanced AI.
Equipping Robots with Versatile Skills for Real-World Industrial Tasks
The team is developing technology for robots to handle heavy lifting, operate machinery, and perform coordinated tasks like walking while manipulating, aiming to equip them with versatile abilities for industrial use.
The team also built a single-legged “hopping” robot that can balance, hop repeatedly, and perform a 360-degree somersault.
Reinforcement Learning Breakthrough Without a Biological Reference Model
What makes this achievement especially notable is that imitation learning was not possible, as no biological reference model existed. The researchers built a reinforcement learning AI controller that optimized balance and reduced landing impact, achieving strong performance.
Professor Park Hae-won highlighted that this milestone shows independence in both hardware and software, with plans to build a full humanoid for industrial use and future collaboration with humans.
Ph.D. candidate JongHun Choe will present the hardware findings at Humanoids 2025 on October 1.
Ph.D. candidates Dongyun Kang, Gijeong Kim, and JongHun Choe will present their AI algorithm work as co-first authors at CoRL 2025 on September 29.
The presentation papers have been made available on the arXiv preprint server.
Read the original article on: Techxplore
Read more: What We Know and Don’t About Painkillers and Autism