IROS 2025 Keynotes - Mechanisms and Controls: Eiichi Yoshida

Why It Matters

Automating contact perception and embedding human motion strategies enable humanoid robots to operate safely in complex, unstructured environments, unlocking new commercial and societal applications.

Key Takeaways

• Human contact-rich motions guide next-gen humanoid control strategies
• Data-driven contact annotation reduces manual labeling effort dramatically
• Tactile sensors enable whole‑body force feedback for stable manipulation
• Inverse optimal control extracts latent cost functions from human motion
• Integrated visual‑tactile imitation learning improves fragile object handling

Summary

The IROS 2025 keynote by Eiichi Yoshida examined how contact‑rich human motions can be harvested to advance humanoid robot mechanisms and control. Yoshida traced the evolution from a handful of humanoid platforms in 2022 to a burgeoning ecosystem of commercial robots, emphasizing that mastering multi‑contact locomotion and manipulation remains the next frontier.

He outlined a hybrid research agenda that blends model‑based planning with data‑driven learning. By instrumenting humans with tactile skins from the Technical University of Munich, the team captures whole‑body force distributions, maps them onto standard human shape models, and uses inverse optimal control to infer latent cost functions. A self‑attention BQVA network trained on motion‑capture datasets reconstructs ground‑reaction forces and automatically annotates foot‑ground contact states, dramatically cutting manual labeling.

Demonstrations included a robot maintaining balance on a narrow beam using full‑body tactile feedback, and a visual‑tactile transformer that combines RGB images with skin data to imitate fragile‑object manipulation from a few tele‑operation demos. The system can predict contact forces and adjust grip softness even for unseen objects, showcasing the power of integrated tactile sensing and imitation learning.

The work signals a shift toward robots that navigate tight spaces, handle delicate items, and transfer skills across embodiments with minimal human supervision. By automating contact annotation and embedding human‑derived cost metrics, developers can accelerate the deployment of safe, adaptable humanoids for logistics, healthcare, and service sectors.

Original Description

"Keynote Title: ""More Contacts in Interactions: Learning Humanoid Motions from Humans""

Speaker Biography

Eiichi Yoshida is Professor of Tokyo University of Science, at Department of Department of Medical and Robotic Engineering Design, Faculty of Advanced Engineering. He received Ph. D degree from Graduate School of Engineering, the University of Tokyo in 1996. He then joined former Mechanical Engineering Laboratory, later in 2001 reorganized as National Institute of Advanced Industrial Science and Technology (AIST), Tsukuba, Japan. He served as Co-Director of AIST-CNRS JRL (Joint Robotics Laboratory) at LAAS-CNRS, Toulouse, France, from 2004 to 2008, and at AIST, Tsukuba, Japan from 2009 to 2021. He was also Deputy Director of Industrial Cyber-Physical Systems Research Center, and TICO-AIST Cooperative Research Laboratory for Advanced Logistics in AIST from 2020 to 2021. Since 2022. He is Fellow of IEEE and Robotics Society of Japan, and member of SICE and JSME. He received several awards including Best Paper Award in Advance Robotics Journal, and the honor of Chevalier l’Ordre National du Mérite from French Government. His research interests include humanoid robotics, human modeling and task and motion planning for robots.

Abstract

Recent advent of artificial intelligence and rapid progress on motion capacity of humanoid robots are bringing a strong attention expecting industrial and social applications. Despite their highly dynamic motion ability, humanoid robots need further improvements in motions involving contacts with all over their body. We humans generate such contact-rich motions with ease while solving the complex problem of combined discrete contact sequence and continuous dynamic motions is extremely hard. In this talk, I will present ongoing research to address the following challenges: collecting whole-body contact motion data from humans, learning behaviors from those data and synthesizing multi-contact humanoid motions.

"