IROS 2025 Keynotes - Humanoid Robot Systems: Wei Zhang

Why It Matters

Zhang’s data‑first, compute‑focused framework lowers the barrier for creating versatile humanoid robots, enabling faster commercialization and broader industry adoption.

Key Takeaways

• Theory insights can hinder learning; prioritize data-first approach.
• Model‑based and learning methods both aim to reduce computation.
• Limax Dynamics released versatile Chong platform and 31‑DOF OI humanoid.
• Unified RL whole‑body controller enables autonomous perception‑driven manipulation.
• Future humanoid AI requires foundation models and efficient reward design.

Summary

Wei Zhang’s IROS 2025 keynote focused on the evolution of humanoid robot systems and introduced his company, Limax Dynamics, as a catalyst for practical, research‑grade platforms. Drawing on his background as a control theorist turned robotics entrepreneur, Zhang outlined the shift from pure theoretical models to data‑driven reinforcement learning (RL) pipelines, coining the term “humanoid physical intelligence.” He highlighted Limax’s two flagship products: the modular Chong platform for local motion research and the 31‑degree‑of‑freedom OI humanoid, both equipped with open SDKs to accelerate development.

Zhang emphasized two core insights. First, theoretical bias can restrict solution spaces; a data‑first principle often yields better policies. Second, theory should be measured against computational cost, not against data, meaning that as computational tools improve, the reliance on handcrafted models diminishes. He advocated a divide‑and‑conquer strategy—splitting tasks into motion generation (VA) and whole‑body control—to cut data requirements while preserving flexibility. The whole‑body controller, built entirely with RL rather than model‑predictive control, can execute perception‑driven manipulation with a single policy, eliminating the traditional hand‑off between locomotion and manipulation.

Concrete examples underscored his points. The OI humanoid demonstrated autonomous local manipulation using one RL policy, and a teacher‑student training scheme reduced training time by roughly 50 %. Reward engineering that prioritizes stability over pure motion tracking enabled robust navigation on uneven terrain without extensive tuning. These results showcase how modular hardware and streamlined RL pipelines can produce smooth, adaptable motions that were previously limited to task‑specific solutions.

The broader implication is a roadmap toward scalable, foundation‑model‑style humanoid AI. By minimizing data costs through modular design and efficient reward structures, researchers and companies can accelerate the deployment of versatile humanoids for manufacturing, logistics, and service applications. This paradigm shift promises lower R&D overhead, faster iteration cycles, and a competitive edge for firms that adopt data‑centric, compute‑efficient approaches to robot intelligence.

Original Description

"Keynote Title: ""Towards Physical Intelligence in Humanoid Robotics"" Speaker Biography

Dr Zhang is a Professor at SUSTech, where he directs the Control and Learning for Robotics and Autonomy (CLEAR) Lab, and the Founder of LimX Dynamics, a fast-growing robotics company building next-generation humanoid and general-purpose robots. Dr. Zhang received his B.S. from the University of Science and Technology of China, and his Ph.D. in Electrical Engineering from Purdue University. He completed a postdoctoral appointment in EECS at UC Berkeley, and previously served as an Associate Professor at The Ohio State University. He is a recipient of the U.S. NSF CAREER Award, and a Pengcheng Distinguished Scholar.

Abstract

Humanoid robots are poised to become a transformative technology, with their societal and industrial roles expanding rapidly. This surge creates exceptional opportunities for both fundamental research and industry innovation in humanoid robotics. At the core of their future capabilities lies physical intelligence—the ability to sense, plan, decide, and execute motion with human-like versatility. We present two complementary foundation models as the backbone of this intelligence: a prefrontal/sensory–cortex-like model for perception, planning, and decision, and a motor-cortex–like model for precise, whole-body control. Both rely on large-scale, multimodal data, but each requires a specialized data pipeline designed to meet its unique functional and temporal requirements. The talk will highlight key industrial achievements shaping today’s humanoid landscape and present cutting-edge academic advances toward the two foundation models of humanoid physical intelligence. I will present a holistic perspective that unites model- and rule-based methods with data-driven approaches, emphasizing key research findings and supporting experimental evidence. I will conclude by highlighting the key research challenges for humanoid robots to reach their full potential in society.

"