Wei Zhang of LimX Dynamics, a control theorist turned robotics entrepreneur, presents at the IROS 2025 humanoid session on robot systems and data-driven physical intelligence.
Zhang describes moving from control and learning theory to applied motion control around 2017, when model-based and learning approaches were both in play. He argues that strong prior assumptions can constrain learning, so researchers should let data guide policy design. He frames theory as a way to reduce computation, whether a method is model-based or data-driven.
Zhang says LimX Dynamics began in 2022 and describes TRON as a configurable platform for local motion research, mobility, and mobile manipulation. The company’s full-size Oli humanoid has 31 degrees of freedom and an SDK. Zhang reports a fully autonomous perceptive manipulation pipeline using one reinforcement learning policy for local motion and manipulation, without a remote operator.
Zhang describes physical AI as mapping task objectives, environmental measurements, and robot state to motor commands. Because collecting enough end-to-end data is costly, he recommends dividing the system into motion generation and whole-body control, while balancing data savings against reduced generality. He describes motion generation as a VLA-like component that can use fast and slow decision systems.
Zhang says whole-body control for humanoids remains task-dependent, and a cross-task foundation model is still a research target. Policies can track references for every joint, selected end effectors, or simpler signals such as body velocity. Simpler references are easier to specify but leave more work for reinforcement learning; Zhang also describes a student-student method that nearly halves training time and a stability reward based on the LIP model.
Zhang proposes expanding manipulation data beyond real, synthetic, and internet sources with reinforcement learning and skill composition. In his account, rewards guide trial-and-error learning to create useful data, while composition combines learned skills into new ones. He identifies internet video pretraining and real-world reinforcement learning as key research directions, and cites the GVF-TAPE CoRL paper for video-pretrained policies and reducing hallucinations.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
LimX Dynamics R&D head Paul Lee discusses humanoid robotics, locomotion, hardware, teleoperation, and practical use cases with The Humanoid Hub at LimX Dynamics in Shenzhen.
LimX Dynamics’ main speaker describes the Shenzhen company’s cognitive operating system for its Oli humanoid and compares its approach with other companies’ efforts to make robots learn and act autonomously.
Kevin Qiu, product manager for the TRON series at LimX Dynamics, discusses its research robots, development tools, and field applications with Chris Matthieu of RealSense.
Jesse Liu of LimX Dynamics describes the company’s humanoid and modular robot platforms, its developer tools, and his view of China’s robotics ecosystem at MACHINA 2026 in France.
