He Zhang
He Zhang is a senior researcher at Tencent RoboticsX Lab (Agent Learning Center) whose work focuses on motion control of simulated characters and dexterous manipulation using physics-based, data-driven, and learning-based techniques. He received his Ph.D. in Computer Science from the University of Edinburgh in 2022 under the supervision of Prof. Taku Komura. He serves as a corresponding author on the Hy-Embodied-0.5-VLA system, an end-to-end Vision-Language-Action robot learning stack spanning data collection, model design, RL post-training, and real-world deployment.
“Tencent has shipped a complete, open-source robot learning stack — data hardware, model architecture, RL fine-tuning, and deployment runtime — that achieves 99% task success on sub-millimeter manipulation tasks and cross-embodiment transfer without ever collecting data on the target robot.”
Source→“Zhang is also a co-author on the FlowPRO paper ([47]) and appears on the Universal Pose Pretraining paper ([22]), indicating a sustained research program at Tencent on VLA pre-training and RL post-training.”
Source→“A new fine-tuning framework that uses human-guided failure correction — without designing reward functions — to push robot manipulation policies from 'good demo performance' to near-deployment-grade reliability, achieving 92–99% success rates on hard bimanual tasks.”
Source→“He Zhang — Tencent Robotics X / Futian Laboratory (Corresponding Author). The corresponding author and likely the principal researcher driving this work.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.