Embodied Intelligence and General‑Purpose Humanoid Robots: Insights from Wang He’s Ant T‑Space Talk

Assistant professor Wang He of Peking University and director of the Peking‑University‑Galaxy‑General Embodied Intelligence Lab introduced his dual role as CEO of the startup Galaxy General, a company developing embodied large‑model humanoid robots and recently completing a 700‑million‑RMB angel round.

He argued that successful embodied AI must consider commercialisation from day one, focusing on data‑rich scenarios, iterative improvement, and prioritising arm capabilities over legs because arms deliver higher immediate value in many tasks.

Wang identified three essential intelligences for embodied AI—spatial, interactive, and hardware—and positioned embodied intelligence as a necessary step toward artificial general intelligence, with applications in manufacturing, services, logistics, and household chores.

The talk highlighted the limitations of real‑world data collection (e.g., Tesla’s costly approach) and advocated large‑scale synthetic data; his team trained grasping and manipulation models on billions of synthetic samples, achieving strong generalisation on unseen objects.

Key technical contributions include the "Lefty" algorithm that lifts RGB images to high‑precision 3D geometry, enabling accurate grasp planning, and an end‑to‑end multimodal model (RGB‑D + language) that can follow natural‑language navigation commands in unseen environments.

Demonstrations covered flexible assembly line tasks, SPS sorting, autonomous retail restocking, garment handling, and multi‑modal perception, illustrating how the robots combine vision, language, and control to operate without hand‑coded behaviours.

Wang concluded that embodied large‑model robots, powered by massive synthetic datasets and integrated spatial‑interactive‑hardware intelligence, will soon serve a wide range of industries and homes, bridging the gap between research breakthroughs and real‑world deployment.