Wang He on Embodied AI Research, Data and Commercialization

Zhang Xiaojun interviews Peking University assistant professor and Galbot founder Wang He about the research history of embodied AI, robot data and commercialization challenges.

Wang distinguishes language, vision and intelligence. He describes intelligence as the ability to respond to an environment according to circumstances, and sees language as a major development in human intelligence rather than its only possible form.

He recalls changes in how research communities used terms such as embodied agent and embodied AI, emphasizing that meanings differ across research traditions.

He contrasts traditional industrial robotics’ emphasis on control precision, repeatability and reliability with embodied AI research on connecting perception and action.

Using examples from simple actions to complex human decisions, he says interaction with an environment can span different dimensions and timescales. This is a conceptual framework.

Looking back at research history, he says navigation became an important task: an agent moves and receives a new view, creating a perception-action loop.

Compared with moving a camera, grasping and manipulation also change objects in the environment and introduce more complicated physical contact.

He recalls research communities advancing visual navigation through tasks, datasets and challenges, while other teams studied grasping and related problems in parallel.

Discussing his PhD, he recalls early work on human video and object interaction as well as difficulty with an initial research direction, changing projects and waiting for results.

He describes an experimental cup demonstration in which a system used interaction information learned from video to move into position. It illustrates a research idea, not a general-purpose home robot.

He reflects on moving from physics modeling to machine learning, where models, experimental data and research methods changed and he had to learn new team skills.

He says programming and research were difficult early in his PhD before he gradually found questions better suited to his long-term interests.

He introduces object-pose estimation: a system must infer an object’s position and orientation relative to a reference frame. Generalizing across objects adds data collection and labeling challenges.

To address limited data, he recalls capturing indoor tabletop scenes and rendering digital objects into them to construct mixed data.

Around his return to China, he considered different research paths and continued working on robots for homes and other practical environments.

He says robot software and hardware influence each other: unreliable hardware limits algorithm deployment, while insufficient algorithms limit hardware’s practical value, so both must improve iteratively.

He emphasizes real tasks and marginal costs: if every new task requires substantial data collection from scratch, the system may be difficult to scale.

On Galbot’s wheeled robots, he says customers care about reliable work, not just appearance. Claims about operating duration and performance are the guest’s company account.

He discusses data differences between vision-language and language models and says robot action data is harder to collect. Model performance should be judged by task-specific evaluation.

He warns of two commercialization difficulties: remaining outside practical settings for too long and failing to reduce marginal cost as tasks are added.

Wang summarizes his aim as “productivity as the product”: robots need to perform valuable work reliably in specific settings, not merely demonstrate capabilities.

On data, he says his team used substantial synthetic data and a small proportion of real data. This was his team’s approach at the time, not an optimal ratio for all systems.

He believes some vision and grasping tasks can transfer from simulation to reality, but results depend on scenes, materials and validation; synthetic data does not automatically eliminate the gap.

He also identifies simulation difficulties for tasks such as tying shoelaces, involving deformable objects and complex contact. Performance must be tested under specific conditions.

Discussing funding and demonstrations, he calls for clearly identifying teleoperation and autonomy so audiences understand what a robot actually did.

He says a high-quality simulation data pipeline needs sustained investment; a proof of concept alone does not establish an efficient operating system.

He discusses potential robot uses in the context of aging populations and labor demand. Estimates of demand and the pace of technical development require independent verification.

On humanoid robots, he says understanding of whole-body control and biological movement remains limited. Embodied AI has long-term research questions, and recent demos should not be treated as complete solutions.