Why this research matters
Bringing intelligent agents from digital environments into the physical world is widely considered a key milestone toward artificial general intelligence (AGI). Rapid advances in deep learning and computer vision have given AI powerful perceptual capabilities. Multimodal large language models (MLLMs) have added capabilities resembling human cognition, reasoning and common-sense understanding. Turning these brain-like abilities into practical actions in the physical world—embodied AI—remains a frontier challenge.
By combining reinforcement learning with Vision-Language-Action models, we aim to bridge perception, cognition and control, enabling robot dogs and robotic arms to adapt autonomously to complex environments. As AI systems interact more closely with society, protecting privacy during training on large volumes of multimodal data becomes increasingly important. We therefore also investigate differential privacy to build embodied AI systems that are intelligent, secure and trustworthy.
Many scientific questions remain open. For example:
- How can efficient VLA models align visual perception, language instructions and robot actions end to end?
- How can reinforcement learning enable robust control and dexterous manipulation for quadruped robots and robotic arms in highly uncertain physical environments?
- How can we reduce or eliminate the Sim2Real gap between simulated training and real-world deployment?
- How can advanced computer vision improve fine-grained spatial perception in dynamic, complex three-dimensional environments?
- How can differential privacy prevent sensitive-data leakage when training large multimodal and embodied-control models?
- How can robot systems continually learn and improve from very few human demonstrations or real-world trials?
Developing capable, generalizable and secure embodied AI requires more researchers with potential and enthusiasm to join this field.
This emerging field is receiving attention from national strategies, local governments and society, and is viewed as the next wave of AI. At a broad level, AI with physical execution and multimodal perception could transform manufacturing and everyday life. At an individual level, skills in deep learning, reinforcement learning, VLA models, robot control and privacy-preserving computation can support a competitive career. Major internet companies, technology firms, and autonomous-driving and robotics companies are actively recruiting embodied-AI researchers and engineers, with a broad range of opportunities and attractive compensation.
