Yunfei Bai

- Lead the Gemini coding agent research team, focusing on Gemini post-training for coding including web dev, mobile app dev, game dev, and long horizon SWE capabilities. - Lead the Gemini coding and agentic capabilities evaluation, and Gemini self-improvement evaluation.

Led GenAI Evaluation Platform team (35+ engineers), built and launched evaluation infrastructure and tools used for Gemini post-training, and for Google GenAI product launches including Gemini App, Search, Vertex AI, Workspace, etc.

- Led data, ML infrastructure and applied ML teams in Project Starline - the next generation of telepresence (https://blog.google/technology/research/project-starline-prototype/). - Established the effort and set technical strategy and roadmap for data collection, synthetic data generation, data infrastructure, ML infrastructure to support novel view synthesis, head tracking development, and delivered robust and performant ML models in products.

- Led technical direction across robot simulation, deep learning, software development, and robot autonomy at Everyday Robot project (https://everydayrobots.com) - Led the team to collaborate with Google Brain and DeepMind on a dozen research projects, including recent breakthrough - PaLM-SayCan, which is an AI research breakthrough that unifies natural language understanding with the physical capabilities of robots (https://ai.googleblog.com/2022/08/towards-helpful-robots-grounding.html) - Founded the simulation team for Everyday Robots Project, built and managed cross geolocation teams with software engineers, researchers, and technical artists - Built robot simulator and enabled robot learning at scale by running 240 million virtual robots in sim in 2021 - Co-Developed PyBullet - widely used in the robotics and AI community (https://pybullet.org/) - Demonstrated 100x data efficiency for robot reinforcement learning by using simulation and Sim2Real (https://ai.googleblog.com/2021/06/toward-generalized-sim-to-real-transfer.html)

Simulating Human Motion of Object Manipulation - Introduced a planning and optimization technique to synthesize continuous animation of human manipulating multiple objects and environment features concurrently at various locations in a constrained environment. - Introduced a physics-based simulation and control technique to synthesize interaction between cloth and rigid bodies for animating cloth manipulation. - Introduced a control technique for manipulating orientation of an object using robot hand with palm and fingers. - Introduced a sampling-based simulation method for solving inverse dynamics in human walking so that zero residual forces can be guaranteed.

Worked on object recognition in videos and images.

Developed artist-directed dynamics simulation technique for 2D animation.

Implemented history-free collision handling algorithms for cloth simulation as part of studio tool.