Pingchuan Ma (Chinese: 馬平川) is an AI Research Scientist at Meta's Superintelligence Labs. His research focuses on the intersection of machine learning, computer graphics, and robotics, with significant contributions in areas such as differentiable simulation, physics-augmented generative models, and multimodal learning. ^{[1] [2]}

Education

Ma earned a Bachelor of Engineering degree in Software Engineering from Nankai University in Tianjin, China, attending from 2015 to 2019. Following his undergraduate studies, he moved to the United States to pursue graduate work at the Massachusetts Institute of Technology (MIT). At MIT, he joined the Computer Science and Artificial Intelligence Laboratory (CSAIL), where he was advised by Professor Wojciech Matusik. He completed a Master of Science (S.M.) in Computer Science in February 2023 and successfully defended his Ph.D. thesis in Computer Science in February 2025. ^{[1] [3] [6]}

Career

In July 2025, Ma joined Meta as an AI Research Scientist in its newly formed Superintelligence Labs, a team assembled to advance foundational AI research. His appointment was part of a significant talent acquisition effort by Meta, which recruited numerous researchers from other leading AI organizations.

Prior to his role at Meta, Ma was a Member of Technical Staff at OpenAI from February to July 2025, where his work centered on multimodal models and post-training techniques. Throughout his doctoral studies at MIT CSAIL from 2019 to 2025, he served as a Research Assistant. Ma's professional experience is supplemented by several research internships at prominent technology labs. He was an intern at the NVIDIA Seattle Robotics Lab from May to December 2024, working with Professor Dieter Fox. In 2021, he interned at the MIT-IBM Watson AI Lab under the guidance of Professor Chuang Gan. His earliest industry internship was at SenseTime Research from May 2018 to February 2019. During his time at MIT, he also served as a teaching assistant for the 6.807/6.839 Advanced Computer Graphics course in the fall of 2022. His research career began with an assistantship at Nankai University, which lasted from April 2016 to June 2019.

Ma's research integrates concepts from machine learning, computer graphics, and robotics. A central theme in his work is the development and application of differentiable physics simulation, which enables the use of gradient-based optimization methods to solve complex physical inverse problems. This approach has been applied to challenges in soft robotics, fluid dynamics, computational design, and system identification. His work also explores the creation of physics-augmented generative models, which combine the expressive power of deep learning with the constraints of physical laws to produce more realistic and controllable outputs. Other key areas of his research include multimodal learning for vision and language, the development of efficient AI systems, and the application of AI to scientific discovery.

He has co-authored numerous papers presented at major AI and computer graphics conferences, including NeurIPS, ICML, ICLR, SIGGRAPH, and ICRA.

A selection of his notable publications includes:

• KAN 2.0: Kolmogorov-Arnold Networks Meet Science (2024): Co-authored with Ziming Liu, Yixuan Wang, Wojciech Matusik, and Max Tegmark, this paper proposes an evolution of Kolmogorov-Arnold Networks (KANs) tailored for scientific discovery by incorporating physical priors and symbolic components into the model architecture.
• LLM and Simulation as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Discovery (2024): Presented at ICML, this work introduces a framework where Large Language Models (LLMs) act as high-level optimizers that propose and refine hypotheses, while differentiable simulations act as low-level optimizers that validate these hypotheses against physical laws.
• DiffuseBot: Breeding Soft Robots with Physics-Augmented Generative Diffusion Models (2023): An oral presentation at NeurIPS, this paper presents a method for co-designing the morphology and control of soft robots using a diffusion model that is guided by a physics-based simulation, enabling the generation of diverse and high-performing robotic structures.
• Learning Neural Constitutive Laws from Motion Observations for Generalizable PDE Dynamics (2023): Published at ICML, this research introduces a method for learning the underlying material properties (constitutive laws) of physical systems directly from video observations, using a neural network to model the relationship between stress and strain within a partial differential equation (PDE) framework.
• SoftZoo: A Soft Robot Co-design Benchmark For Locomotion In Diverse Environments (2023): Presented at ICLR, this work introduces a comprehensive benchmark for designing and evaluating soft robots, providing a standardized environment for co-optimizing their shape and control policies across various terrains.
• RISP: Rendering-Invariant State Predictor with Differentiable Simulation and Rendering for Cross-Domain Parameter Estimation (2022): An oral presentation at ICLR, this paper proposes a framework for creating "digital twins" of real-world objects by estimating their physical properties from video. It integrates differentiable simulation and rendering to bridge the gap between the visual appearance of an object and its underlying physical dynamics.
• DiffAqua: A Differentiable Computational Design Pipeline for Soft Underwater Swimmers with Shape Interpolation (2021): Published in ACM Transactions on Graphics (SIGGRAPH), this paper details a fully differentiable pipeline for designing efficient soft-bodied underwater swimmers. The system optimizes the shape and actuation of a robot to maximize swimming speed.
• Efficient Continuous Pareto Exploration in Multi-Task Learning (2020): Presented at ICML, this work introduces an algorithm for efficiently mapping the Pareto front in multi-task learning problems, allowing for a more complete understanding of the trade-offs between different objectives.
• Fluid Directed Rigid Body Control using Deep Reinforcement Learning (2018): Published in ACM Transactions on Graphics (SIGGRAPH), this early work demonstrates how deep reinforcement learning can be used to control the motion of rigid bodies within a fluid simulation, enabling objects to navigate complex fluid environments autonomously.

The above list represents a selection of Ma's contributions to the fields of AI, robotics, and computer graphics. ^{[1] [2] [3] [4] [5] [6] [7]}