Meng Song

I completed my PhD at UC San Diego, where I was advised by Prof. Manmohan Chandraker.

Prior to that, I obtained my Master's degree at the Robotics Institute, Carnegie Mellon University, working with Prof. Abhinav Gupta and Dr. Daniel Huber.

Email  /  CV  /  Scholar  /  Twitter  /  Github  /  LinkedIn

Research

My research is motivated by the goal of developing a mathematical construct of the intelligent agent from first principles. My recent work has primarily focused on answering the question "What is a good representation of states and goals in decision-making problems?" I explored this problem under three different learning paradigms: reinforcement learning, imitation learning, and unsupervised learning.

PPO and BC generalize differently in visual navigation: BC imitates successful trajectories, while PPO combinatorially stitches together past experiences, including failures, to solve new tasks and achieve stronger generalization.

Learning through interaction is a foundational principle in both human and animal learning. In a broad sense, intelligent agents can be formulated as goal-directed systems interacting with an uncertain environment.

Despite the generality of this definition, a key challenge in computationally grounding it lies in how to effectively set up and represent goals and purposes. This dissertation explores this question through the lens of various machine learning paradigms.

A novel probabilistic world model trained with contrastive learning. The learned latent space enables subgoal discovery, asymmetric transition modeling, and supports highly efficient planning without requiring any inference-time search.

A novel prompting method that enables interpretable zero-shot generalization in unseen robotics tasks without requiring demonstrations and surpasses few-shot baselines.

RL-based prompt optimization approach outperforms a wide range of finetuning and prompting baselines on text classification and style transfer tasks.

A PPO agent learns physical concepts such as mass and friction by actively interacting with the environment and mastering everyday skills, instead of passively observing physical processes.

A novel framework for creating large-scale photorealistic datasets of indoor scenes, enabling broad applications in inverse rendering, scene understanding, and robotics.

A novel 6-DoF grasp detection method using a single-shot grasp proposal network trained on automatically generated synthetic data, significantly outperforming state-of-the-art methods.

A novel geometric approach that automatically segments and parses a complex network of thin structures from low-quality 3D point clouds.

A novel geometric approach for extracting and modeling tree trunk landmarks from noisy 3D point clouds of cluttered forested environments.

A novel geometric approach for extracting and modeling tree trunk landmarks from noisy 3D point clouds of cluttered forested environments.

Thoughts and Talks

Teaching

Academic Service

Template stolen from Jon Barron.