Philip (Yizhou) Huang

I am a Ph.D. student in the Robotics Institute at Carnegie Mellon University, advised by Prof. Jiaoyang Li, where I work on multi-robot task and motion planning. This summer, I am a Research Scientist Intern at NVIDIA's Seattle Robotics Lab, developing GPU-accelerated planning for multi-arm systems.

My recent work focuses on making multi-robot-arm systems useful in the physical world by closely combining accelerated planning systems with learning — for example, generating LEGO designs and assembling them end-to-end with a bimanual system.

Before CMU, I completed my M.Sc. in Computer Science at the University of Toronto, co-advised by Prof. Florian Shkurti and Prof. Tim Barfoot, and my B.A.Sc. in Engineering Science, also at the University of Toronto. I have been fortunate to learn from many wonderful mentors along the way, including Prof. Fabio Ramos (NVIDIA), Prof. Angela Schoellig, and Prof. Amir Degani.

Email  /  CV  /  Resume  /  Google Scholar  /  LinkedIn  /  Twitter  /  Github

My research builds planning and learning systems for teams of robot arms to work together on long-horizon, real-world tasks. The throughline across all three thrusts is coordinating multiple robots:

• Accelerated, high-quality motion planning that is both fast enough and good enough for the real world.
• Integrated task and motion planning with learned manipulation skills for multi-robot assembly.
• End-to-end systems that turn high-level goals (e.g., a natural-language prompt) into coordinated physical execution.

To tackle these problems, I draw inspiration from a range of areas, including multi-agent path finding, generative models, task and motion planning, and robot learning.

If you have any questions / want to collaborate, feel free to reach out and send me an email! I am very excited to talk with more people and learn about your work!

The first generative LEGO assembly framework that designs and construct product from natural language prompts

Accelerating multi-robot collision checking and planning 10x-100x through CPU SIMD vectorization.

Dataset and evaluation of multi-robot shortcutting, a important yet under-investigated post-processing step in multi-robot arm motion planning.

An asynchronous planning and execution framework designed to safely and efficiently coordinate multiple robots arms to achieve cooperative assembly.

Leveraging parallel and differentiable simulation to efficiently search for multiple diverse plans with gradient-based variational inference.

Field testing and system descriptions of our GPS-, vision-, and sonar-enabled autonomous vessel for water-quality monitoring in freshwater lakes.

A robust route-planning algorithm uses satellite images as a corase map to plan water sampling routes for autonomous surface vessels (ASV) given environmental disturbances.

Task-conditioned hypernetworks can be used to continually adapt to varying environment dynamics in lifelong model-based reinforcement learning, with a fixed-size replay buffer.

System design and development of the winning self-driving car in the AutoDrive Challenge, as well as lessons learned.

A skill graph representation that unifies semantic planning, execution grounding, data collection, and iterative skill refinement for scalable and adaptive robotic assembly.

In continual learning, self-tuning network (STN) is a memory-efficient and easy-to-implement hypernetworks architecture with strong performance on many benchmarks.

RL can be applied to the Minimum Latency Problem by using a graph attention network to encode stochastic policy for constructively building partial paths, yielding solutions which are comparable to state-of-the-art, hand-engineered methods.

Re-implemented the SIGGRAPH 96 paper from David Baraff that introduced a linear-time sparse solver with Lagrangian multiplers in MATLAB. Verified that the time complexity of our re-implemented sparse solver is linear with serial chains and trees.

We measured the quantitative performance of recent language models for text style transfer, using three metrics and third-party models for fair comparison.

Re-implementation of the paper PIXOR: Real-time 3D object Detection from Point Clouds using PyTorch.

Indoor localization and navigation of a Crazyflie nano-quadcopter with a RGB-D camera for pollinating sunflowers.

Designing software framework and simulations for flying a swarm of 9 Crazyflie nano-quadcopters indoors.

This template is from Jon Barron's website. Here is the source code