Felix Yanwei Wang

I am on the job market for research scientist positions starting in spring 2025.

I am a final-year EECS PhD student at MIT CSAIL, working with Julie Shah on robot learning, specifically, inference-time model customization through human interactions. I am also a fellow at MIT's Work of the Future group focusing on generative AI.

Before MIT, I did my MS in robotics at Northwestern University and researched with Todd Murphey and Mitra Hartmann on active sensing. I completed my undergraduate degree in physics at Middlebury College with Richard Wolfson.

Outside research, I enjoy theatre and backpacking. I thru-hiked PCT in 2019.

Email / Google Scholar / LinkedIn / CV / X

• [Dec 2024] Invited talk at MIT EI Seminar - "Inference-Time Policy Customization through Interactive Task Specification"
• [Oct 2024] Invited talk at University of New Hampshire - "Inference-Time Policy Alignment through Human Interactions"
• [Jul 2024] Organized GenAI-HRI workshop at RSS 2024.
• [Mar 2024] Invited talk at Brown University Robotics Seminar - "Interactive Task and Motion Imitation"
• [Mar 2024] Invited talk at University of Utah - "Interactive Task and Motion Imitation"
• [Jan 2024] Check out our Generative AI newsletter from MIT's work of the future group.
• [Oct 2023] Temporal logic imitation was awarded the Best Student Paper at IROS 2023 Workshop.
• [Jan 2023] PBS News coverage of our human robot interaction exhibition at MIT museum.

Imagine driving with Google Maps, where multiple routes unfold before you. As you take turns and change plans, it adapts instantly recalculating to match your shifting preferences. My research goal is to bring this level of interactivity to embodied AI, empowering humans to steer pre-trained policies at inference-time.

We propose Inference-Time Policy Steering (ITPS), a framework that leverages human interactions to zero-shot adapt pre-trained generative policies for downstream tasks without any additional data collection or fine-tuning.

We present the Versatile Demonstration Interface (VDI), a collaborative robot tool designed to enable seamless transitions between data collection modes—teleoperation, kinesthetic teaching, and natural demonstrations—without the need for additional environmental instrumentation.

This work learns grounding classifiers for LLM planning. Our end-to-end explanation-based network is trained to differentiate successful demonstrations from failing counterfactuals and as a by-product learns classifiers that ground continuous states into discrete manipulation mode families without dense labeling.

We present a continuous motion imitation method that can provably satisfy any discrete plan specified by a Linear Temporal Logic (LTL) formula. Consequently, the imitator is robust to both task- and motion-level disturbances and guaranteed to achieve task success.

We prompt large language models to augment a domain-specific dataset to train specialized small language models that outperform the general-purpose LLM.

By learning how to pan, tilt and zoom its camera to focus on random crops of a noise image, an embodied agent can pick up navigation skills in realistically simulated environments.

We installed an interactive exhibition at MIT Museum that allows non-robot-experts to teach a robot an inspection task using demonstrations. The robustness and compliance of the learned motion policy enables visitors (including kids) to physically perturb the system safely 24/7 without losing a success gaurantee.

This research project studies the application of two information gain methods in a search problem: 1. Infotaxis - Search guided by entropy minimization 2. Ergodic exploration - Search guided by proportional coverage.

This research project studies how active sensing, i.e., choosing what data to collect, can improve data efficiency for decision-making under uncertainty. Inspired by the active whisking behavior of rats, we use simulated rat whisker sensory signals as a model for spatial-temporal data to learn policies that first collect observations and then classify object shapes.

This research project studies multi-agent distributed algorithms concerning coordination, segregation, and locomotion, with a hardware implementation of robust localization with cheap sensors on a low-cost underactuated system.

For my undergraduate physics thesis at Middlebury College, I did a computational fluid dynamics (CFD) simulation of Festo's Dual-Wing generator using COMSOL CFD package.

Miscellaneous class projects.

We compare the effectiveness of relaxation methods for warm starting trajectory optimization through soft contact.

Template