My research centers on perception and sensor fusion for autonomous systems, risk-sensitive locomotion and navigation in unstructured environments, embodied AI for assistive robots, and collaborative multi-agent systems.
Read more about my work on risk-sensitive locomotion and navigation, embodied scene understanding, generative occupancy for exploration, and multi-agent coordination under communication constraints. You can also see my talks, or my projects on their own page.
Research threads
This work uses distributional reinforcement learning, which models the whole distribution of outcomes rather than just the expected return, to learn navigation and locomotion policies for quadruped agents that account for risk directly.
A robot tidying a home has to answer a question that object recognition alone cannot: where does this object belong? A book could reasonably go on a shelf, a desk, or a nightstand, and what breaks the tie is the inhabitant's routine, which is almost never stated outright. ECHO is a large synthetic dataset for studying this. It pairs dense RGB-D scans of indoor scenes with recordings of an embodied human carrying everyday objects to where they belong, plus a one-sentence description of the routine that implies the destination, and poses contextual object placement as a benchmark task.
Papers
In review2026
3,805 human-annotated object-carrying episodes across 159 floors of 115 HM3D scenes. No single input modality is enough to predict where an object belongs.
Robots normally plan only over geometry they have directly measured, so they stall when they turn a corner or enter a new room. This thread trains diffusion models to predict the geometry that has not been seen yet, fuses those predictions into a live occupancy map without ever overwriting observed space, and uses the result to keep exploring instead of pausing.
Papers
IROS2024
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Predicts occluded and out-of-view geometry in real time from a single RGB-D camera, without ever overwriting observed space.
ICRA2025
IEEE International Conference on Robotics and Automation (ICRA)
73% faster runtime with minimal accuracy loss, extending prediction across the whole map rather than just around the robot.
When a team of robots explores somewhere the radio does not reach, the hard problem stops being perception and becomes deciding where and when to transmit. This work plans transmission locations from signal strength and payload size, so agents share what they have found without backtracking to meet.
Papers
In review2026
Across 480+ simulated cave environments, cut distance travelled by up to 58% and exploration time by up to 52%.
Talks
