We present a novel LLM-powered framework with DAG architecture for adaptive communication in human-robot collaboration. A dual-module system—Coordinator (strategic) and Manager (tactical)—enables dynamic transition between passive and active interaction modes based on task complexity.

We introduce a fine-grained benchmark with 22 layouts to assess proactive adaptability and temporal responsiveness, and the MonTA framework using a hierarchical LLM approach (fast monitor + slow adapters). The system achieves real-time adaptation and 75% reasonability in human expert evaluations, with significant gains in low-teaming-fluency scenarios.

Submitted to Science Robotics

We present a scout-rover cooperation framework for online terrain strength mapping and traversal risk estimation. The approach was validated in planetary-analog environments at the NASA Ames Lunar Simulant Testbed and White Sands National Park, through multi-institutional collaboration (USC, UPenn, NASA Marshall, NASA Ames).

Submitted to Science Advances

We present LASSIE, a framework in which legged robots treat every step as an experiment—using proprioceptive sensing during locomotion to advance mobility and data collection in planetary-analog exploration.

We propose inverse resistive force theory (I-RFT) to learn granular terrain properties through robot-terrain physical interactions, enabling data-driven characterization of deformable substrates.

PSANE learns GP-based traversability from leg–terrain interactions with uncertainty-aware safe-set certification, and combines multi-objective frontier subgoal selection with a reactive controller for safe navigation on deformable terrain.

We show that bio-inspired tail oscillation enables fast crawling on deformable granular terrains through effective locomotion strategies that leverage body-tail interactions.

We introduce a proprioceptive sensing method to estimate mud properties from actuator signals and an RFT-based model of substrate strength. A flipper-driven robot achieves adaptive locomotion on varying muddy terrains, with real-time adaptation that prevents failures in complex, deformable environments.

We present a force modeling approach measuring shear and extraction forces and a mudskipper-inspired robot for terrestrial locomotion on muddy substrates. We identify two failure mechanisms (slippage vs. entrapment) and non-monotonic performance with respect to water content, with optimal performance at 25–26%. Adaptation strategies increase robot speed by more than 200%.

🏆 Best Paper Finalist (5%) at HRI '24.

We develop models to understand expert sampling strategies in scientific exploration and create web interfaces for interactive exploration.

We develop frameworks for understanding human sampling objectives in scientific exploration to enable robot-assisted scientific decision-making. Presented at ICRA Workshop (2022) and American Geophysical Union Fall Meeting (2021).