Deep Learning

Whole-body control of G1 and Spot robots for loco-manipulation, using GAIT scene reasoning.

Leverage online semantic scene graph construction to build a persistent memory for vision-language-action models, enabling open-vocabulary navigation.

Training foundational models to reproduce embryogenesis — predicting both cell-level gene expression dynamics and 3D morphogenesis.

End-to-end robot learning pipeline for language-guided drone navigation — from MPC expert demonstrations to sim-to-real deployment via behavioral cloning and DAgger.

A multi-agent LLM system that automates gene panel selection for large-scale spatial transcriptomics.

A reinforcement-learning framework for adaptive gene panel selection that preserves single-cell structure while minimizing panel size.

A surrogate modeling and optimization framework for pore-scale permeability inference in supersonic parachute fabrics.