Embodied AI — reinforcement learning and active perception.

• Designed a two-phase MDP framework that separates goal inference from goal-conditioned planning; recovers the optimal strategy when observations are informative.
• Built a feature-based heuristic model (linear regression) that captures most human behavior and isolates the contribution of goal inference.
• Introduced a mixed-strategy model (logistic regression over weighted combinations) that explains residual variance beyond the optimal strategy.
• Vectorized single-goal heuristic features to extend to a multi-goal setting without retraining.

Computational cognitive science — planning, decision-making, memory and habits, brain-computer interfaces.

• Built behavioral-computational models of partner selection under ability vs warmth trade-offs using reinforcement-learning frameworks.
• Parallelized parameter-search for fitting candidate models to behavior; improved parameter-precision from 0.1 to 0.001.
• Built end-to-end pipeline (preprocessing, training, evaluation, inference) for reconstructing Chinese semantic representation from EEG/fMRI signals using deep learning.
• Used Distributed Data Parallel (DDP) on a multi-node multi-GPU cluster; compressed model training time from 4 days to 1.5 days on identical data and model.

• Designed a two-phase MDP that separates goal inference from goal-conditioned planning.
• Built a feature-based heuristic baseline (linear regression) for human behavior.
• Introduced a mixed-strategy model (logistic regression) explaining residual variance.
• Vectorized single-goal features for multi-goal extension.

• Parallelized parameter-search for model fitting; parameter precision 0.1 → 0.001.

• {"End-to-end pipeline" => "preprocessing, training, evaluation, inference."}
• DDP-based multi-node multi-GPU training; 4 days → 1.5 days on identical model & data.

• {"5× faster teleoperation" => "motion-rate factor 0.01 → 0.05 while keeping trajectories smooth."}
• ~90% fewer interruptions during long sessions via buffering, deadband tuning, reconnection protocol.
• End-to-end gripper open/close control in the recorded trajectory stream.

• Designed and implemented the Franka Research 3 driver in the hardware abstraction layer (HAL).
• Lets LLM agents take over Franka Research 3 without per-user environment setup.

• End-to-end pipeline for world-model data on Franka Research 3.

• Active perception research (Matlab).

• {"Course project" => "intelligent analysis of music scores / tabs (Python + Web)."}