Mind the Residual Gap: Probabilistic Downscaling under Real-World Bias

• Identifies residual target misspecification as a fundamental source of under-dispersion in mean–residual probabilistic downscaling under real-world bias.
• Proposes ReMatch, combining deterministic mean regression with residual diffusion and optimal-transport-based residual distribution matching.
• In publication.

Sparse-to-Field Reconstruction via Stochastic Neural Dynamic Mode Decomposition

• Proposes a probabilistic DMD–Neural ODE model that reconstructs continuous spatiotemporal fields from only 10% sparse observations with uncertainty quantification.
• Recovers interpretable Koopman modes and continuous-time eigenvalues, learns distributions over dynamics across realizations.
• Published in the Proceedings of L4DC 2026, project website

Single-Instance Sampling for Computationally Efficient and Accurate Real-Time Task Space MPPI Control

• Real-time MPPI controller that achieves high-frequency task-space manipulation by redesigning sampling and horizon structures.
• Establish that constant-perturbation rollouts remain theoretically compatible with MPPI weighting.
• Published in the Proceedings of IEEE Transactions on Robotics, supplementary video

Distilling Realizable Students from Unrealizable Teachers

• Policy distillation under asymmetric imitation learning setting.
• Propose two new IL/RL algorithms robust to state aliasing.
• Published in the Proceedings of IROS 2025, project website

Subspace-wise Hybrid RL for Articulated Object Manipulation

• Develop skills for operating equipment (e.g., valve, switch, gear lever…) at industrial sites with a manipulator.
• Learn skills while minimizing human-engineered features using reinforcement learning.

Whole-body motion planning of dual-arm mobile manipulator for compensating for door reaction force

• Address challenges of door traversal motion planning
• Unified framework for door traversal, from approaching, opening, passing through, and closing the door with dual-armed mobile manipulator
• Decision making for optimal contact point planning with RL.
• Presented in ICRA 2025 workshop

A reinforcement learning approach to dynamic trajectory optimization with consideration of imbalanced sub-goals in self-driving vehicles

• Challenge to skewed sub-goal distribution for goal-conditioned RL controller.
• Enable adaptive sub-goal planning and efficient reward learning via MPC-synchronized rewards.
• Published in the Proceedings of Applied Sciences

Reinforcement Learning for Autonomous Vehicle using MPC in Highway Situation.

• Addresses the challenge of reward shaping for continuous RL controllers by using MPC reference.
• Published in the Proceedings of ICEIC 2022