Research

DexHOT: Learning Hand–Object Configuration Transfer for Dexterous Manipulation

In submission 2026

We address dexterous manipulation through hand–object configuration transfer, jointly specifying hand articulation and object pose relative to the wrist. A single relation-conditioned policy learns reusable transfers from synthetic configuration pairs and composes them into long-horizon behaviors, with evaluation on unseen objects and a real dexterous hand.

PaperWebsite Code soon

OpenDexGrasp: Open-Vocabulary Task-Oriented Dexterous Grasping

[CoRL 2026] Conference on Robot Learning 2026 Oral (1 of 21 oral presentations among 687 accepted papers, selected from 2,094 submissions; 3.1% of accepted papers · 1% of submissions)

We address open-vocabulary task-oriented dexterous grasping by integrating language instructions, multi-view observations, and object geometry in a shared perception–action representation. Joint affordance grounding and grasp generation are trained with synthetic data and human demonstrations to produce task-consistent grasps that generalize to unseen object categories.

Paper Website Code soon

RefineBridge: Generative Bridge Models Improve Financial Forecasting by Foundation Models

[ICASSP 2026] International Conference on Acoustics, Speech and Signal Processing 2026 Oral (1 of 866 oral presentations among 4,589 accepted papers, selected from 11,200 submissions; 19% of accepted papers · 8% of submissions)

We address financial forecasting with a context-conditioned Schrödinger bridge that refines predictions from frozen time-series foundation models. The learned stochastic transport progressively improves initial forecasts using historical observations, with accuracy gains across multiple financial benchmarks and prediction horizons.

Paper Oral Slides Code soon