Currently ongoing research (not yet published)
We present a generative extension of HPSTM, enabling structured human motion to be sampled from noise using flow matching and anatomical constraints. Extensions to broader action spaces and physical robot deployment are left for future work.
Work in progress. View on arXiv
Leveraging a monocular RGB stream and a novel Transformer-based smoothing module (HPSTM), this end-to-end ESFP pipeline estimates 3D human pose via ROMP, enforces anatomical consistency through forward‑kinematics, and maps refined trajectories to a low‑cost 4‑DoF uArm in real time. By jointly predicting per‑joint uncertainty and applying dynamic filtering, it delivers smooth, reliable vision‑to‑robot imitation for desktop manipulation tasks.
Currently ongoing research (not yet published)
A PyTorch implementation for 3D human pose smoothing using a Transformer-based temporal encoder and manifold constraints (HPSTM) to ensure physically plausible poses. Inspired by SmoothNet and ManiPose.
Work in progress. View on arXiv
Leveraging pre-trained GANs and U‑Net, this framework combines a global anomaly detection module with iterative mask refinement under adversarial loss to accurately segment brain tumors on multi‑modal MRI. By incorporating synthetic image augmentation, it overcomes limited annotated data and achieves superior lesion‑wise Dice and HD95 performance on the BraTS benchmark, reducing reliance on fully labeled datasets for real‑world clinical use.