Robotics & AI — I build research systems that ship as real libraries: visuo-tactile perception, 3D Gaussian splatting, and tools that make AI verifiable instead of plausible.
| what it does | the number that matters | |
|---|---|---|
🧩 splatreg   |
Register Gaussian splats — align & merge two 3DGS scans into one SE(3)/Sim(3) frame, recover scale, dedupe the overlap. CLI + pure-PyTorch API. Docs. | 91.5% official 3DMatch recall · splat-merge Chamfer 10.3 → 2.0 mm vs naive concat · photometric refine 5° → 0.36° where geometry under-constrains |
📐 mathlas    |
Airtight math an AI uses over MCP — 3.7M-theorem search, PSLQ constant ID, OEIS, real Lean kernel checks, applicability checklists. No LLM inside, no API key. claude mcp add mathlas -- uvx mathlas-mcp |
Beats TheoremSearch on its own 110 human queries (59.1 vs 45.0 Hit@20) · mathlib formal search Hit@5 0.96 · zero false positives across every verification tier |
⚡ HiCache++   |
Training-free diffusion acceleration by exponential (Dynamic-Mode-Decomposition / Prony) feature forecasting — a drop-in basis upgrade to TaylorSeer/HiCache, plus a holdout auto mode that can't make things worse. |
Exact on the feature-ODE class — error stays flat in skip horizon where polynomial bases diverge (4.7e-9 vs 6.5 at H=8) · auto detects basis misfit 120/120 |
🛣️ CERT-FLOW   |
Certified route planning under drifting costs: every replanning round emits a conformal certificate LB ≤ OPT ≤ UB on the optimal route, directs paid sensing at the edges that shrink the certified gap fastest, and proof-gates fast preprocessing. | Coverage at or above the claimed level on every reported condition incl. replayed METR-LA / PEMS-BAY traffic · 223 tests · 16 reproduction pipelines · 7 theorems incl. an impossibility result |
⚙️ The 13-adapter acceleration family — HiCache/HiCache++ deployed on real 3D generators
| model | Hermite (HiCache) | DMD (HiCache++) | headline |
|---|---|---|---|
| TRELLIS v1 | faster-trellis | faster-trellis-plus-plus | 2.85× at equal-or-better F-score vs Fast-TRELLIS |
| TRELLIS.2-4B | hermit-trellis2 · fast-trellis2 | hermit-trellis2-plus-plus | carved-hybrid sampler |
| Hunyuan3D-2 mini | hunyuan2-plus | hunyuan2-plus-plus | exactly lossless at interval-5 (F 0.794 = 0.794) |
| Hunyuan3D-2.1 | hunyuan2.1-plus | hunyuan2.1-plus-plus | DMD leads +0.13 F at i5, +0.24 at i6 |
| SAM 3D Objects | sam3d-plus | sam3d-plus-plus | geometry-lossless (F1 = 1.000) to interval-6 at 1.56× |
| Fast-SAM3D | fastsam3d-plus | fastsam3d-plus-plus | stacked accelerations |
🫳 GaussianFeels — real-time visuo-tactile 3D-Gaussian SLAM for in-hand manipulation. One object-centric Gaussian map serves tracking, reconstruction, rendering, and manipulation-facing geometry: 0.83 mm ADD-S in simulation / 3.37 mm on real hardware, ~7.6× faster than the neural-field baseline, 94% sim-to-real reconstruction retention. M.S. thesis (SNU Soft Robotics & Bionics Lab) — release upcoming.