Forenly AI Lab is our open, reproducible R&D track toward Scene→Sim (Splat-to-Sim): turning a real-world scene into a sim-ready digital twin for humanoid robot training.
Each study stress-tests a frontier 3D-reconstruction or robot-learning method on a public benchmark — and every result is runnable on a free Kaggle T4, so anyone can verify it.
Featured — Study 1: PPISP ablation (OFF vs ON). Running NVIDIA's PPISP on a multi-camera capture, at 30k iterations the no-PPISP baseline collapses into a hazy blur while PPISP stays sharp and color-correct: +9.5 dB PSNR.
Pipeline: capture → COLMAP poses → 3DGRUT + PPISP (visual) → 2DGS/GOF mesh (collision) → MuJoCo (sim-ready).
Live, with reproducible code: forenly.ai/lab
0
11