imit8

Humanoid robots that learn by watching.

Who we are

imit8 builds the perception and skill-extraction stack that lets a humanoid robot learn a manipulation task from a single short clip of a person doing it. We treat video as the training set and keep the modular pre-trained components inspectable — no end-to-end black box between the human demonstration and the robot's execution plan.

imit8 is a research venture of ShedInnovations LLC (California, USA), the same holding behind hpc-labs.ai. Founder: Frantz Lohier, PhD.

Research in preparation

Two companion papers are being prepared for arXiv release in the next few days. Together they argue for a training-light compositional approach to humanoid manipulation: compose pre-trained AI modules under geometric glue, then measure what a single human demo adds on top of a scene prior.

arXiv · upcoming

F.R.A.N.T.Z.: 3D Scene Reconstruction and Hand-Trajectory Extraction from Monocular Video via Training-Light Compositional AI

An open, modular, inspectable perception pipeline that composes MiDaS / DAV2 (depth), YOLO (detection), MediaPipe (hands), MOG2 (background) under geometric glue — per-frame optical-flow pose chain, two-bone analytic IK, COLMAP sparse-cloud reuse for gsplat initialisation. We measure the pipeline across three regimes: novel-view synthesis against COLMAP, pose quality against ground truth on five TUM fr3 dynamic sequences, and detector-generation modularity. We report both wins and catastrophic failures honestly, and scope ORB-SLAM3 / DROID-SLAM / MASt3R-SLAM baselines as the explicit next experiment.

Frantz Lohier, PhD · targeted for IROS 2027 · preview PDF (9 pp) · arXiv ID pending

arXiv · upcoming

What Does a Single Human Demonstration Add Over a Scene Prior? A Controlled Study

Controlled measurement of what a Layer-3 affordance descriptor extracted from one 5-second human-video clip adds over a no-demo hand-coded planner, under matched scaffolding across five MuJoCo Franka Panda tasks. Three default-sufficient control tasks (T1–T3): the demo adds nothing measurable. Two structural tasks: T4 (top-knob mug) exits the default regime; T5 (articulated-arc knob turn) is the paper's one clear positive demo-vs-default result (paired-McNemar p = 0.0026 strict / p = 3.6 × 10⁻⁵ relaxed). A short HOT3D fine-tune of DAV2 Metric brings Layer-1 wrist RMSE from 128.81 cm out-of-box to 13.71 cm pooled on held-out clips — a 4-9× in-distribution reduction.

Frantz Lohier, PhD · targeted for CoRL 2027 · preview PDF (14 pp) · arXiv ID pending

Preprints will link here once posted. Code is held back pending a patent review; reach out for the source data references used in the papers.