imit8

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.: A Modular Monocular-Video Perception Pipeline with Drop-In Pose, Depth, and Detector Modules and a Head-to-Head Benchmark on First-Person Egocentric Video

An open modular monocular-video perception pipeline — pre-trained MiDaS / DAV2 depth, YOLO detection, MediaPipe hands, MOG2 background, essential-matrix pose chain or ORB-SLAM3 drop-in, two-bone IK, COLMAP sparse-cloud reuse for 3DGS initialisation — plus a reproducible head-to-head benchmark across the swap points the typed PoseStage / DepthStage / detector interfaces enable. On 10 held-out HOT3D first-person egocentric clips, ORB-SLAM3 wins ATE 9/9 against the essential-matrix chain in both keyframe mode (4–300× margins) and per-frame mode via a patched SaveTrajectoryEuRoC that accepts monocular (2.5–116× margins at 61–98% per-frame coverage). The chain is best read as a 100%-coverage baseline for the pre-initialisation and tracking-lost regime, not a competitor on absolute pose; the paper's contribution is the modular architecture plus the measurement infrastructure that forced this comparison into the open.

Frantz Lohier, PhD · targeted for IROS 2027 · preview PDF (12 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.

Policy Studio early access

imit8-studio v0.1.0 — the Phase 1 retargeting layer, perception pipeline scaffolding, and evaluation harness from the F.R.A.N.T.Z. Policy Studio product. Shipped under an early-access licence to invited design partners.

Downloads below are IP-restricted at the edge; if you reach a 403 Forbidden, email hello@imit8.ai with the address you’re connecting from.

What ships: the typed PoseStage/DepthStage/Retargeter protocols; the ParallelJawRetargeter for xArm 6; Savitzky-Golay + SLERP trajectory smoothing; the contact-point-click calibration CLI; the arc-descriptor fit for the knob-turn task family; the three-metric evaluation harness; 36 unit + end-to-end tests that pass without a torch install.

What doesn’t: real-arm xArm SDK readback, Nuitka-compiled runtime, encrypted Layer-3 weights, phone-home licence server, non-xArm embodiments, non-arc task families. See CHANGELOG.md inside the tarball for the v0.2.0+ roadmap.