Products and research shipping out of the Lab
Technical writing from the Realarity team. Engineering decisions behind our products, experiments we ran, and our perspective on where Physical AI is going.
June 2026 · Engineering
From scan to GPU-accelerated RL: causal contracts that train
Authored causal contracts go silent in vectorized RL training. The Realarity Contract Engine turns them into GPU tensor rules — 10,368× faster at 65k environments, effect|cause 1.00 from iteration 1.
June 2026 · Engineering
Sim-to-real is a calibration problem
Sim-to-real is a calibration problem, not a generation problem. The beliefs about friction and compliance live in a few hundred bytes of JSON. This is how Veron patches them.
May 2026 · Research
Cross-robot transfer: same drawer, Franka YES, Allegro NO
Most asset formats answer feasibility with a confidence number. We answer it with a verdict and a reason.
May 2026 · Perspectives
What we actually need from REArtGS-style articulated reconstruction
The unmet need is not a better reconstruction algorithm. It is a benchmark. This post argues for one.
April 2026 · Research
Three trials we ran, and what they actually told us
Three arXiv papers, the smallest faithful version of each, against our pipeline in an afternoon. Two of them surprised us.
April 2026 · Engineering
Capturing failure as a first-class artifact
Treating failure as a presence rather than an absence cut our refinement loop from twelve iterations to two.
February 2026 · Engineering
Why your "kinematic body" breaks Isaac Lab articulations
Four small USD authoring details that each cost us a day. None of them are documented anywhere a search engine would find.
February 2026 · Engineering
The 16-test gauntlet every asset has to pass
Sixteen deterministic tests, each added because a bundle without it once shipped broken. This is the gauntlet.
January 2026 · Engineering
The kitchen lighter: contact-rich physics built from 90 photographs
The lighter, the drawer, and the game controller through the full Realm pipeline. Contact-rich physics, calibrated values, no placeholders.
January 2026 · Engineering
What a neural object actually contains
A walk through the neural object schema — the typed Pydantic bundle that ships every Realm asset.
December 2025 · Perspectives
Physical AI is scaling effort, not knowledge
Physical AI is scaling effort, not knowledge. Every team starts from zero. We are building the memory layer that makes it compound.
December 2025 · Research
Robotics motion planning from clinical data
A pipeline that turns raw iPad motion capture into structured motion data suitable for robot learning.
November 2025 · Research
JEPA world model for human motion
A Joint-Embedding Predictive Architecture that learns human movement dynamics from clinical motion capture, predicting future state in a latent space.
November 2025 · Research
From LiDAR scan to walkable 3D world
Turn an iPad Pro into a room scanner that produces a first-person walkable digital twin from a single 60-second walk-around.