Caliper — a robotics engine in Rust.
Robotics software makes you choose between a datacenter and a junk drawer.
Caliper is my bet that most arm work needs neither: one deterministic Rust
engine — kinematics, IK, dynamics, contact sim, datasets, training verdicts —
behind a desktop app, a CLI, and import caliper. All three faces
run the exact same code.
break the limits
early measurements, v0.1, apple silicon — method in the docs
It tells you why
The prevailing workflow in robot learning is: export a URDF from CAD, record demos, start training, and hope. Caliper assumes your robot description is broken, your dataset is flawed, and your policy will do nothing on deploy — and instruments all three. An asset doctor catches the inertia your converter dropped. A dataset doctor catches variance collapse before the GPU bill. A policy autopsy reads a failed training run and names the culprit. Every finding has a stable code, a plain-English message, and a fix hint — like the toy report above, but real.
Why small is the point
GPU-sim platforms want tens of gigabytes and a driver matrix; the ROS stack wants a workspace you're afraid to update. Caliper is one signed download, runs on a CPU laptop, and puts a robot on your screen in under a minute. Determinism isn't a feature flag — the whole engine is built on it, which is what makes the doctors possible at all.