THE JAX*Y SUITE · OPEN SOURCE · MIT

Every robot is a
differential equation.
We build the tools that treat it like one.

Three libraries built on JAX. Simulate hybrid dynamical systems with gradients flowing through the physics. Calibrate the model to the unit on your bench, from its own telemetry. Compile the result to embedded C that's checked, step for step, against the workstation build.

$ pip install jaxonomy jaxterity jaxility
SIMULATE jaxonomy CALIBRATE jaxterity DEPLOY jaxility

jaxonomy

SIMULATION ENGINE · HYBRID DYNAMICAL SYSTEMS
  • Model any physical system as block diagrams or acausal component networks — electrical, mechanical, thermal, hydraulic — with ODE/DAE solvers, zero-crossing events, and state machines.
  • Gradients flow through the simulation. jax.grad of a closed loop matches finite differences to ~1e-6, so controllers and physical designs are tuned by descent instead of by sweep.
  • Compiled and batched with jit and vmap; exports FMI 2.0 co-simulation FMUs that pass the official validator with zero findings.
IN THE BOX
150+ library blocks LQR · MPC · Kalman · EKF/UKF Pantelides index reduction event-time (saltation) gradients adjoints through DAEs multirate substepping balred · POD–DEIM · Koopman neural corrections inside DAEs FMI 2.0 export
machinavitalis/jaxonomy docs ↗

jaxterity

ROBOTICS LAYER · PER-UNIT CALIBRATION & TWINS
  • Your URDF says every joint is frictionless. Your servos disagree. Jaxterity fits actuator dynamics from the robot's own telemetry — on our SO-101 arm, that cut shoulder torque error by 45% on held-out motions.
  • A calibrated model becomes an attested twin: fitted parameters, off-body physics (servo heating, battery sag), and the unit's identity under one handle. Recalibrate, and every prior sign-off is invalidated.
  • Predict and veto before the robot moves. The twin flags the hold that will reach a servo's thermal ceiling, with time-to-ceiling, and the maneuver that will sag the battery into brownout.
IN THE BOX
URDF / MJCF → MJX articulated dynamics reverse-mode AD through limits & friction per-unit sysid with identifiability scoring calibrated actuator & sensor blocks SafetyMonitor & preflight attestation handles whole-body control QP self-check co-processor on a Raspberry Pi
machinavitalis/jaxterity

jaxility

DEPLOYMENT COMPILER · ATTESTED EMBEDDED C
  • One lane: JAX → CasADi → acados → plain C for Arm targets. If an op isn't covered, the build stops and names it at compile time — before anything reaches the robot.
  • Deployed code is measured, not trusted. On a real Raspberry Pi 5, the compiled MPC's outputs match the workstation build step-for-step to ~1e-15, and the solve takes ~100 µs of a 1 kHz control cycle.
  • Every build writes a hash-chained manifest — sources, dependencies, toolchain — that jaxility verify re-checks end to end. Same inputs, byte-identical manifest.
IN THE BOX
OCP templates: LQR · MPC · WBC · centroidal URDF → CasADi dynamics generator no-malloc arena runtime pinned Arm toolchain in Docker BLAKE3 hash-chained manifests HIL parity harness ONNX → LiteRT policy lane dual-path arbiter with safety clamp
machinavitalis/jaxility

The model is differentiable. The robot is measured.
The binary is checked.