MoonBit · pure standard library
Trajectory generation, feedback control, state estimation, grid planning and drive kinematics in one library. Five packages, one flat surface, and nothing outside the standard library.
01 / Modules
Each package is a separate MoonBit package with its own moon.pkg,
so a project that only needs a filter takes only the filter. The root package
re-exports all seven, which is how the examples address them.
Dense vectors and matrices over Double: products, transpose, LU, QR, Cholesky and the symmetric eigendecomposition.
The traits every algorithm is written against, and the records they pass around: a plant, a controller, an estimator, a trajectory, a mechanism, a state space with its validity oracle, and a planner.
35 itemsPID with conditional-integral anti-windup, feedforward, LQR and integral LQR, iterative LQR, linear MPC over a box-constrained program, pure pursuit, Stanley, computed torque, impedance and a barrier filter.
78 itemsA linear Kalman filter, an extended filter and an unscented one over any plant, a particle filter, a complementary filter, an alpha-beta-gamma tracker and a backward smoothing pass.
41 itemsTrapezoidal and jerk-limited profiles that are time optimal from any boundary state, cubic and quintic polynomials, a natural cubic spline, a clamped B-spline and a synchroniser for several axes.
32 itemsSampling planners over any state space, the Dubins and Reeds-Shepp car spaces with every one of their words, dynamic-window and potential-field local planners, and a bridge to the graph package's grid searches.
63 itemsDifferential, bicycle and Ackermann bases, mecanum and omnidirectional wheels, an n-link planar arm with a damped inverse, a delta cell, a SCARA and a gantry.
61 items02 / Capabilities
The shapes below are the library's own output: a jerk-limited move, a Riccati gain settling, a filter pulling an estimate out of noise, and a shortest route through a grid.
A rest-to-rest profile for a distance and a peak velocity, acceleration
and jerk. at(t) returns position, velocity, acceleration and
jerk at any instant, and the profile is the minimal-time one for those limits.
lqr iterates the discrete Riccati equation, with the
tolerance and the iteration cap optional. The gain drives the closed loop;
integral LQR adds the states that remove a standing error.
Predict and update in place. The extended filter differentiates the models numerically, which keeps a nonlinear plant to two closures and one cone of covariance.
grid_path returns the route as cell indices through the
graph package's own search, or refuses when there is none
when the goal is walled off.
03 / Surface
The root package re-exports the public types and free functions of all
five modules. Import it under an alias — the examples use
mctl — and every name is one level deep. The module packages
stay importable on their own.
Angles are radians, lengths are metres, time is seconds, and every scalar
is a Double.
import {
"moonbitstack/moonmcl" @mctl,
}
// A jerk-limited move over twelve metres.
let profile = @mctl.scurve(distance=12.0, v_max=2.0, a_max=1.0, j_max=4.0)
let end = profile.sample(profile.duration())
assert_eq((end.position - 12.0).abs() < 1.0e-6, true)
assert_eq(end.velocity.abs() < 1.0e-6, true)
// A route across an eight by eight grid with a wall in it.
let grid = @mctl.Grid::new(8, 8)
for row in 0..<6 {
grid.block(4, row)
}
let route = @mctl.grid_path(grid, start=grid.node(0, 0), goal=grid.node(7, 7))
assert_eq(route.length() > 8, true)04 / Where it fits
The library carries no allocation-heavy machinery and no numerical dependencies: it is a few thousand lines of MoonBit that compile to wasm, JavaScript and native code alike.
Profile a move, run it through a drive's inverse kinematics, and close the loop with a gain from the Riccati equation. The planner and the controller share one notion of time.
Deterministic, allocation-light stepping suits a fixed-timestep loop; the same code runs in the browser through the JavaScript backend.
Every algorithm is a page of readable MoonBit with the equations in the doc comments, so a controller can be swapped for another without a rewrite.
No I/O, no threads and no global state: the filters and controllers hold their state in the value you own, which keeps them easy to place on a fixed budget.