I am working on a javascript-web based simulation of a spring-cart-pendulum system in order to apply some math I've recently learned into a neat small project. I found a sample image on the internet as to what I am trying to model: enter image description here

I understand how to derive the equations of motions by finding the Lagrangian and then from there deriving the Euler–Lagrange equations of motion:

Given that the mass of cart is $M$ and mass of ball on pendulum is $m$, length of pendulum $l$, distance of cart from spring (with constant $k$) equilibrium $x$ and pendulum angle $\theta$ we have the following equations of motion:

$$x'' cos(\theta) + l\theta'' - gsin(\theta) = 0$$ $$(m+M)x''+ml[\theta''cos(\theta) - (\theta')^2sin(\theta)] = -kx$$ (A derivation can be found here, although this is not the point of the question ).

I have isolated the equations for $x''$ and $\theta''$ as follows:

$$x'' = \frac{-kx - ml[-sin(\theta)(\theta')^2-\frac{g}{l} cos(\theta)sin(\theta)]}{m+M - mcos^2(\theta)}$$

$$\theta'' = \frac{-gsin(\theta)-cos(\theta)x''}{l}$$

I then used Euler's method by using previous/initial conditions to solve for $x''$ and then for $\theta''$ (since it depends on $x''$. I then solved for $x' = x''t + x'$, $x = x't + x$ (likewise for $\theta$), where $t$ is the stepsize.

This approach seems to work well for initial conditions where $\theta$ or $\theta'$ is small (have yet to test playing around with $x$'s). However, when I set the $\theta$'s to bigger values, then my spring goes crazy and all over the screen.

Now I have never physically set up a spring-cart-pendulum system so maybe my simulation is correct ( although I have my doubts), but I was wondering what other methods I can use for numerical and more accurate approximation. I could always set the step size for Euler approximation to something smaller, but that would effect the speed and functionality of my simulation.

I have heard about things like Runge–Kutta but not sure how to apply it to a second order system of DEs. Does anyone have any advice, reading or examples that I can look at to learn from for more accurate approximation?

Thank you for your time.



2 Answers 2


You can always transform a system of higher order into a first order system where the sum over the orders of the first gives the dimension of the second.

Here you could define a 4 dimensional vector $y$ via the assignment $y=(x,θ,x',θ')$ so that $y_1'=y_3$, $y_2'=y_4$, $x_3'=$ the equation for $x''$ and $y_4'=$ the equation for $θ''$.

In python as the modern BASIC, one could implement this as

def prime(t,y):
    x,tha,dx,dtha = y
    stha, ctha = sin(tha), cos(tha)

    d2x = -( k*x - m*stha*(l*dtha**2 + g*ctha) ) / ( M + m*stha**2)
    d2tha = -(g*stha + d2x*ctha)/l  

    return [ dx, dtha, d2x, d2tha ]

and then pass it to an ODE solver by some variant of

sol = odesolver(prime, times, y0)

The readily available solvers are

from scipy.integrate import odeint
import numpy as np

times = np.arange(t0,tf,dt)
y0 = [ x0, th0, dx0, dth0 ]

# odeint has the arguments of the ODE function swapped
sol = odeint(lambda y,t: prime(t,y), y0, times)

Or you build your own

def oderk4(prime, times, y):
    f = lambda t,y: np.array(prime(t,y)) # to get a vector object
    sol = np.zeros_like([y]*np.size(times))
    sol[0,:] = y[:]
    for i,t in enumerate(times):
        dt = times[i+1]-t
        k1 = dt * f( t       , y        )
        k2 = dt * f( t+0.5*dt, y+0.5*k1 )
        k3 = dt * f( t+0.5*dt, y+0.5*k2 )
        k4 = dt * f( t+    dt, y+    k3 )
        sol[i+1,:] = y = y + (k1+2*(k2+k3)+k4)/6

    return sol

and then call it and print the solution as

sol = oderk4(prime, times, y0)

for k,t in enumerate(times):
    print "%3d : t=%7.3f x=%10.6f theta=%10.6f" % (k, t, sol[k,0], sol[k,1])

Or plot it

import matplotlib.pyplot as plt

x=sol[:,0]; tha=sol[:,1];
ballx = x+l*np.sin(tha); bally = -l*np.cos(tha);
plt.plot(ballx, bally)

where with initial values and constants

m=1; M=3; l=10; k=0.5; g=9
x0 = -3.0; dx0 = 0.0; th0 = -1.0; dth0 = 0.0;
t0=0; tf = 150.0; dt = 0.1

I get this artful graph:

trajectory of the point of the spring pendulum

  • $\begingroup$ I appreciate the answer, but I still have some questions!I am coding in javascript for a more 'real time' solution. A sample simulation can be found here. I cannot use the specific libraries that you mentioned because they are for python. Even if there was a library for javascript, I would like to keep myself away from using it because I am trying to learn a little bit more about the numerical solutions to the system of second order DEs in order to scale it to other simulations. I'm trying to understand the theory behind $\endgroup$
    – stantheman
    Commented Nov 23, 2016 at 23:14
  • $\begingroup$ the solutions. I was looking for some sort of reading on solving systems such as this something along the lines of your rk4 example but with more detail. Cheers $\endgroup$
    – stantheman
    Commented Nov 23, 2016 at 23:17
  • $\begingroup$ Some basic example of RK4 in javascript: stackoverflow.com/questions/29830807/runge-kutta-problems-in-js . The problem is that lists in JS are not vectors. One would have to implement vector arithmetic using list operations or loops. Then in principle the fixed time step methods work the same way as above. -- There is a numericjs project with the standard integrators. $\endgroup$ Commented Nov 24, 2016 at 0:02

In JavaScript one can implement a universal variant of the Euler or RK4 step using the axpy array operation introduced by BLAS/LAPACK

function axpy(a,x,y) { 
// returns a*x+y for arrays x,y of the same length
    var k = y.length >>> 0;
    var res = new Array(k);
    while(k-->0) { res[k] = y[k] + a*x[k]; }
    return res;

This blindly assumes that the input arrays have the same length.

function EulerStep(t,y,h) {
    var k = odefunc(t,y); 
    // ynext = y+h*k
    return axpy(h,k,y);

function RK4Step(t,y,h) {
    var k0 = odefunc(t      ,               y );
    var k1 = odefunc(t+0.5*h, axpy(0.5*h,k0,y));
    var k2 = odefunc(t+0.5*h, axpy(0.5*h,k1,y));
    var k3 = odefunc(t+    h, axpy(    h,k2,y));
    // ynext = y+h/6*(k0+2*k1+2*k2+k3);
    return axpy(h/6,axpy(1,k0,axpy(2,k1,axpy(2,k2,k3))),y);

The ODE function would be correspondingly

function odefunc(t,y) {  
    var x = y[0], tha = y[1], dx = y[2], dtha = y[3];  
    var stha = Math.sin(tha),  ctha = Math.cos(tha);

    var d2x = -( k*x - m*stha*(l*dtha**2 + g*ctha) ) / ( M + m*stha**2)  
    var d2tha = -(g*stha + d2x*ctha)/l  

    return [ dx, dtha, d2x, d2tha ]

Now one can do one integration step per frame, or multiple time steps with accordingly reduced time step.

The more complicated variant is to use an optimal time step for a required error level and interpolate the states for the time points of the frames, which could mean that there is no integration step for several frames. This is essentially what the dopri5/rk45/ode45 or lsoda solvers do, the use internally computed time steps and interpolate the values for the given list of times.


You must log in to answer this question.