Simulation fundamentals: physics engines, time steps and real-time factor

Simulation lets a robot team drive a thousand routes overnight, crash into walls without cost, test behaviour in rain that never comes to the test site and run the same scenario again and again until a bug is understood. It is also easy to misuse: a simulation that looks convincing can be wrong in ways that matter, and a team that trusts it blindly ships software tuned to a world that does not exist. This lesson explains what a robot simulator actually computes, how numerical integration, time steps and contacts shape the results, what the real-time factor means, and how the main simulators compare. Lesson 6 covers building models and worlds, and lesson 8 covers closing the gap between simulation and reality.

What a robot simulator does

A robot simulator combines several models. A physics engine advances the state of rigid bodies (positions, orientations, velocities) under forces, joint constraints and contacts. Sensor models generate what cameras, lidars, IMUs, encoders and GPS receivers would report, usually with added noise. Actuator models turn commands into forces or velocities, sometimes including motor limits and delays. A rendering engine draws images for cameras and for people watching. An interface layer connects all of this to the robot software, in ROS 2 typically through bridge nodes that publish simulated sensor topics and subscribe to commands, together with a simulated clock.

The physics engine is the core. At each time step it gathers the forces on every body, detects which bodies are touching, computes the contact and constraint forces that prevent interpenetration and keep joints together, and integrates the equations of motion forward by one step. Rigid-body dynamics follows the same Newton–Euler equations met in ROB-300; the difficulty lies in contacts and in doing it fast and stably for many bodies.

Numerical integration and stability

Simulators advance the state by discrete time steps h, replacing the continuous equations with an update rule. The simplest is explicit (forward) Euler: new position from old velocity, new velocity from old acceleration. It is cheap but unstable for stiff or oscillating systems: for an undamped spring it steadily adds energy until the motion explodes. Semi-implicit (symplectic) Euler updates velocity first and then uses the new velocity to update position. It costs the same but keeps energy bounded for oscillators, which is why many physics engines use it. Higher-order methods such as fourth-order Runge–Kutta (RK4) are more accurate per step but cost several force evaluations per step and do not mix easily with contact constraints.

import math
# Undamped mass-spring: m x'' = -k x, m = 1 kg, k = 100 N/m (natural frequency 10 rad/s)
k, m, x0 = 100.0, 1.0, 0.1
E0 = 0.5 * k * x0**2
def energy(x, v): return 0.5 * m * v**2 + 0.5 * k * x**2
def explicit(x, v, h): return x + h * v, v - h * k / m * x
def semi(x, v, h):
    v = v - h * k / m * x
    return x + h * v, v
def rk4(x, v, h):
    f = lambda x, v: (v, -k / m * x)
    a1 = f(x, v); a2 = f(x + h/2*a1[0], v + h/2*a1[1])
    a3 = f(x + h/2*a2[0], v + h/2*a2[1]); a4 = f(x + h*a3[0], v + h*a3[1])
    return (x + h/6*(a1[0] + 2*a2[0] + 2*a3[0] + a4[0]),
            v + h/6*(a1[1] + 2*a2[1] + 2*a3[1] + a4[1]))
for h in (0.001, 0.01, 0.05):
    row = []
    for name, step in (("explicit", explicit), ("semi-implicit", semi), ("RK4", rk4)):
        x, v = x0, 0.0
        for _ in range(int(10 / h)):          # simulate 10 s
            x, v = step(x, v, h)
        row.append(f"{name} {energy(x, v) / E0:9.4g}")
    print(f"h={h:5.3f} s  energy ratio after 10 s: " + "  ".join(row))

h=0.001 s  energy ratio after 10 s: explicit     2.718  semi-implicit     1.004  RK4         1
h=0.010 s  energy ratio after 10 s: explicit 2.096e+04  semi-implicit     1.043  RK4         1
h=0.050 s  energy ratio after 10 s: explicit  2.41e+19  semi-implicit    0.8074  RK4    0.9588

The output shows the pattern clearly. Explicit Euler gains energy at every step size and explodes at the larger ones; semi-implicit Euler keeps the energy bounded, oscillating around its true value rather than drifting away; RK4 is extremely accurate at small steps but starts to lose energy as the step grows. A useful rule for explicit schemes is that the step must be small compared with the fastest natural period in the system. Stiff contacts and stiff joint controllers create very fast dynamics, which is why physics engines use implicit or semi-implicit formulations and why simulations can blow up when a gain is raised or a contact is made very stiff.

Contacts and friction

Contacts are the hardest part of robot simulation. Real contacts are stiff, short-lived and governed by friction that switches between sticking and sliding. Physics engines model them in two broad ways. Penalty methods treat contact as a very stiff spring and damper that pushes bodies apart in proportion to how far they interpenetrate; they are simple but need small steps and allow some penetration. Constraint-based methods formulate contact as a complementarity problem: contact forces must be non-negative, bodies must not penetrate, and forces exist only where bodies touch. These are solved at each step, often approximately and iteratively, with parameters such as the number of solver iterations and error-reduction terms that affect accuracy and stability.

Friction is usually modelled with Coulomb's law, the friction force limited to μ times the normal force, approximated by a pyramid in many solvers. Wheels, grippers and feet depend critically on friction, and simulated grasps or wheel slip can differ greatly from reality if μ, contact stiffness or the contact geometry are wrong. A common symptom is jitter: objects that should rest quietly vibrate or creep because the solver never quite converges. Remedies include smaller steps, more solver iterations, softer contact parameters, better mass ratios between touching bodies and simpler collision shapes.

Simulated time and the real-time factor

A simulator keeps its own clock. Each physics step advances simulated time by h; how long that takes in wall-clock time depends on the computer and the scene. The real-time factor (RTF) is the ratio of simulated time to wall-clock time: an RTF of 1.0 means the simulation keeps pace with reality, 0.5 means it runs at half speed and 5.0 means five times faster. Robot software in simulation must use the simulated clock (use_sim_time in ROS 2, with the simulator publishing on the /clock topic); otherwise timeouts, controllers and filters behave wrongly whenever the RTF is not exactly 1.

Running faster than real time is valuable for testing: a 30-minute mission at an RTF of 6 takes 5 minutes. But the robot software must then compute six times faster too, or it will appear to the simulated world as if it were slow. Many test setups therefore run in lock-step, where the simulator waits for the software to finish each control cycle, which gives deterministic results at the cost of speed.

Determinism

A simulation is deterministic if the same inputs always produce exactly the same outputs. Determinism is precious for debugging and for regression tests, but it is easy to lose: multithreaded physics solvers, random sensor noise without fixed seeds, rendering on GPUs, robot software that runs asynchronously with the simulator and messages that arrive in different orders all introduce variation. Teams aim for determinism in unit-level and short integration tests (fixed seeds, lock-step stepping, single-threaded physics) and accept controlled variation in larger scenario tests, judging those by statistics over many runs rather than by one exact trace.

The simulators in use

Several simulators are common in robotics at the time of writing. Gazebo (formerly Ignition) is the open-source simulator most closely tied to ROS: its LTS releases are paired with ROS 2 LTS distributions, Harmonic with Jazzy and Jetty, released in September 2025, with Lyrical, and the ros_gz packages bridge topics and the clock. It uses pluggable physics engines (DART by default, with others available) and supports many sensors. NVIDIA Isaac Sim, built on the Omniverse platform, offers photorealistic GPU rendering and GPU-accelerated physics and is popular for perception and learning work. MuJoCo, developed for contact-rich control and open-sourced by DeepMind in 2022, is fast and accurate for articulated bodies and widely used in reinforcement learning. Webots, open source since 2018, is easy to start with and common in education. Commercial and domain-specific tools cover autonomous driving, drones and industrial cells.

Choosing a simulator is an architectural decision worth an ADR (lesson 2). The right choice depends on what must be trusted: wheel and navigation behaviour, camera realism, contact-rich manipulation or the speed to run millions of episodes. Many teams use more than one, for example Gazebo for nightly integration tests of the full ROS 2 stack and a GPU simulator for training perception models.