Engineering practice · topic F1 · 40 hours
Physics simulation (MuJoCo / Isaac)
Trial and error in simulation costs three orders of magnitude less than crashing the real robot. Simulation is also the only realistic path to RL training.
What to learn
- The modeling pipeline: from URDF to MJCF / USD, and which parameters you must fill in by hand
- Contact models and solver parameters: friction, restitution, iteration count
- Where the sim-to-real gap comes from: latency, friction, compliance, sensor noise
- Cost estimation for parallel simulation and large-scale data collection
The lesson · plain words
Physical simulation is «faking the world with numerical integration»: every step of dt approximates real physics. Contact (collision, grasping) is where things blow up most easily — it is essentially an ultra-stiff spring, and a slightly too-large dt makes the numerics diverge.
This page builds your intuition for «step size and stability», the first lesson of all simulation work.
dt and explosions: contact is an ultra-stiff spring
Drag the sim-step slider below: make dt large and the ball instantly explodes off screen — the contact stiffness is Kc=40000 N/m, and dt must be small enough to resolve its oscillation period. That is the mathematical nature of a «sim explosion»: the numerical integrator cannot keep up with the fastest physical process.
Explicit Euler explodes most easily; implicit / semi-implicit Euler is far more stable — MuJoCo and Isaac both use semi-implicit internally. You do not need the derivation, but you must know «when it explodes, first cut dt or switch integrators».
Interactive figure · drag it and see
Division of labor: MuJoCo vs Isaac
MuJoCo: accurate contact dynamics, small scenes, fast RL training — the academic standard. Isaac: thousands of GPU-parallel environments, photorealistic rendering, large-scale training — but its contact accuracy is below MuJoCo. Choose by task: fine manipulation research goes to MuJoCo, large-scale RL goes to Isaac.
A universal engineering calculation: one sim step must finish inside the real control period (1 kHz control → sim stepping at 1 kHz at minimum, plus rendering). A simulation that cannot hit real-time rate makes both training and validation untrustworthy.
Three uses of simulation, each with its own trap
① Algorithm validation (planning/control): the trap is an inaccurate model — do F2's system identification first. ② RL training: the trap is overfitting the sim — you need domain randomization (G5). ③ Regression testing: the trap is fragile scenarios — use fixed seeds so results are reproducible.
The acceptance question for any simulation system: «where does this sim differ from the real robot?» If you can list the differences (friction, latency, sensor noise), you understand simulation; if you cannot, you only know which buttons to press.
How deep to go
You can load a model in MuJoCo and run a control loop against it (MJCF or URDF + MJX); explain what dt, solver iteration count, and restitution each affect; and keep the sim stepping in sync with the control loop at a fixed rate. Only then do F2 and F3 connect.
Three questions — do you really get it?
1. What is the most common cause of a sim «exploding»?
- ANot enough CPU
- BA step size too large to resolve the stiff contact dynamics
- CA mistake in the model file
Show answer
B. Contact = an ultra-stiff spring. Explicit integration is step-size sensitive; cut dt or switch to an implicit integrator.
2. Which simulator is the first choice for fine manipulation research (grasping, insertion)?
- AIsaac (large parallel scenes)
- BMuJoCo (contact accuracy)
- CGazebo (richest ecosystem)
Show answer
B. MuJoCo's soft contact model accuracy is industry-recognized; Isaac's strength is massive parallelism and rendering.
3. What does a fixed random seed do in simulation regression testing?
- ASpeeds up the simulation
- BSaves memory
- CMakes results reproducible — a failure you can reproduce is one you can fix
Show answer
C. The value of regression testing is reproducibility. Fixed seed + fixed scenario = a catchable bug.
Keep going deeper
Links without a marker stay on this site; the rest point to the Chinese edition until its translation batch lands.
How it fails in the field
- Contact parameters filled in casually in sim; on the real robot the friction is completely wrong and it walks like it is on ice
- Actuator latency not modeled in sim; the trained policy oscillates forever on hardware
You have learned enough when
The Chinese edition carries the full topic page for Physics simulation (MuJoCo / Isaac), including the complete outline, depth requirements per career track, and further derivations: robotguide.dev/k/F1/.