Embodied AI Glossary中文

Physics Engine

物理引擎Essential

The software core that computes forces, collisions, and motion step by step according to physical laws.

A physics engine is software that approximately simulates a physical system, first used heavily in games and film. On every simulation step, it first runs collision detection to find which objects are touching, then a constraint solver computes contact forces, friction, and joint forces, and finally an integrator updates positions and velocities. Engines can be grouped by their goal into real-time engines, which favor speed, and high-accuracy engines, which favor precision. Common ones in robotics include MuJoCo, PhysX, Bullet (PyBullet), ODE, and the newer Newton. It's worth distinguishing from a simulator: a physics engine only handles how things move, while a simulator such as Isaac Sim or Gazebo wraps it with rendering, sensors, scene editing, and a programming interface. The contact model and step size it uses directly determine how accurate and stable the simulation is, and are one of the main sources of the sim-to-real gap.

ExampleThe same robot-arm grasping task, run in MuJoCo versus PhysX, can produce different contact forces and success rates — which is why papers usually state which engine they used.

Related
Simulator · MuJoCo (Multi-Joint dynamics with Contact) · PhysX · Contact Model · Constraint Solver · Simulation Timestep
Sources
Wikipedia: Physics engine
google-deepmind/mujoco (GitHub)

See it in the full glossary →