Embodied AI Glossary中文

Contact Model

接触模型Common

The physics engine's rule for how much force and friction two touching objects generate.

Collision detection only answers whether and where two objects are touching; a contact model answers how much force results from that contact: a normal force that keeps objects from interpenetrating, and a tangential friction force that resists sliding, capped by a friction cone whose size is the friction coefficient times the normal force. The classical approach is a hard-contact model, which writes “no force without contact, and no interpenetration once there is contact” as a complementarity condition, reducing to a Linear Complementarity Problem (LCP) that becomes NP-hard once friction is included, so different engines rely on various approximations. MuJoCo instead uses soft contact: it allows a small amount of interpenetration and defines the contact force as the unique solution to a convex optimization problem, with parameters like solref and solimp used to tune how soft or stiff it is. How accurate the contact model is directly affects whether contact-heavy tasks — grasping, peg insertion, legged locomotion — can transfer successfully from simulation to a real robot.

ExampleIn MuJoCo, setting a geom's condim to 3 computes only the normal force and in-plane sliding friction; setting it to 4 adds torsional friction, which the official documentation says helps simulate soft fingertips for steadier grasping; setting it to 6 further adds rolling friction.

Also called
contact dynamics
Related
Soft Contact Model · Linear Complementarity Problem · Friction Cone · Constraint Solver · Collision Detection · Coulomb Friction
Sources
MuJoCo Documentation: Computation - Soft contact model
Isaac Sim Documentation: Simulation Fundamentals

See it in the full glossary →