A physics party game simulates a body instead of animating one, then gives you controls too coarse to steer it precisely. The distance between what you meant and what the simulation did is the content. Nearly every design decision in the genre falls out of a single verb.
You latch onto a ledge, a sofa, or another player's face, and the instant you do, two bodies become one physics problem that neither player is fully driving. Gang Beasts puts a hand on each shoulder button, so lifting someone is a two-handed commitment that leaves you unable to defend yourself while you carry them. Human: Fall Flat gives you a grip per hand and arms that follow the camera, which makes climbing a question of aim rather than timing. Heave Ho reduces the character to two hands and a torso: you cross a chasm by grabbing a wall, then a friend's wrist, then swinging them ahead of you.
What grab buys a designer is incident without systems. No combat, no aiming and no score is needed for two players to generate a story — one verb that targets other players turns everybody else into terrain.
The lineage is a run of single-player games that made the controls the antagonist. QWOP (Bennett Foddy, 2008) maps four keys to a sprinter's two thighs and two calves. Surgeon Simulator (Bossa Studios, grown from a 48-hour Global Game Jam build in 2013) puts a finger on each key and an operating theatre under your unsteady hand. Octodad — a DePaul student project before Octodad: Dadliest Catch — asks you to move an octopus's limbs one at a time while passing as a suburban father. Each takes away a competence you already have and sells it back one limb at a time.
The tuning is the craft, and the target is a character that's predictably unpredictable: wobbly enough that walking is funny, consistent enough that skill exists and improvement shows. Miss in one direction and you have an ordinary game with a silly walk. Miss in the other and outcomes stop tracking inputs, at which point nobody's win means anything. It holds together only because everyone flails under identical rules — that fairness is what keeps the laughter pointed at the situation rather than at the game.
The level plays it straight so the bodies don't have to. See-saws, conveyors, spinning arms, tiles that drop away, doors that swing back: each is a plain mechanism doing exactly what it advertises, and the comedy is what a ragdoll does on meeting one. An arena that tries to be funny itself just puts two jokes in competition.
Pace is authored here too. Eliminations want to be quick, visible and legible to everyone watching, because going out is a punchline and a punchline has timing. Fall Guys (Mediatonic, 2020) built the strongest version of that structure — shows that begin at sixty and cut the field round by round, so nobody sits out long and the crowd thins on a schedule. The format also assumes an audience, which is a real structural advantage: this is a genre watched at least as much as it is played.
A rigid-body simulation is not reproducible in practice. Solver iteration order, variable frame times and accumulated floating-point error mean identical inputs don't produce the same fall twice, and across two networked machines they are not even trying to. Two consequences follow, and both are structural rather than fixable.
For the builder, fairness cannot be tested into a level. You can't run a round a thousand times, read a win rate off it and adjust, because the thing you are measuring has no stable answer — playtesting tells you how a hazard feels, but never quite what it does. For the player, a defeat can't reliably be traced to a mistake. That is harmless while the outcome is a joke and corrosive the moment a round matters, which is why competitive scenes here stay small and why the meta is costumes and rotating events rather than a ladder.
Physics puzzles sit on the opposite side of the same property: they need the simulation to behave the same way every time, because their contract is that a solution can be repeated. Point the wobble at jobs instead of opponents and you arrive at co-op chaos, where the uncooperative sofa is the antagonist rather than the punchline. Human: Fall Flat is the case worth knowing on that border: it shipped in 2016 as a single-player puzzle game and gained the multiplayer it is now known for afterwards.
Type: Physics party game (an indirectly steered simulated body; the sim writes the failure, not a script)
Core loop: Grab, shove, clamber and fall through an arena built so all four go wrong
Perspective: Third-person 3D, whole-body ragdoll — a handful in a room, or dozens in a show
Win / lose: Last standing, first across, or the whole group through; you lose to a ledge as often as to a player
Session: Rounds of a couple of minutes, dealt back to back with no downtime
Progression: None mechanical — costumes, events and new arenas
Decide early whether anyone is supposed to win. If they are, you owe players a simulation consistent enough to learn from, and that constrains everything above it. If they aren't, you owe them a reason to come back on the eleventh night, and costumes have never been that reason on their own.