A social mini-game (SMG) is a short, local interaction in a physically constrained environment in which a few agents (robots and/or humans) compete for a shared space that only one of them can use at a time (e.g. going through doorways, narrow hallways, unsignalized intersections etc.). SMGs exhibit high agency where small actions can significantly change the interaction dynamics, coordination requirements, and potential for deadlocks.
SMGs can be formalized as a multi-agent variation of a spatially constrained, time-optimal, motion planning problem in which a frequent failure mode is the freezing robot problem in which robots stop moving because every available motion looks unsafe.
Humans solve SMGs quickly, without speaking and usually without slowing down much. We want robots to do the same, and we want to be able to prove that they will.
The following are some exciting open problems in robotics and autonomy towards solving SMGs:
Our main result on this question. We formulate social mini-games in discrete time and derive barrier functions that certify collision avoidance and rule out deadlock at the same time, with no communication between robots.
A single neural controller that handles safety, liveness, and robustness together, and that intervenes on the nominal plan as little as it can. The robot deviates only where the interaction actually requires it.
Takes the same guarantees to high-dimensional inputs. The controller runs directly on dense observations instead of a clean state estimate, and stays fully decentralized in cluttered environments.
Real robots have bounded actuators, and most safety guarantees quietly assume they do not. This work learns neural integral control barrier functions that respect input limits, and scales to large teams.
Yielding is also a fairness problem. If the controller always picks the same robot to give way, the team is still safe, but one robot pays for everyone else's progress. FACA assigns priorities that change over time, so the cost of yielding is spread across the team without giving up agility. It is the control-side answer to the question GameChat asks in language.
Classical multi-agent path finding assumes agents execute whatever plan they are handed. SocialMAPF drops that assumption: agents are strategic, so we compute an ordering each agent has an incentive to follow, fast enough to run online.
Game-theoretic ordering for human drivers at intersections, roundabouts, and merges. Priority is settled by an auction among the drivers rather than by a central authority, which is closer to how these situations actually resolve.
Humans differ in how much risk they will take, and a planner that assumes the average human fails against the tails. We make risk preference an explicit parameter of the game, so the robot's plan degrades gracefully when the other driver is more aggressive than expected.
Real-time priority assignment and control at unsignalized intersections with mixed traffic, where no signal exists to impose an order and the agents must find one themselves.
Adds fuel efficiency to the objective. Ordering that is optimal for throughput is not automatically optimal for energy, and at an intersection the gap between the two is measurable.
All of the above needs reward functions, and hand-designing them is guesswork. Here we recover them from real pedestrian crowds instead — a multi-agent inverse RL problem in which none of the behavior is labeled and the crowd is genuinely unstructured.
A VLM watches the scene and tunes the barrier function online. The safety certificate still comes from the CBF; the model only decides how conservative to be, and does it fast enough to stay in the control loop.
Robots negotiate in natural language. Each robot runs its own LLM, they exchange a short dialogue about whose task is more urgent, and they agree on an order. The outcome is socially optimal rather than merely collision-free: a robot with nothing to deliver yields to one that is running late. GameChat argues for priority in language; FACA enforces it in the controller and keeps it fair over repeated encounters.
Where the term comes from, what counts as one, and how the existing algorithms relate to each other. This is the best single entry point to everything on this page.
A community effort across more than twenty groups on what social navigation should be measured against. Much of the disagreement in the field turns out to be disagreement about metrics.
A simulator for multi-robot learning and navigation in shared human spaces, built so that policies can be trained and compared on the same scenarios.
Our open implementation of the social mini-game scenarios and the baselines that solve them. We use it in our own papers and in the graduate multi-robot navigation course.