When two people approach each other in a narrow hallway, they often perform a "social dance"—both stepping in the same direction, then mirroring each other again, until one person stops to let the other pass. In robotics, this is the result of symmetric collision avoidance: if every agent assumes an equal burden for safety, the system can become indecisive or overly conservative.
A growing body of research suggests that the key to fluid, human-like navigation is asymmetric responsibility. Instead of treating every agent as equally responsible for avoiding a collision, we can mathematically encode social norms, roles, and physical capabilities to distribute the "safety burden" unevenly.
The Limits of Symmetry
Traditional safety frameworks, such as Reciprocal Velocity Obstacles (RVO) or worst-case Hamilton-Jacobi (HJ) reachability, often treat agents as identical entities. This leads to two major failure modes:
- The Frozen Robot Problem: If a robot assumes every human is a "worst-case" adversary, it may decide that no movement is safe and stop entirely.
- Unrealistic Conservatism: In scenarios like highway merging or racing, assuming a 50/50 split of responsibility ignores the "rules of the road," where the merging vehicle or the follower typically carries more responsibility to avoid contact.
As Julio Erasmo Godoy and Stephen J. Guy argue in Moving in a Crowd, a robot must "read the room." By identifying whether a neighbor is reactive or non-reactive, the robot can decide whether to rely on reciprocity or take 100% of the responsibility for safety itself.
Encoding Responsibility as Control Constraints
How do we turn the abstract concept of "responsibility" into a mathematical constraint? Andrea Bajcsy and Marco Pavone propose using asymmetric control constraints within an HJ reachability framework.
In Towards the Unification and Data-Driven Synthesis of Autonomous Vehicle Safety Concepts, they show that safety can be framed as a differential game. By restricting the "allowable actions" of one agent more than another, they can encode social priorities. For example, a vehicle already in a lane might be given a larger "safe set" of actions, while a merging vehicle is constrained to a smaller set, effectively forcing the merger to bear the burden of the interaction.
Learning Dynamic Allocations
Responsibility isn't static; it shifts with the context of the maneuver. A car driving straight has little responsibility for a car in the adjacent lane, but the moment it begins a lane change, its responsibility spikes.
Isaac Remy and Karen Ka Yan Leung address this in Learning responsibility allocations for multi-agent interactions. They use Control Barrier Functions (CBFs)—mathematical tools that ensure a system stays within a safe set—combined with differentiable optimization. By observing human trajectories, their model learns to "back out" how much responsibility each agent was assuming at any given moment. This allows autonomous systems to predict how humans will allocate the safety burden in real-time.
Agility and Roles
Asymmetry is also a powerful tool for coordinating heterogeneous teams. In AG-CVG, Nare Karapetyan and Dinesh Manocha demonstrate that coordination becomes much simpler when responsibility is assigned to the most capable agent. In a team consisting of a slow ground robot (UGV) and an agile drone (UAV), they assign all "detour responsibility" to the drone. The UGV continues its optimal path without interruption, while the UAV manages all spatial adjustments for docking, reducing coordination overhead.
Similarly, in high-stakes environments like autonomous racing, Rishabh Saumil Thakkar and David Fridovich-Keil found that assigning the primary burden of collision avoidance to the following agent allows for more assertive, realistic behavior. As detailed in Hierarchical Control for Multi-Agent Autonomous Racing, this asymmetry prevents the lead agent from being forced into overly conservative maneuvers, mimicking the "etiquette" of professional human drivers.
Summary
By moving away from symmetric safety models, researchers are creating robots that can navigate complex social environments. Whether through formal HJ reachability, learned CBFs, or role-based assignments, asymmetric responsibility provides a path toward autonomous agents that are safe without being timid, and assertive without being reckless.