CoopUQ

CoopUQ is an uncertainty-aware multi-robot collaborative navigation framework, whose core modules include cross-robot spatiotemporal feature fusion prediction and an uncertainty-driven constrained reinforcement learning policy.

CoopUQ uses a learned attention mechanism for cross-robot feature fusion, where robots adaptively weight the information provided by peers according to the confidence of their own viewpoints.

CoopUQ's prediction module outputs prediction uncertainty for each grid cell position and maps these uncertainties into cost penalties in reinforcement learning.

CoopUQ's risk sensitivity is adjustable: the policy can be more aggressive in in-distribution scenarios, while elevated uncertainty in OOD scenarios automatically triggers more conservative behavior.

CoopUQ was evaluated in multiple Unity3D-based simulation environments, covering in-distribution and OOD settings, including different layouts, pedestrian densities, and abrupt pedestrian behaviors.

CoopUQ outperforms representative baselines in both prediction performance and navigation safety.

CoopUQ's real-robot experiments validated its robustness across multiple scenarios.

The research team open-sourced CoopUQ's code and demo video, with the project page at https://coopuq.github.io/.

Tight coupling (such as the cost-constraint coupling adopted by CoopUQ) embeds prediction outputs into the decision optimization objective for joint training; the advantage is that the policy can automatically degrade to conservative behavior under distribution shift, at the cost of higher training complexity and more sensitive inter-module interface design.