Capability Composition in Heterogeneous Teams via Barrier-Guided Diffusion Sampling
Robots can learn separately and still generate cooperative plans without demonstrations of teamwork.
The method enforces known physical constraints while sampling trajectories from generative models trained on each agent alone. The authors prove it can generate every team trajectory satisfying those constraints; the demonstrations determine how likely each trajectory is. Numerical results on two cooperative tasks show constraint-satisfying trajectories without retraining when agents join or leave. The researchers also implement the method on a multi-robot testbed.
ArXiv · AI/CL/LG's note
The method enforces known physical constraints while sampling trajectories from generative models trained on each agent alone. The authors prove it can generate every team trajectory satisfying those constraints; the demonstrations determine how likely each trajectory is. Numerical results on two cooperative tasks show constraint-satisfying trajectories without retraining when agents join or leave. The researchers also implement the method on a multi-robot testbed.
ArXiv · AI/CL/LG's note
score 4