A density-based garbage collection algorithm of swarm robots is proposed in this paper to gather the garbage with patchy-like distribution. Using the kernel function-based estimation method, a density field is established to characterize the spatial distribution of garbage and other robots in robot"s sensing range, and it then is used to coordinate the collective motion of swarm robots to take the garbage collection task. Specifically, an edge-surrounding behavior is designed to perform a ring-shape encirclement around the garbage patchy, a radial shrinkage behavior is established to collect the coarse-grained garbage, and a vortex shrinkage is designed to clear and gather the fine-grained garbage. Finally, the scattered garbage is gathered into high-density cluster by these two-stage collective behaviors of swarm robots, i.e., surrounding the garbage patchy and shrinking the encirclement. Numerical simulations and real swarm-robotic experiments show that the proposed garbage collection algorithm is very efficient and adaptive for gathering either the single-patchy garbage, the multi-patchy garbage, or even the dynamic-patchy garbage.