Protein docking algorithms are assessed by evaluating blind predictions performed during 2007-2009 in Rounds 13-19 of the community-wide experiment on critical assessment of predicted interactions (CAPRI). We evaluated the ability of these algorithms to sample docking poses and to single out specific association modes in 14 _targets, representing 11 distinct protein complexes. These complexes play important biological roles in RNA maturation, G-protein signal processing, and enzyme inhibition and function. One _target involved protein-RNA interactions not previously considered in CAPRI, several others were hetero-oligomers, or featured multiple interfaces between the same protein pair. For most _targets, predictions started from the experimentally determined structures of the free (unbound) components, or from models built from known structures of related or similar proteins. To succeed they therefore needed to account for conformational changes and model inaccuracies. In total, 64 groups and 12 web-servers submitted docking predictions of which 4420 were evaluated. Overall our assessment reveals that 67% of the groups, more than ever before, produced acceptable models or better for at least one _target, with many groups submitting multiple high- and medium-accuracy models for two to six _targets. Forty-one groups including four web-servers participated in the scoring experiment with 1296 evaluated models. Scoring predictions also show signs of progress evidenced from the large proportion of correct models submitted. But singling out the best models remains a challenge, which also adversely affects the ability to correctly rank docking models. With the increased interest in translating abstract protein interaction networks into realistic models of protein assemblies, the growing CAPRI community is actively developing more efficient and reliable docking and scoring methods for everyone to use.