The RVO algorithm is inherently local and will get stuck in local optima (like in this case). That agent's can be locked is not something that the RVO algorithm was originally built to handle, and it does affect the quality in situations like these.
The good news is that I do have another local avoidance algorithm in development that handles these cases a lot better. It is based on how Starcraft 2 does it.
Here is a demo of it https://www.youtube.com/watch?v=wBUdMXeeej0 (when the agents start attacking other agents then they are marked as locked, but agents behind them still manage to move around them to get to the front line).
I'm not really sure when it will be released though, I have a lot of stuff going on right now. I could perhaps upload a beta version pretty soon.