I wrote a custom implementation of the Radau integrator, its been heavily modified. I have a lot of additional physics, it supports the non-gravitational models that JPL Horizons defaults to, so diurnal yarkovsky at least. I've been using it to study dust and small object dynamics, as they get pushed around by the sun a lot.

It does an OK job for impactors, but the integrator is tuned heavily for performance, and the tolerance defaults are not great for impactors.

I match jpl horizons for apophis to a few km, they have a lot more intense earth gravitational model then I care to implement, and by default I only include the 5 heaviest main belt asteroids, they have many more. That was the sweet spot for accuracy vs speed for me, overall accuracy goal is less than a few km over a decade.

The goal is to be able to handle the huge influx of new asteroids that the catalog will have due to LSST and eventually NEO Surveyor (which I worked on for 3 years). Most systems I know have been throwing hardware at the problem, I tried to make fast and efficient enough software that we can use it on a laptop for 5-10 million asteroids.