If you can't do it via i2c, i doubt the Arduino would solve any problems you're actually having.
Computationally, an Arduino is so many times slower than the pi that you would need to be very clever about what you do on that to even measurably reduce the load on the RPI. I don't think your motor controller could even deal with as much commands per second that the load on the RPI would even be in the single digit percentages.
The only advantage of a second processor in the system would be that you could dedicate that to low-latency emergency stop (and other safety-critical) functionality, so that you have less strict requirements on in how many milliseconds the pi must have reacted to something. Having to guarantee low latencies on an application processor costs performance, because instead of doing the computationally most efficient of not interrupting your CPU-intense tasks, you will tell your Linux that it needs to pause whatever process is currently running to service the low-latency software, no matter whether there's actually something to do our not.
But USB is a notoriously high-latency bus, and Arduino a notoriously ill-suited operating system for hard real-time requirements that your proposed architecture just won't have that advantage. It also sounded like you weren't planning to let the Arduino react to any observations it has on its own, but to just use it as long of command buffer - that has no advantages whatsoever.
Honestly, talking USB will probably be more additional CPU load than your saving by not having the RPI do a few realtime priority tasks. So, I'd say, for this round, scratch that Arduino, build it directly on the RPI. Should it later turn out that you need to offload some of the control load off the RPI, then you'll actually have some data on how much that is, what the critical latencies are, and then you can devise an architecture that solves the problem that you'll then be able to quantify. Adding an external microcontroller later is easy. Building a heterogeneous multiprocessor system that you don't actually need is hard.
