Statistical testing for bimodal/multimodal sample

Let's simulate some simple examples, to clarify what you may want to do.

Let's start with a bimodal sample, which would look like this:

It is definitively bimodal, but would you say your device worked well? Maybe not, because the densities around 0 are still quite high (in many instances, it either did not "push" the result at all, or not much).

Now, let's look at another bimodal distribution;

I think we could agree that your device worked much better this time (much fewer observatiosn at or around 0).

And let's look at a 3rd example

I think we can agree that your device now works "perfectly".

But...

1. All 3 are bimodal (and I do not think that declaring one "more bimodal" than the other is sensible)
2. In all 3 cases, the mean is 0, and the median is 0 (the variances are different, but that will not lead us any place useful), and so no t-test, or median test, etc. will let you discern a difference.

So it is not a t-test (for means), or a median test, or a test of multi-modality that you need. Instead it is a test looking at what proportions of your observations are within a range $[-\epsilon, +\epsilon]$, with epsilon of your choosing (in your context, how far from 0 must the object be for you to say it is meaningfully far from 0).
And you also need to define another parameter $p$, which is the proportion of observations in the $[-\epsilon, +\epsilon]$ interval (aka "0 interval") which, in your context, you would consider meaningless (in other words, you would be satisfied with the performance of your device, even if such a small proportion was in the "0 interval".

Then the test you can use is a binomial test. You said that you can "easily" collect 1000 datapoints, so the (relative) low power of a binomial test is no longer an issue.
Then dichotomize your observations; inside the "0 interval", or outside. And compute the upper bound of the CI of the proportion observed in the "0 interval". If it is below $p$, you have demonstrated, at the selected significance level, that your device would allow, at most $100.p\%$ of the results in the "0 interval", and so 'gets the job done'.

Yes, dichotomizing a continuous variable is (usually appropriately) frowned upon; but you can collect large enough samples so that the objections do not matter in practice. And all the continuous proportion tests I know of are based on normal approximations, or normal assumptions, but a bimodal distribution is anything but normal...

An alternative to what jginestet suggested is to take the absolute value of the measure, and then see how far it is from 0. You could do a statistical test on this, seeing if it was different from 0. This would work well in the three examples in jginestet's answer, and it doesn't require dichotomizing the data.