We describe the on-orbit characterization of the HgCdTe detectors aboard NICMOS. The flat-field response is strongly wavelength dependent, and we show the effect of this on the photometric uncertainties in data, as well as the complications it introduces into calibration of slitless grism observations. We present the first rigorous treatment of the dark current as a function of exposure time for HgCdTe array detectors, and show that they consist of three independent components which we have fully characterized - a constant component which is the true dark current, an 'amplifier glow' component which results from operation of the four readout amplifiers situated near the detector corners and injects a spatially dependent signal each time the detector is non-destructively read out, and finally the 'shading', a component well known in HgCdTe detectors which we show is simply a pixel dependent bias change whose amplitude is a function of the time since the detector was last non-destructively read out. We show that with these three components fully characterized, we are able to generate 'synthetic' dark current images for calibration purposes which accurately predict the actual performance of the three flight detectors. In addition, we present linearity curves produced in ground testing before launch. Finally, we report a number of detector related anomalies which we have observed with NICMOS some of which have limited the observed sensitivity of the instrument, and which at the time of writing are still not fully understood.