Bear with me on this subject because I find it really interesting. I think I will give this a try myself soon.
I should probably start by asking:
What do you wish to have your robot "see"? The candle flame or the heat from the candle?
This starting point should help narrow down the right sensor for your robot and tune the circuit properly.
I started out by looking at the general consensus of the blackbody temperature of a paraffin candle. I found that 1930 Kelvin is the key number used in most calibrations. Knowing that number, I used Wiens Law to find the peak spectrum for that temperature.
Wiens Law: lamdamax
is the peak frequency in nm and Tk
is the objects temperature, Kelvin.
is roughly 1500nm - or - 1.5um, so, if you want to "see" the candle at it's maximum frequency of emittance then a sensor sensitive to the longer IR waves would appear to be suitable. In this case the thermopile would seem appropriate for they all seem to be more sensitive in the high IR range. The TPA81's range is stated to be 2-22um, which in this case is slightly above our optimal temperature. The thermopile I steered you toward earlier s rated at 5.5 to 15um so I was actually in error.
Now, the visible yellow glow of a candle is just the particulates of soot and stuff burning up. This is from the exited hydrocarbon atoms shooting their photons at us. Since this color is generally orange-ish, it's probably somewhere around 650nm. If we wanted our robot to "see" the visible flame itself then it appears a photdiode would be more appropriate. Digikey lists photodiodes that range from around 200 to 1100nm.
Using a photodiode pair would reduce the amount of legwork on your part and you could just about plug them into your existing circuit plus they seem cheaper. Using a thermopile would require some extra work matching an appropriate op amp to the device as I don't believe the LM386 to be quite suitable for this application. Mainly because the thermopile can swing negative if the sensor reads a temperature cooler than the sensor itself. This would require a dual +/- power supply or for you to use a voltage divider to place a virtual ground to the negative side of the op amp. There are single supply op amps on the market as well.
An unrelated note... All the time I was thinking about this today I was wondering if multiple detectors (spanning the UV to IR wavelengths) could be used to sense if an object was moving away from or approaching the robot by interpreting the shift in wavelength. Thats probably a pretty cosmic thought
Hope I was able to help some.