5
$\begingroup$

With New Horizons on final approach to Ultima Thule (2014 MU69) it has been stated that some or all of the images returned may end up being blank due to the camera being pointed incorrectly because of uncertainties in MU69's position (given the over 12 hour round-trip time to NH all camera pointing is pre-programmed).

In order to make sure it doesn’t miss the target, the spacecraft will have to scan its instruments across a region of space in which it is most likely to find the tiny world, taking a lot of data on empty space to make sure to get data on the object.

Source: Emily Lakdawalla, What to Expect When New Horizons Visits 2014 MU69, Ultima Thule, The Planetary Society

Allowing the probe itself to independantly (and perhaps optionally, at its discretion) choose targets of interest1 or fine-tune camera pointing2 for at least of subset of the flyby could improve the quality of the data returned. Has any research been done in this area and what were the findings? Do any current or planned probes do this?


1 Say, an unexpected object spotted while slewing the camera platform to a new, pre-programmed target.

2 i.e. finding the precise location of a known target within a search area.


Case study: New Horizons

According to this source, both the "Command and Data Handling" and "Guidance and Control" computers on NH use 12 MHz Mongoose-V (MIPS R3000) 32-bit RISC CPUs (actually redundant pairs of them) with 15 MIPS each and either 0.75MB (C&DH) or 16MB (G&C) of RAM plus 16×512MB flash (on the C&DH); the processors are interconnected by a 480kb/s bus. Note that on NH only the C&DH processor has direct access to the science payload (i.e. the cameras). This might well be sufficient for simpler AI tasks such as known-target search, possibly even if done as a lower-priority background job. Having worked with MIPS processors, it is my opinion that NH could be up to the task for at least the "known-target search" example though RAM might be a bit tight; a clean-sheet design using NH-era technologies could be fully capable of doing it.

$\endgroup$
  • 1
    $\begingroup$ What "visual" systems would you use for spotting these targets? What corpus would you use for training the AI? $\endgroup$ – JCRM Dec 28 '18 at 9:35
  • $\begingroup$ so it would already have to have the item in its smaller than 1 degree FoV Camera to spot it? and exposure times are around 10 seconds, so it would require two and a half hours to image a 30 degree fov $\endgroup$ – JCRM Dec 28 '18 at 9:57
  • $\begingroup$ ... and given it ndoesn't have to be done in real time, it can be done in the days leading up to the encounter and preprorammed $\endgroup$ – JCRM Dec 28 '18 at 10:44
  • 1
    $\begingroup$ Without a sufficient corpus, are you going to gamble on the AI getting it right, or on your predictions which a HI had a chance to check? $\endgroup$ – JCRM Dec 28 '18 at 10:53
  • $\begingroup$ Let us continue this discussion in chat. $\endgroup$ – Alex Hajnal Dec 28 '18 at 11:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.