What this assignment does
The essay compares the convolutional and transformer lineages for object detection, and asks which is the better choice specifically for perception on a mobile robot — a constraint that changes the answer.
Its strongest move is to challenge the evidence it has just used. Having compared architectures on standard benchmarks, it argues that those benchmarks may not transfer to a robot's viewpoint at all.
How it is built
Foundations, the convolutional lineage, the transformer lineage, comparative evaluation, implications for deployment, conclusion. The two lineages get symmetrical treatment, which is what makes the comparison table legible.
The deployment section reframes everything around latency, determinism and embedded hardware support — the constraints that actually decide the question for a robot rather than for a leaderboard.
Annotated extracts
Three passages, each with a note on why it earns marks and what a weaker answer does in the same place — enough to teach the move, not enough to stand in for the assignment.
Benchmark comparisons must be interpreted cautiously. COCO consists largely of internet photographs, whereas a mobile robot observes the world from a low, moving viewpoint with motion blur, variable lighting and a different object distribution. Geiger, Lenz and Urtasun (2012) showed with the KITTI benchmark that methods ranking highly on general datasets did not necessarily perform best on driving data, illustrating the risk of domain shift.
Why this marks well
The essay turns on its own evidence. Having used benchmark results to compare the architectures, it now argues that the benchmark's image distribution does not match the application — and names the specific mismatches: viewpoint height, motion blur, lighting, object distribution.
The KITTI citation is doing proper evidential work. It is not an example of domain shift in the abstract; it is a documented case where leaderboard rank failed to predict performance on the application domain. That converts a reasonable worry into a demonstrated risk.
This is the paragraph that would get quoted back approvingly in feedback. Students are taught to support claims with evidence; far fewer are taught to ask whether the evidence measures the thing they care about.
What a weaker answer does here
A weaker answer reports mAP figures from COCO, declares the higher one better, and recommends it. The numbers are real and the recommendation is unsafe, because nothing has established that COCO performance predicts robot performance.
For a mobile robot, latency directly affects safety: at 1.5 m/s, a 100 ms perception delay corresponds to 15 cm of travel before the robot can react. The end-to-end nature of DETR-family models is attractive because removing NMS makes inference time less dependent on the number of objects in the scene.
Why this marks well
One arithmetic step converts an abstract engineering quantity into a safety distance. 100 ms is meaningless to most readers; 15 cm of uncommanded travel is not. That sentence is the whole argument for why latency dominates accuracy here.
The architectural consequence then follows directly: removing non-maximum suppression makes inference time less scene-dependent, which matters for a robot because worst-case latency — not average — is what bounds safe speed. Reasoning about determinism rather than throughput is a genuinely postgraduate distinction.
Note the balance in what follows: CNNs retain the advantage in quantisation and accelerator support. Neither architecture is allowed to win outright.
What a weaker answer does here
A weaker answer states that real-time performance is important for robotics and quotes frames per second for each model. FPS without hardware, batch size or precision is close to meaningless, and the essay elsewhere says so.
Inductive bias: strong (locality, translation equivariance) for CNNs; weak, learned from data, for vision transformers. Data requirement: moderate for CNNs, trains well from ImageNet; high for pure ViTs, reduced by pre-training and hybrid designs. Small-object accuracy: strong with feature pyramids; historically weaker, improving. Global context and occlusion reasoning: limited receptive field in early layers; strong from early layers.
Why this marks well
The comparison is organised by property rather than by architecture, which forces a genuine like-for-like judgement on each row instead of two separate descriptions.
Crucially, the first row explains the rest. Strong inductive bias is why CNNs need less data; weak inductive bias is why transformers need more but reason globally. A table whose first row causes the others is an argument in tabular form.
The temporal hedges are accurate rather than evasive — “historically weaker; improving” is the honest state of small-object detection for ViTs, and dating a claim is better than overstating it.
What a weaker answer does here
A weaker answer tabulates accuracy, speed and model size for a handful of named models. It compares implementations rather than architectures, so nothing in it generalises beyond the specific versions benchmarked.
What makes this a strong answer
It questions whether its own evidence base measures the right thing, and supports that challenge with a documented case rather than a worry.
Abstract engineering quantities are converted into consequences a reader can feel — 100 ms becomes 15 cm of travel.
The comparison is built on properties that explain each other, so the table reasons rather than merely records.
Request the full document
Request the full document
The extracts above are the teaching material and they stay on this page. The complete 2,050-word document is sent person to person, by a human, after a short conversation about what you are working on — so that we can point you at the right parts of it rather than hand over a file and wish you luck.
Both links below already name this sample (RAI-03), so you do not have to describe it.
WhatsApp · +44 7537 126 976 · Email · mia@globalprojectshelp.com
Mon–Sat, 9–22 GMT. No form, no mailing list, no download gate.
Full documents are never published, hosted or linked on this site.
Using this sample honestly
This is a worked example. It sits in the same place as a worked example in a textbook or a past paper with an examiner’s report attached: read it, take the move, then write your own paragraph without it open in another window.
Submitting any part of this document as your own work would be plagiarism — the same offence as copying from a journal article, and detected the same way. If you draw on it, cite it. If a deadline has got to the point where submitting someone else’s paragraph feels like an option, say so when you message us; that is a problem with a legitimate answer and we would rather work on it than watch it happen.
Our full position, including how to declare mentoring support to your institution, is on the academic integrity policy page.
Where to go next
- More MSc Robotics and Artificial Intelligence samples The full set of annotated model answers for this course.
- All samples Every annotated model answer in the library, with discipline filters.
- How requesting works What you receive, what you do not, and the conditions attached.
- Academic integrity policy What we do, what we will not do, and how to declare mentoring support.
Model answer written by a Global Projects Help mentor · © Global Projects Help
Published 8 October 2026 · Questions? mia@globalprojectshelp.com