What this assignment does
The review covers deep reinforcement learning for robotic manipulation: the algorithm families, the reality gap between simulation and hardware, and the safety problem created by an exploratory learning process running on a physical arm.
It is organised around the reality gap rather than around papers or dates, so each contribution is introduced by the problem it attacks and judged on how completely it solves it.
How it is built
Foundations, algorithm families, sim-to-real transfer, safety, emerging directions, then implications for an industrial deployment. The deployment section is what converts a review into an argument — the literature has to pay out into a recommendation.
Within sim-to-real the structure repeats: the source of the gap, the technique that addresses it, then a critical evaluation of what the technique leaves unsolved.
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.
Domain randomisation is conceptually simple and has produced impressive results, but it is not a complete solution. Randomisation ranges must be chosen by the engineer; ranges that are too narrow fail to cover reality, whereas ranges that are too broad make the task needlessly hard and can produce conservative policies.
Why this marks well
This is what critical reviewing looks like. The technique is credited, then the hidden dependency is named: domain randomisation moves the engineering judgement rather than removing it, because somebody still has to choose the ranges.
The failure modes are given in both directions — too narrow fails to cover reality, too broad produces conservative policies. A one-sided criticism usually means the writer has found an objection; a two-sided one means they understand the mechanism.
“Not a complete solution” is also the correct strength of claim. The method is not debunked, it is bounded, which is the honest position given the results it has produced.
What a weaker answer does here
A weaker answer describes domain randomisation, cites Tobin et al. and Peng et al. for impressive transfer results, and moves to the next technique. The reader learns the field's successes and none of its open problems, which is precisely what a critical review is assessed on.
Reinforcement learning is inherently exploratory, and exploration on physical hardware can damage the robot, the environment or nearby people. García and Fernández (2015) distinguish two broad categories of safe RL: modifying the optimisation criterion, for example by penalising risk or imposing constraints, and modifying the exploration process, for example by incorporating external knowledge or a safety supervisor that overrides dangerous actions.
Why this marks well
The safety problem is derived from the method rather than appended as an ethics paragraph. Exploration is not a side effect of RL, it is RL — so the hazard is structural, and the review says so in its first clause.
The taxonomy that follows is then genuinely useful: two categories, each defined by what it modifies, each with a concrete instance. A reader can classify a new paper using it, which is the test of whether a review's framework was worth constructing.
Many students treat safety as a closing obligation. Placing it as a full section with its own literature signals that it is an engineering constraint on deployment, which is how industry treats it.
What a weaker answer does here
A weaker answer adds a paragraph noting that safety and ethics must be considered when deploying robots. Unobjectionable, detached from the algorithms, and worth almost nothing.
Domain randomisation tackles the gap by training across a wide distribution of simulated conditions, so that the real world appears to the policy as just another sample. Tobin et al. (2017) randomised textures, lighting and camera positions to train an object-localisation network entirely on synthetic images that transferred to a real robot. Peng et al. (2018) extended the idea to physical parameters such as mass, friction and latency.
Why this marks well
The mechanism is given in a single clause — the real world becomes just another sample from the training distribution — before any citation appears. Explaining the idea in your own words first, then evidencing it, is what stops a review reading as a chain of summaries.
The two citations are then sequenced to show the idea developing: visual properties first, then physical parameters. “Extended the idea” states the relationship between the papers explicitly, which is the connective tissue a literature review is actually marked on.
What a weaker answer does here
A weaker answer gives a paragraph to Tobin et al. and another to Peng et al., each accurately summarised, with no statement of how the second relates to the first. The reader gets two summaries and no field.
What makes this a strong answer
It is organised by problem, not by paper. The reality gap gives the review a spine, and every technique is introduced as an attempt on it.
Relationships between papers are stated rather than implied — extended, complementary, challenged. That is the single clearest marker of a review that has been thought about rather than assembled.
Safety is treated as a structural consequence of how RL works, not as a closing ethical gesture.
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-02), 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