ROB 498 / 599 · Fall 2026 · Paper seminar
One per assigned paper, due before the lecture that covers it. Four questions, about 400 words total. This is not a summary. We have all read the paper.
Each one is asking for a judgement, not a fact you can copy out of the paper. The contrasts below are the difference between a response that earns full credit and one that reads like an abstract with the sentences reordered.
What is this paper asserting that was not already established? One or two sentences, in your own words, stated so that someone could disagree with it.
Which experiment actually carries the claim, and how far does it carry it? Not every result in the paper carries weight; most tables are context for one or two that matter.
Where does the argument rest on something that might not hold? Name one, and say what breaks downstream if it fails.
One experiment. Say what you would measure and what result would change your mind. That last clause is what separates an experiment from a wish.
Copy this into your submission. Keep the headings. They are what makes a stack of thirty responses readable, and the graders read by section.
Paper: Author et al., Short Title, Venue Year Reader: Your Name (uniqname) Seminar: I / II / III / IV Discussed with: names, or "nobody" ## 1. The claim What is asserted here that was not established before? One or two sentences, in your own words. State it so that someone could disagree. ## 2. The evidence Which experiment carries the claim, and how far? Name the comparison that would collapse the claim if it came out the other way, and say whether it isolates the claim or confounds it. ## 3. The weakest assumption One assumption the argument rests on that might not hold. What breaks downstream if it fails? Do not use one the paper already lists. ## 4. What you would run next One experiment. What you would measure, what you expect, and what result would change your mind.
Discuss the papers freely. That is the point of a seminar. Write your response alone, and name anyone you discussed it with in the header. Doing so costs you nothing; its absence, when we can tell, costs a lot.
On a paper that is not on the reading list, so that nobody's response is written for them. It is the paper the volumetric-fusion lecture builds from, and every assigned paper in Unit 1 is downstream of it.
Newcombe et al., KinectFusion: Real-Time Dense Surface Mapping and Tracking, ISMAR 2011.
A commodity depth camera and a GPU are enough to build a dense, drift-free model of a room-scale scene in real time, provided you fuse every frame into one persistent volume and track against that rather than against the previous frame. The rival being argued against is frame-to-frame odometry, which was the standard and which drifts.
The comparison that carries the claim is frame-to-model against frame-to-frame ICP: it is the only result that separates the paper's actual contribution (tracking against an accumulated surface) from the two things that came free with the hardware (dense depth, GPU throughput). It is convincing on the drift question. It is weaker than it looks on reconstruction accuracy, which is shown qualitatively (renders and a few closeups) with no quantitative comparison against ground-truth geometry, so “how good is the model” is largely taken on the reader's eye.
That the scene is static. The TSDF update is a running weighted average over every observation of a voxel, which is only meaningful if all of those observations are of the same surface. Nothing in the formulation detects that they are not. A person walking through the scene does not produce an error the system can flag; it produces a smeared surface that is indistinguishable, downstream, from a real one. Everything built on the map (planning, collision checking) inherits that silently.
Record one sequence of a static room, then the same trajectory with a person walking a fixed path through it. Fuse both and measure surface error along that path as a function of how long the person occupied each region. I expect error to grow roughly with dwell time and not to recover, because the weighted average has no mechanism to forget. If instead it recovers once the person leaves, with later observations outvoting earlier ones, then the static-scene assumption is much softer than the formulation suggests, and the practical fix is a weighting schedule rather than explicit dynamic-object handling.
Each response is marked out of 3. 3: all four questions answered with judgement, the assumption is one the paper does not concede, the experiment has an outcome attached. 2: complete and correct, but one or more answers restate the paper rather than assess it. 1: submitted, largely summary. 0: not submitted, or written about a paper you did not read.
Your lowest response is dropped, no questions asked. Late responses are accepted until the end of the semester. Submit on time when you can, though: the discussion is the point, and the response is what makes it possible.