- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
VALERANT automatically explores and constructs a 3D point-cloud map from a single game screenshot. At each step, a frozen Matrix-Game 3.0 generates videos for five candidate actions. MASt3R-SLAM reconstructs camera motion and scene geometry, and the system selects a branch that improves coverage and forward progress while passing collision checks. The selected result is integrated into the map. Floor estimation, complementary collision detection, and checkpoint rewind address geometric drift, wall penetration, and local stagnation. Experiments include qualitative results, module ablations, and a visual-preference comparison with HY-World 2.0.
Critical strengths
- The evaluation includes a relevant baseline. The paper compares against HY-World 2.0, renders both methods' outputs as 3DGS scenes, and reports VLM and human preferences.
- The system requires no additional training, and the appendices provide substantial implementation information.
Critical weaknesses
- Quantitative results are limited to VLM and human visual-preference comparisons with a single baseline, HY-World 2.0. The numbers of test inputs and comparisons are not reported, so the evaluation scale cannot be assessed. The main paper primarily presents a few qualitative examples, without quantitative validation of final-map navigability, collisions, or exploration coverage.
- Consistency when revisiting generated regions is not evaluated. The world model uses the latest 16 frames as visual context. Although the paper claims to construct a persistent map, it does not specifically test whether scene structure remains consistent upon revisiting generated regions or whether the reconstructions before and after a revisit align correctly.
- The module ablations provide examples without quantitative results. Figure 3 only shows one scene for each of floor estimation, collision detection, and rewind.
Questions for authors
- After generation, can authors conduct relevant experiments for path planning to prove "Navigable"?
- How many wins and comparisons produce the 80.00% and 83.33% entries in Table 1? How many inputs were used, how were they selected, which VLM version was used, and how many human participants took part?
- Can the three modules be disabled individually on the existing input set, reporting collision rates, stalled steps, and coverage under the same exploration budget?
- When continued exploration brings the agent back to a previously generated region, are newly generated frames consistent with the existing map? How is this consistency evaluated, and how are any conflicting reconstructions handled?