Motivation
Results
Yes. Removing it, or using another video's frames, takes Gain to about 0.
In the paper (appendix, Bank Contents) the 100 s model shows the same pattern. Web photographs, first 60 s, rollout seed 0:
Table 1: Bank contents at inference, 100 s model (paper). Starting-pose and mid-path returns; web photographs, first 60 s, rollout seed 0.
| bank contents (100 s, paper) | start pose, Revisit / Gain | mid-path, Revisit / Gain |
|---|---|---|
| Top-K (ours) | 0.377 / 0.213 | 0.440 / 0.319 |
| Empty | 0.176 / 0.053 | 0.105 / 0.019 |
| K random frames | 0.153 / 0.055 | 0.207 / 0.109 |
| K frames, another episode | 0.071 / 0.012 | 0.073 / 0.016 |
This response repeats the experiment for the 200 s and 400 s models (all returns together):
Table 2: Inference-time bank ablation on the 200 s and 400 s models (Revisit / Gain, first 60 s, 20 web-photo trajectories).
| model | top-6 retrieval | random bank | no bank | other video's bank |
|---|---|---|---|---|
| 200 s, Revisit / Gain | 0.467 / 0.359 | 0.228 / 0.122 | 0.121 / 0.041 | 0.138 / −0.002 |
| 400 s, Revisit / Gain | 0.396 / 0.285 | 0.184 / 0.088 | 0.083 / 0.013 | 0.132 / 0.005 |
Examples
Videos from this response's runs (400 s model), same first frame and camera path, 60 s. Only the bank content differs. Compare what each walk shows when it returns to an earlier place.
Walk 05
Walk 13
Walk 15
Setup
- Trained models unchanged; only the bank content at inference differs.
- 20 web-photo walks, first 60 s, one rollout seed. The top-6 column is the paper's five-seed mean.
Conclusion
- The 200 s and 400 s models read the bank. No bank, or another video's bank, takes Gain to about 0 (200 s: 0.359 → 0.041 / −0.002; 400 s: 0.285 → 0.013 / 0.005).
- Same-video random frames recover only about a third (Gain 0.122 and 0.088), so the retrieved frames matter, not just any frames from the video.
- This does not explain why 400 s is below 200 s. It shows the 400 s model uses its memory, nothing more.