Datasets:
How far ahead can you predict on DROID? A measured horizon curve
I ran a "predict that nothing changes" baseline against this dataset. Copy the last observed frame and call it the prediction to see how fast a trivial predictor falls behind. Sharing the curve because it says something useful about evaluating on DROID.
Setup: 64 rollouts, 3 context frames, 48-step horizon, scored on the moving regions only (a static background otherwise inflates every pixel metric). Ran on a laptop, no GPU.
Dynamic-region SSIM over the horizon:
| step | 1 | 3 | 6 | 12 | 18 | 24 | 28 | 36 | 47 |
|---|---|---|---|---|---|---|---|---|---|
| SSIM | 0.873 | 0.797 | 0.676 | 0.446 | 0.350 | 0.260 | 0.204 | 0.192 | 0.216 |
Three regimes:
- steps 1–3 — near-perfect; a trivial baseline and a real model score the same
- steps 4–24 — error climbs steeply and monotonically; this is where models are separable
- step 28+ — floored at ~0.20 SSIM and ~10.3 dB PSNR, oscillating with no trend; the prediction has decorrelated and everything ties again
So the horizon worth evaluating on here is roughly 8–24 steps. Shorter and models tie near perfect; longer and they tie at the floor.
For contrast, the same run on a 30fps SO-101 recording gives a flat dynamic-region curve over 6 steps (0.972 → 0.950, wandering rather than falling). DROID's 15fps and more varied scenes are what make it usable for this at all.
Left is the prediction, right is what actually happened, same 48 steps.
On the camera-perspective thread - since a viewpoint switch mid-episode would show up as model error when it isn't, I checked for it. Across 16 sampled episodes, 2 contain a frame-to-frame jump 8–10x their own median.
I inspected the largest (ep=0, step 28) frame by frame and it's continuous fast camera approach, not a viewpoint change. The same objects persist and move consistently across the jump. So I did not reproduce the alternation described in that thread in this sample, though 16 episodes is a small sample and I'd welcome a pointer to a specific episode where it happens.
Caveats: n=64. Scores aggregate with the interquartile mean and bootstrap CIs, so a minority of outlier episodes doesn't drive the numbers. The pixel metrics are mutually consistent but LPIPS disagrees with them on the cross-dataset comparison. And this is a trivial baseline, so it marks where a do-nothing predictor becomes separable. A real model stays correlated longer and would push the upper end out.
Tooling is open source if anyone wants to reproduce it or run it on their own data: https://github.com/BuceaGeorgia/worldproof