Temporal synthetic spatial-perception calibration

Labnote 006: Temporal synthetic spatial-perception calibration🔗

Question🔗

Can a compact temporal model infer useful local spatial state from four low-resolution grayscale observations plus ego-motion, and does it outperform a parameter-matched model that sees only the final frame?

Method🔗

Blender 5.2.2 generated 384 deterministic four-frame sequences: 256 train, 64 validation, and 64 held-out test scenes with disjoint seed ranges. Each frame contained a 64 by 64 grayscale observation, an exact 16 by 16 ray-cast depth target, a depth-discontinuity mask, and ego-motion. File digests, tensor shapes, label derivation, and the complete split manifest were validated and frozen before training.

The matched 171,682-parameter conditions used the same encoder, ConvGRU, and prediction heads. The single-frame condition received only the final frame; the temporal condition received all four observations and ego-motion. Both trained for 20 epochs with seed 1701 on an RTX 4070. The test split was evaluated once after both runs finished.

Results🔗

The temporal condition modestly improved depth error and discontinuity detection while modestly worsening free-space overlap. The run completed with exit code zero and empty stderr. The frozen result and scheduler receipt are retained under experiments/labnote_006/.

Interpretation🔗

This is a mixed result. Short temporal context was useful for two of the three precommitted spatial metrics, but did not dominate the single-frame baseline. The result supports continuing to a separately frozen transfer probe rather than treating temporal memory as an unconditional improvement.

Intended boundary🔗

The experiment predicts local depth, free space, and depth discontinuities on held-out procedural geometry. It does not establish complete mesh reconstruction, general game understanding, navigation competence, or transfer to real game imagery. A cross-game probe requires a later frozen protocol.