Physical AI Vision Retrieval 025 - Raw Research Artifact

Generated at: 2026-07-16T16:07:45Z Classification: Research-only Dataset: lerobot/libero_10_image at main Encoder: google/siglip-base-patch16-224

Scope

This run uses real LIBERO front-camera frames, a real SigLIP encoder on an NVIDIA GPU, and the real ZeptoDB Agent Memory HTTP path. It measures task-level episode retrieval, not VLA action generation or simulator success.

Data And Runtime

Tasks	Memories	Held-out queries	Frame	GPU	Embedding dims	Peak GPU MiB	Encoder ms/image
10	190	100	middle	NVIDIA L40S	768	539.2	1.478

Retrieval Quality

Variant	Path	Recall@1	Recall@5	MRR
image	local cosine	0.970	1.000	0.985
image	ZeptoDB	0.700	0.980	0.810
image_text	local cosine	1.000	1.000	1.000
image_text	ZeptoDB	0.900	1.000	0.926

ZeptoDB Latency

Variant	Insert p50 ms	Insert p95 ms	Search p50 ms	Search p95 ms
image	1.261	1.858	1.037	1.362
image_text	1.622	2.168	1.032	1.361

Acceptance

Criterion	Status
10 held-out queries per task	pass
Real CUDA encoder produced embeddings	pass
Image + instruction ZeptoDB Recall@1 >= 0.80	pass
Image + instruction ZeptoDB Recall@5 >= 0.95	pass
All 10 tasks represented	pass
Image + instruction ZeptoDB search p95 < 30 ms	pass
Temporary AWS resources deleted	pass

Result

Overall status: pass.

Interpretation

A pass proves that real visual embeddings can be created on the EKS GPU path and retrieved through ZeptoDB within this bounded workload. It does not prove faster or more accurate VLA policy inference; that remains the next experiment.