eetmie/xvla-base-onnx

X-VLA 0.9B split ONNX

Twelve FP16 ONNX graphs derived from lerobot/xvla-base, traced with LeRobot 0.6.1. Unmodified base weights — a starting point to fine-tune your own robot on, not a policy that drives anything as it stands. Split into twelve so each TensorRT engine builds within an 8 GB Jetson Orin Nano.

Shapes are static in the graphs, so changing one means re-exporting: 3 image-view slots at 50 tokens each, 50-token language budget, 30-action chunk, 10 denoise steps, 20-dim state and action in ee6d. All three view slots are exported valid, which is what keeps this general — a 1- or 2-camera robot zero-pads the rest at no cost.

Contents

file	what
vision_0..3.onnx	DaViT vision tower
text_encoder_0..2.onnx	Florence-2 text encoder
cond.onnx	conditioning assembly
denoise_0..3.onnx	policy transformer, 24 blocks as 6+6+6+6
bundle.json	graph inputs/outputs and the shapes above
MANIFEST.sha256	SHA-256 of every graph and of bundle.json
tokenizer/	facebook/bart-large @ cb48c1365bd826bd521f650dc2e0940aee54720c — the text encoder takes token ids

Retained Orin Nano result

Pinned clocks (MAXN_SUPER), JetPack R39.2.1, all three view slots fed a deterministic procedural scene, FP16, ONNX Runtime 1.24.0 on TensorRT 10.16.2.10 — all twelve graphs on TensorRT:

p50	p95	rate	resident
415.94 ms	418.01 ms	2.40 Hz	4741 MB

Denoising is 295.5 ms of that and vision 111.8 ms.

Parity

Export-time CPU parity measured cosine 0.999993 for actions. There is no on-device cross-backend measurement yet: the PyTorch reference harness hits a sequence-length mismatch before inference. Treat the export-time figure as supporting evidence until that audit is complete.

The latency row above also comes from the procedural observation source rather than camera frames — sound for timing, since the transformer does the same work whatever the pixels are, but it says nothing about predicted action values.

Runtime, limitations, and reproducible benchmark: jetson-orin-nano-vla.