Spacecraft Pose Estimation Benchmark

Mission Critical

Odin v0 calculates a spacecraft's 6-DoF pose in 0.82 ms — 7.8× faster than TensorRT on Jetson GPU (6.4 ms). System power draw is 30.4% lower (10.3 W vs 14.8 W). Timing jitter is 43.6× tighter, enabling stable Kalman filter integration for GNC control loops.

Live Comparison: Jetson vs. Odin v0

Left: Jetson Only | Right: Odin v0 Hybrid — Playback at 0.25× speed

Key Metrics

Benchmark: 2,880 frames · SPIN synthetic dataset · Pose-ResNet50

Metric	Odin v0	NVIDIA Jetson	Delta
AI Model Inference Latency	0.82 ms	6.4 ms	7.8× faster
End-to-End Latency	7.19 ms	23.04 ms	3× faster
End-to-End Throughput	139.1 FPS	43.4 FPS	3.2× higher
System Power	10.3 W	14.8 W	30.4% lower
Latency Jitter (σ)	0.40 ms	17.44 ms	43.6× tighter

---

Why Determinism Matters for Docking

In autonomous proximity operations, the pose estimate feeds a Kalman Filter. Latency variance (jitter) degrades filter accuracy, causing "slippage" in the control loop.

Odin's 0.40 ms jitter (vs. 17.44 ms) provides a stable temporal reference for the GNC system, enabling faster and safer approach maneuvers.

---

Pipeline Architecture

Stage Breakdown (per frame)

Phase	Latency (ms)	% of Total
H2D Transfer	0.04	5%
Model Inference	0.59	84%
D2H Transfer	0.00	1%
Post-processing	0.08	10%
Total	0.71	100%

---

Conclusion: Benefits for Space DPUs

• Kalman Integration: 0.40 ms jitter is below the noise floor of standard MEMS IMUs — pose estimate models as a fixed delay with no adaptive timestamp correction.
• Power Compliance: At 10.3 W, the system fits within a standard 15 W small satellite payload allocation. 30.4% lower than Jetson-only at 14.8 W.
• Scalability: Deeper backbones (ResNet101+) scale linearly while PCIe overhead stays constant.