Efficient Onboard Spacecraft Pose Estimation with Event Cameras and Neuromorphic Hardware
Authors: Arunkumar Rathinam, Jules Lecomte, Jost Reelsen, Gregor Lenz, Axel von Arnim, Djamila Aouada
Organizations: University of Luxembourg · Fortiss GmbH, Munich, Germany · Technical University of Munich, Munich, Germany · Paddington Robotics, London, United Kingdom
Reliable relative pose estimation is a key enabler for autonomous rendezvous and proximity operations, yet space imagery is notoriously challenging due to extreme illumination, high contrast, and fast target motion. Event cameras provide asynchronous, change-driven measurements that can remain informative when frame-based imagery saturates or blurs, while neuromorphic processors can exploit sparse activations for low-latency, energy-efficient inferences. This paper presents a spacecraft 6-DoF pose-estimation pipeline that couples event-based vision with the BrainChip Akida neuromorphic processor. Using the SPADES dataset, we train compact MobileNet-style keypoint regression networks on lightweight event-frame representations, apply quantization-aware training (8/4-bit), and convert the models to Akida-compatible spiking neural networks. We benchmark three event representations and demonstrate real-time, low-power inference on Akida V1 hardware. We additionally design a heatmap-based model targeting Akida V2 and evaluate it on Akida Cloud, yielding improved pose accuracy. To our knowledge, this is the first end-to-end demonstration of spacecraft pose estimation running on Akida hardware, highlighting a practical route to low-latency, low-power perception for future autonomous space missions.
Figures & tables
Event
Keypoint
Pose Metrics
Rej
Rep.
PCK↑
ET[m]↓
ETnorm[.]↓
ER [°] ↓
Epose[.]↓
V1 Float Models
LNES
0.9424
0.1114
0.0128
1.3793
0.0367
1
E2F
0.9439
0.1065
0.0122
1.3759
0.0361
1
2DHist
0.9303
0.1018
0.0117
1.4752
0.0373
0
Quantized
Table 1: Comparison of pose estimation performance of AKIDA V1 models across different event representation.
Event
Keypoint
Pose Metrics
Rej
Rep.
PCK↑
ET[m]↓
ETnorm[.]↓
ER [°] ↓
Epose[.]↓
V2 Float Models
LNES (5)
0.9669
0.0719
0.0080
0.9084
0.0236
0
LNES (8)
0.9669
0.0704
0.0078
0.8301
0.0220
3
E2F (5)
0.9684
0.0722
0.0080
0.9252
0.0239
0
E2F (8)
0.9684
0.0690
0.0076
0.8261
0.0217
3
Table 2: Comparison of pose estimation performance of AKIDA V2 models across different event representation
Event
Dynamic Power
Idle Power
Speed
Rep.
[mW] ↓
[mW] ↓
[fps] ↑
Jetson Orin Nano (Floating Point)
LNES
2491
3997
321
E2F
2519
3983
330
2DHist
2650
4009
349
Akida V1 (Deployed Quantized)
Table 3: Comparison of power and latency metrics for Akida V1 and Jetson Orin Nano. Tests are run for 4000 samples per representation and average values are given.
Figure 1: Cumulative distribution of SPEED scores on the SPADES dataset. (a) shows the impact of quantization on various representations for Akida V1, while (b) compares the performance of quantized Akida V1 and V2 architectures.
Figure 2: Qualitative results on LNES. Rows show samples with high (a-b), median (c-d), and low (e-f) SPEED scores.
Figure 3: Comprehensive Quantization and Error Analysis: Top two rows display the performance impact of quantization across different representations. Bottom two rows Mean localization and orientation error binned by ground-truth object distance for full-precision (FP) and quantization-aware trained (QAT) models between V1 and V2 architectures. Shaded regions denote ± 1 standard deviation.
Figure 4: Scatter distribution of per-sample localization error versus orientation error for full-precision (FP) and quantization-aware trained (QAT) models across Akida-V1 and Akida-V2 architectures. Each point represents one test sample, coloured by input representation.
Spacecraft pose estimation recovers a spacecraft’s position and orientation from images and other sensor data for navigation, rendezvous, and servicing.