Reflexes in the Dark

How Brain-Inspired Hardware Is Teaching Spacecraft to See
It is 2029. Two spacecraft drift toward each other at the edge of lunar orbit, roughly 400,000 kilometers from Earth. There is no GPS here. The radio lag back to mission control stretches to seconds. That lag is far too long for a human operator to issue corrections in real time. One craft must dock with the other autonomously, using only what its onboard sensors can perceive.
The sunlit side of the target vehicle blazes white. The shadowed side is pure black. A standard camera would burn out on the highlights or lose the target in the darkness. Onboard one of those spacecraft, however, a different kind of sensor fires. It does not take photographs. Instead, it reacts, pixel by pixel, to every shift in light. Tens of thousands of micro-signals race into a chip the size of a thumbnail. Spiking neurons then process the geometry of the scene in milliseconds, consuming just milliwatts of power.
That sensor is an event camera. That chip is a neuromorphic processor. As of 2026, a group of researchers spanning four institutions (the University of Luxembourg, Fortiss GmbH, the Technical University of Munich, and Paddington Robotics) has shown this combination can run a full spacecraft pose-estimation pipeline in real time, end to end, on commercially available neuromorphic silicon.
A Different Kind of Sight
To understand why this matters, it first helps to understand why cameras struggle in space.
Standard frame-based cameras capture the full brightness of every pixel at fixed intervals. Think of a cinema film recording 24 still images per second. This approach works well in stable lighting. Space, however, is not stable. The sun in orbit produces light roughly ten times more intense than sunlight at Earth’s surface. There is no atmosphere to soften it. A spacecraft in the foreground can appear blinding white against a target vehicle whose shadowed face is jet black. When a fast-moving debris object tumbles into frame, the shutter speed needed to freeze its motion darkens the background. When the shutter slows to capture dim areas, the moving object becomes a blur.
These effects are not edge cases. They are the routine operating conditions any proximity operations system must contend with.
There is also a deeper structural issue. Frame-based sensors suffer from High Dynamic Range constraints. In financial terms, a frame-based camera effectively goes bankrupt at extreme contrasts. It blows out highlights and destroys the signal it was meant to capture. The sensor cannot simultaneously expose for the blazing sunlit hull and the jet-black shadow behind it. One side of the equation always collapses.
Event cameras solve this at the physics level. Rather than capturing every pixel at a fixed interval, each pixel independently fires an output event the moment its light level changes by more than a set threshold. The result is a stream of timestamped change signals, not a sequence of frames. Because the background of space is static, a near-perfect black void, those signals represent almost exclusively the motion of the target spacecraft. Noise is reduced at the hardware level. Latency is measured in microseconds rather than milliseconds. Furthermore, because each pixel reacts to light change rather than to absolute brightness, the sensor is not blinded by specular highlights or confused by hard shadows. It cares only about the delta, not the scale. In that sense, it is an inflation-proof sensor in an environment where frame-based systems routinely go bankrupt.
Diagram 1 · Section: A Different Kind of Sight
How Sensors Handle Extreme Contrast in Space
The Power Budget Problem
Even if event cameras give a spacecraft better eyes, they still need a brain. In space, every watt spent on computation is a watt taken away from something else — propulsion, communications, life support, or thermal control. Engineers refer to this as the SWaP constraint, meaning size, weight, and power. It is one of the most severe engineering realities of spaceborne systems.
A high-end GPU capable of running a deep neural network in real time can consume hundreds of watts. For a cubesat, that figure is simply not available. For larger craft, the sustained draw of a conventional compute cluster running across a multi-month mission is a serious performance penalty in a fixed energy budget.
This is precisely where neuromorphic processors offer a genuinely different calculus.
Neuromorphic chips process information the way biological neurons do, through sparse event-driven spikes rather than dense synchronous matrix multiplications. When no new information arrives, the processor stays quiet. When a signal comes in, only the neurons relevant to that signal activate. Because space is mostly static and the target spacecraft occupies only a fraction of the sensor’s field of view, the vast majority of neurons in a neuromorphic chip spend most of their time dormant — and consuming almost nothing. In energy terms, this is the marathon runner of compute architectures. It burns fuel only when the terrain demands it.
Enter Akida
BrainChip’s Akida platform is one of the most mature neuromorphic inference chips commercially available. The AKD1000 converts conventional convolutional neural networks into event-driven spiking neural networks and executes them with low-precision weights and activations. Importantly, the chip already has space heritage. BrainChip reported a launch of Akida technology into low Earth orbit in 2024. That launch gives the platform a credibility that purely experimental architectures cannot claim.
The research group used Akida as the target deployment hardware for a full spacecraft pose-estimation pipeline. Their results, published in April 2026, are the first end-to-end demonstration of this capability on Akida hardware — moving the technology from simulation to measured, real-world silicon benchmarks.
What Is 6-DoF Pose Estimation?
Before going further into the engineering, it is worth being precise about what pose estimation actually means.
A spacecraft’s pose describes both where it is and how it is oriented in three-dimensional space. Position is defined by three spatial coordinates: up and down, left and right, forward and back. Orientation is described by three rotational angles, specifically pitch, yaw, and roll. Together, these six numbers form what engineers call the six degrees of freedom, or 6-DoF. Knowing the 6-DoF pose of a target vehicle is the prerequisite for any docking maneuver. Without it, the chaser spacecraft cannot know whether the target’s docking port is facing toward it, away from it, or somewhere between.
Diagram 2 · Section: What Is 6-DoF Pose Estimation?
Six Numbers That Define Where a Spacecraft Is — and How It Is Oriented
Target spacecraft in 3D space
3 translational axes — position
3 rotational axes — orientation
Computing 6-DoF pose from a single camera is a genuinely hard problem. There is no stereo baseline to infer depth from disparity. The system must reason about geometry entirely from the 2D projection of a 3D object onto a flat sensor. The standard approach, which the researchers adopted, is to first detect a set of known reference points on the target called keypoints and then apply a geometric solver to work backward from their 2D positions to the full 3D pose.
Three Representations, One Pipeline
The researchers worked with the SPADES dataset, a large-scale collection of simulated and real event streams captured specifically for spacecraft pose estimation. The dataset holds nearly 180,000 synthetic event frames drawn from 300 unique trajectories, as well as 15,500 real frames from a Prophesee EVK4-HD event camera. The synthetic portion was used for training and evaluation, divided into training, validation, and test sets of roughly 107,000, 36,000, and 36,000 frames respectively.
Because event cameras output a stream of individual events rather than images, some form of conversion is needed before a convolutional neural network can process the data. The researchers evaluated three such conversion strategies.
The first, Event-to-Frame or E2F, assigns each pixel the polarity of the most recent event within a time window. Positive means brightness increased. Negative means it decreased. A neutral default applies when no event occurred. The result is a sparse, polarity-coded image. The second, the 2D Histogram or 2DHist, counts the total number of events per pixel within the window and normalises the result. This captures activity density without preserving the direction of change. The third, Locally Normalised Event Surfaces or LNES, is the most temporally rich of the three. It accumulates events in two separate channels, one per polarity, while weighting each event by its normalised timestamp. The result encodes not just where events happened and their polarity, but roughly when within the accumulation window they occurred.
Each representation involves different trade-offs. E2F is sparse and computationally light. 2DHist captures activity density. LNES preserves more of the temporal structure of the event stream. Moreover, as the researchers found, whether that extra temporal information actually helps depends very much on the hardware running the inference.
Two Generations, Two Approaches
The pipeline uses a compact MobileNet-style neural network to predict the 2D image coordinates of eight keypoints corresponding to the corners of the target spacecraft. Once those keypoints are located in 2D, a standard geometric solver, specifically a Perspective-n-Point or PnP algorithm, recovers the full 6-DoF pose.
For Akida V1, the silicon generation available as physical hardware, the researchers used a direct coordinate regression approach. The network’s final layer outputs 16 scalar values — one x-coordinate and one y-coordinate for each of the eight keypoints. The model contains roughly 1.88 million parameters. Critically, Akida V1 operates at 4-bit weight and activation precision, the lowest quantization level tested in the study.
For Akida V2, evaluated via BrainChip’s cloud simulation platform, the researchers took a different route entirely. Rather than predicting keypoint coordinates directly, the V2 model outputs eight spatial heatmaps, one per keypoint, each representing a probability distribution over possible keypoint locations. The model uses an encoder-decoder architecture with transposed convolutions, totalling 1.72 million parameters. Akida V2 supports 8-bit quantization throughout, doubling the representational precision of V1.
The distinction between coordinate regression and heatmap regression turns out to matter enormously, and not just for accuracy. It determines how well each approach survives quantization — the process of compressing a neural network’s numerical precision for deployment on low-power hardware.
The Quantization Divide
Quantization-aware training, or QAT, is the process of training a neural network while simulating the numerical precision it will have at deployment. A network trained in 32-bit floating point, then naively converted to 4-bit integer arithmetic, can lose substantial accuracy. Small weight values get rounded to zero. Subtle gradient information is destroyed. QAT compensates by exposing the network to quantization noise during training, giving it a chance to adapt.
Even with QAT, the V1 models show a steep performance drop at 4-bit precision. In floating-point, all three event representations achieve keypoint accuracy scores above 93 percent. After 4-bit quantization, that figure drops to around 33 percent for E2F and 31 percent for 2DHist. The LNES representation effectively collapses. Its accuracy falls to just 20 percent. Over 2,300 out of 35,000 test samples are rejected because the solver cannot produce a valid pose estimate. The corresponding pose error for quantized LNES on V1 worsens by more than 594 percent relative to the floating-point baseline.
The V2 result, by contrast, is strikingly positive. After 8-bit quantization, the heatmap models do not degrade. They improve. The quantized E2F model on V2 achieves a SPEED score of 0.0208, which is slightly better than its floating-point counterpart. Quantized LNES on V2 achieves 0.0207. The keypoint accuracy for all three quantized V2 representations exceeds 97 percent. Rejection counts fall to single digits.
The researchers explain this contrast clearly. Coordinate regression is brittle to quantization because errors in the network’s weights translate directly into errors in predicted pixel coordinates. There is no buffering mechanism. Heatmap regression, on the other hand, transforms the problem into something closer to spatial classification. Instead of predicting a single coordinate, the network predicts a probability distribution across many spatial locations. Rounding errors degrade the sharpness of that distribution without necessarily shifting its peak by much. As a result, the heatmap approach absorbs quantization noise at an architectural level.
Diagram 3 · Section: The Quantization Divide
Why Heatmap Regression Survives Quantization — and Coordinate Regression Does Not
Comparing Akida V1 (4-bit direct regression) vs Akida V2 (8-bit heatmap regression)
float
QAT
float
QAT
Power and Speed on Physical Hardware
Beyond accuracy, the researchers benchmarked the Akida V1 hardware against an NVIDIA Jetson Orin Nano, a common reference platform for edge inference in aerospace and robotics. Running floating-point models via TensorRT, the Jetson achieves between 321 and 349 frames per second. Its dynamic power consumption sits between 2,491 and 2,650 milliwatts. Its idle power hovers around 4,000 milliwatts.
The AKD1000 achieves 16 to 17 frames per second. Dynamic power sits between 917 and 928 milliwatts. Idle power is 927 milliwatts.
The raw throughput gap is large, roughly twenty-fold in the Jetson’s favour. That gap, however, does not tell the whole story. The researchers note that the AKD1000’s interface with the host computer has not yet been optimised. The current throughput figure therefore does not reflect the chip’s intrinsic processing speed. The power figures, on the other hand, are direct measurements. The Akida V1 chip consumes three to four times less power in both active and idle states.
This brings into focus the more meaningful metric for space systems, specifically performance per watt, or the return on each unit of energy expended. By that measure, the Akida holds a substantial advantage. Where the Jetson delivers high throughput at high cost, the Akida delivers adequate throughput at a fraction of the energy. For a mission lasting months or years on a fixed solar and battery budget, that ratio is the figure that matters.
The Akida V2 cloud simulation shows a different throughput story. At the hardware’s native clock of 25 MHz, the V2 heatmap model achieves 2.18 to 2.33 frames per second. This is slower than V1 due to the more complex encoder-decoder architecture. Nevertheless, the researchers project that at 400 MHz, a plausible operating frequency for future hardware, the V2 model would reach approximately 37 frames per second. That is sufficient for real-time pose tracking during proximity operations.
Sparsity also plays a direct role in power consumption. The 2DHist and E2F representations produce sparser activation patterns than LNES. Consequently, fewer neurons fire per inference cycle and the dynamic power draw is lower. On a neuromorphic platform, sparsity is not just an abstract computational property. It is a direct dial on thermal output.
What the Errors Tell Us
Beyond aggregate metrics, the researchers provide a qualitative analysis pointing to illumination as the dominant driver of pose error. High-quality pose estimates consistently correspond to samples where spacecraft edges are clearly defined. Poor estimates, by contrast, correlate with high-noise regions, sparse event clusters at long range, and extreme illumination conditions that produce either saturated or near-empty event frames.
This finding is consistent with the original motivation for using event cameras. Even so, it suggests the remaining performance gap is not primarily an artifact of the network architecture or the neuromorphic deployment. Rather, it is a sensor-level phenomenon driven by the physics of how light interacts with spacecraft surfaces at varying distances and angles.
From Lab to Orbit
The broader context for this research is the growing operational demand for autonomous proximity operations. On-orbit servicing missions, active debris removal, and lunar rendezvous scenarios, including the complex docking maneuvers anticipated for NASA’s Artemis architecture, all require a chaser vehicle to estimate the pose of a non-cooperative target. There is no GPS available. There is no real-time ground support. And the lighting conditions are often exactly those where frame-based cameras perform worst.
SpaceX’s Dragon capsule already docks autonomously with the International Space Station. However, it operates in low Earth orbit under relatively predictable lighting. Lunar missions targeting the south pole of the Moon introduce far more severe conditions. The sun sits perpetually near the horizon, casting extremely long shadows. Much of the terrain in that region includes permanently shadowed craters that receive no direct sunlight at all. In those zones, a frame-based camera does not merely perform poorly. It may be functionally unable to see the target. A neuromorphic system built on event cameras, therefore, is not simply an alternative in that environment. It may be the only viable approach.
For a lunar Gateway station or a long-duration surface lander, onboard vision systems must operate across years with a fixed power budget. A sub-watt inference chip that stays dormant between events and wakes only when the scene changes is a meaningfully different engineering proposition than a GPU running at full load continuously.
The fact that Akida hardware has already been launched into orbit gives the platform a credibility that simulation-only results cannot match. Furthermore, the fact that the researchers ran their benchmarks on physical AKD1000 silicon rather than software models alone means the power and throughput figures are real measurements. They are not projections.
What Comes Next
The researchers are candid about the current state of the work. The V1 hardware shows that end-to-end execution is feasible and that power consumption is genuinely low. The V2 simulation shows that accuracy at neuromorphic precision can match floating-point performance. What remains is the convergence of these two results on physical V2 hardware. The authors plan to take that step as the chip becomes available.
Future directions they identify include recurrent architectures that would process not just individual event frames but the temporal sequence of events across multiple time steps. This would more fully exploit the temporal information encoded in the event stream. The authors also note the potential for tighter integration between the detection and keypoint regression stages, training them jointly rather than as sequential modules.
The deeper question this work raises is about the trajectory of onboard AI for space systems. The neuromorphic computing field has spent more than a decade building chips that process information sparsely, asynchronously, and at very low power. Space systems have simultaneously faced mounting pressure to perform more computation autonomously, closer to the sensor, without the weight or power overhead of conventional compute clusters. These two trajectories are now close enough to intersect. The SPADES dataset, the Akida hardware, and this group of researchers and engineers have made that intersection concrete and measurable.
The spacecraft of the near future will not radio Earth to ask what they are looking at. They will know — in milliseconds, on milliwatts, in the dark.

Responses