How Researchers Trained AI to Simulate ISS Rendezvous Maneuvers - Space Portal featured image

How Researchers Trained AI to Simulate ISS Rendezvous Maneuvers

Linking up with a space station hurtling at 17,500 mph demands extraordinary precision—imagine merging onto a freeway blindfolded. A new AI approach c...

By Dr. Stella Cosmos
September 24, 2026

Stanford Engineers Teach Spacecraft to "Dream" Their Way to the Space Station

Docking with the International Space Station may appear deceptively straightforward when viewed from the polished highlight reels of mission control celebrations. In reality, it is one of the most demanding feats of precision engineering ever attempted by humanity. Imagine hurtling down a highway at 28,000 km/hr and attempting to parallel park into an open garage on a multi-billion dollar laboratory traveling at precisely the same velocity — all while the laws of orbital mechanics conspire to make every intuitive instinct wrong.

In the counterintuitive realm of orbital mechanics, firing your thrusters forward doesn't accelerate you toward your target — it raises your orbit and causes you to drift above it. There is no air friction to naturally bleed off excess speed. Sensor systems can be blinded in an instant by a glinting solar panel or a sudden eclipse. And the stakes could scarcely be higher: a miscalculated impact between two spacecraft doesn't just endanger the crews on both vehicles — it risks generating a catastrophic Kessler Syndrome-triggering debris field capable of wiping out dozens of operational satellites and even raining hazardous fragments down on populated areas below.

For decades, aerospace engineers have navigated these extraordinary challenges through a combination of meticulously hard-coded physics equations and the skilled hands of human pilots. But a compelling new paper from researchers at Stanford University, currently available in pre-print on arXiv, proposes a fundamentally different approach — one inspired, remarkably, by the way the human mind dreams.

The Challenge of Rendezvous and Proximity Operations

To appreciate why this new research matters, it helps to understand what engineers call Rendezvous and Proximity Operations (RPO) — the complex choreography of bringing two spacecraft together in orbit. This is not a single maneuver but a carefully sequenced series of orbital adjustments, sensor sweeps, and corrective burns that can unfold over hours or even days.

Traditionally, spacecraft navigation computers rely on a class of algorithms known as Guidance, Navigation and Control (GNC) systems. These systems integrate data from multiple sources — GPS receivers, star trackers, inertial measurement units, and LiDAR rangefinders — to build a continuously updated picture of the vehicle's position, velocity, and attitude in space. A mathematical workhorse called the Extended Kalman Filter (EKF) sits at the heart of most modern GNC systems, elegantly fusing noisy, uncertain sensor data into reliable state estimates and translating them into precise thruster burn commands.

The Extended Kalman Filter has served spaceflight remarkably well since its origins in the Apollo program, but it has a significant blind spot: it is poorly suited to processing high-bandwidth visual data. As spacecraft close to within a few hundred meters of their target, visual information becomes increasingly critical — and increasingly complex. This is where engineers have traditionally turned to dedicated computer vision algorithms to track the target spacecraft's pose (its position and orientation in three-dimensional space).

Unfortunately, space is a hostile environment for conventional computer vision. The extreme contrast between sunlit surfaces and deep shadow, the unpredictable glint of sunlight off reflective solar arrays, and the absence of any stable background reference can all cause traditional visual tracking algorithms to fail catastrophically. A system that works perfectly in one lighting condition can become completely disoriented seconds later as the geometry of sunlight shifts.

Why Reinforcement Learning Wasn't Enough

Recognizing these limitations, researchers began exploring whether Artificial Intelligence — specifically Reinforcement Learning (RL) — could do better. Reinforcement Learning is the same class of algorithms that famously powered computers to superhuman performance in games like chess, Go, and Dota 2. In RL, an agent learns by interacting with an environment, receiving rewards for desirable outcomes and penalties for undesirable ones, gradually refining its policy over millions of simulated trials.

Applied to spacecraft proximity operations, RL showed genuine promise. Agents could learn to interpret visual data and execute docking maneuvers in simulated environments without having physics equations explicitly programmed into them. But a fundamental limitation quickly became apparent: RL agents are brittle. They excel within the specific scenario distribution they were trained on, but generalize poorly to novel situations.

Change the rules — such as repositioning the target docking port to an unfamiliar location on the ISS, or introducing an unexpected obstacle — and a trained RL agent can be rendered almost completely ineffective, having no internal model of why physics works the way it does, only a memorized map of what actions worked in the past.

This brittleness is more than an academic inconvenience. Real-world spacecraft operations are saturated with unexpected events: a visiting vehicle docked where the approach corridor should be clear, a malfunctioning thruster introducing an unanticipated attitude drift, or a sensor glitch corrupting a critical measurement. Any autonomous system intended for operational use must be capable of reasoning about novel situations — not just pattern-matching against a memorized playbook.

Enter the "World Model": Teaching AI to Dream

The Stanford team's solution draws on an exciting frontier of machine learning research known as World Models. The concept can be illustrated through a deceptively simple analogy: when an outfielder sprints across a baseball diamond to catch a fly ball, they are not consciously solving the differential equations of projectile motion. Instead, they rely on a rich internal simulation — a mental model built from a lifetime of visual experience — that continuously predicts where the ball will be and guides their movements accordingly.

In the language of machine learning, a World Model is an AI architecture that learns a compressed, latent representation of the environment's underlying physics directly from raw sensory experience. Rather than having equations explicitly programmed in, or learning a simple stimulus-response mapping through reinforcement, a World Model builds an internal simulator that can be used to "imagine" — or dream — potential future states of the world. This internal simulator can then be queried rapidly and cheaply to evaluate the likely consequences of different actions before any of them are actually executed.

The concept has deep roots in neuroscience as well. Research into the hippocampus and prefrontal cortex suggests that biological brains use similar mechanisms during sleep — replaying and recombining experiences to build more robust internal models of causal reality. The analogy between AI "dreaming" and biological dreaming, while imperfect, captures something genuinely meaningful about the computational strategy involved.

The Out-of-this-World Model (OWM)

The Stanford team's implementation of this concept is called the Out-of-this-World Model (OWM). Built on the framework of Dreamer — a prominent World Model architecture from the machine learning literature — the OWM ingests raw visual data from a simulated spacecraft camera and learns to build an internal latent model of the spacecraft's dynamics and its relationship to the target docking port.

Crucially, the OWM doesn't just generate a single predicted future — it generates dozens of probabilistic future trajectories simultaneously, evaluating each one for its likelihood and desirability. This probabilistic approach is essential for robust autonomy: rather than committing blindly to a single predicted outcome, the system continuously weighs the uncertainty in its predictions and selects actions that perform well across a range of plausible futures.

  • Probabilistic dreaming: The OWM generates and evaluates multiple potential futures simultaneously, selecting actions that are robust to uncertainty.
  • Latent-space planning: All simulation occurs in a compressed internal representation, making it dramatically faster than simulating full physics in real time.
  • Visual input: The model processes raw camera imagery directly, without requiring hand-engineered feature extractors.
  • Model-based adaptation: Because the world model captures underlying physics rather than specific scenarios, it generalizes more effectively to novel docking configurations.

The result is a system that can "step" its way toward a target docking configuration by continuously imagining plausible futures, selecting the actions most likely to achieve the desired outcome, executing them, observing the result, and updating its internal model accordingly — a closed loop of perception, imagination, and action that mirrors in intriguing ways how biological intelligent systems navigate complex environments.

AstroJAX: Solving the Training Bottleneck

Teaching a World Model is a computationally intensive undertaking. Generating the hundreds of thousands of simulated flights needed to train the OWM would have required weeks of continuous computation on a conventional CPU-based system — an unacceptable bottleneck for iterative research and development.

To solve this problem, the Stanford researchers developed a custom simulation library they named AstroJAX, built on top of Google's JAX framework for high-performance numerical computing. AstroJAX is specifically designed to run astrodynamics simulations on Graphics Processing Units (GPUs) — the massively parallel processors originally developed for video gaming that have since become the workhorse hardware of modern AI research.

By exploiting the GPU's ability to execute thousands of parallel computations simultaneously, AstroJAX enabled the team to run vast numbers of simulated docking scenarios in a fraction of the time that would have been required on conventional hardware. This infrastructure innovation was not merely a convenience — it was a prerequisite for making the entire research program feasible.

Results: A Promising but Imperfect Leap Forward

The performance of the OWM, when evaluated against a comparable Reinforcement Learning baseline, was striking on several dimensions:

  • Training efficiency: The OWM required only 500,000 training iterations to master docking maneuvers, compared to 25,000,000 for an equivalent RL system — a 50-fold improvement in sample efficiency.
  • Generalization: The OWM performed significantly better when asked to dock at previously unseen docking port configurations on the ISS, demonstrating the real-world value of its internalized physical understanding.
  • Robustness to perturbation: When researchers deliberately introduced an unexpected obstacle — a docked capsule blocking the intended approach corridor — the OWM adapted more gracefully than the RL baseline.
  • Overall docking success rate: Across all ISS docking ports, the OWM achieved a 53% success rate, compared to the RL baseline's 29% — a substantial improvement, though one that underscores how much further development is required before operational deployment.

The most persistent challenge for the OWM emerged during close-range final approach, where success rates fell disproportionately. The researchers attribute this partly to the high penalty weighting applied to collision events during training — a deliberate conservative choice that may have made the model excessively cautious at close range. Adjusting this weighting in future iterations is expected to improve close-approach performance meaningfully.

"The OWM's ability to generalize to novel docking configurations with minimal additional training is perhaps its most significant capability," the authors note — a property that would be invaluable in real operations where conditions rarely match training scenarios perfectly.

Broader Implications for the Future of Space Autonomy

The significance of this research extends well beyond the specific problem of ISS docking. The coming decades will see an unprecedented expansion of human and robotic activity in Earth orbit and beyond, driven by commercial space stations, satellite servicing missions, lunar Gateway construction, and deep-space exploration initiatives. All of these endeavors share a common need: spacecraft that can reliably operate in close proximity to other objects without continuous human oversight.

NASA's On-orbit Servicing, Assembly, and Manufacturing (OSAM) program and the ESA's ClearSpace-1 active debris removal mission are just two examples of near-term projects that will demand robust, generalizable autonomous proximity operations capabilities. As the number of active satellites surpasses tens of thousands in low Earth orbit — a trend driven by commercial megaconstellations — the ability to autonomously service, relocate, or de-orbit spacecraft will transition from a luxury to a necessity.

World Model approaches like the OWM offer a compelling pathway toward the kind of flexible, physics-aware autonomy these missions will require. Unlike rigid rule-based systems or narrowly trained RL agents, a well-developed World Model carries within it a generalizable understanding of how objects move in space — an understanding that can, in principle, be transferred across a wide range of vehicles, targets, and operational scenarios.

The Road to Operational Deployment

Despite the genuine promise of these results, significant challenges remain before AI systems like the OWM could be trusted with operational spacecraft proximity operations — and especially before they could be deployed on crewed missions. A 53% success rate, while impressive relative to the RL baseline, falls far short of the near-perfect reliability that human life demands. Real-world sensor noise, hardware failures, atmospheric drag variations, and the irreducible complexity of orbital environments all represent additional hurdles that simulation-based training may not fully capture.

The path forward will likely involve hybrid architectures that combine the physical reliability of traditional GNC systems with the adaptive intelligence of World Models — each compensating for the other's weaknesses. It will also require extensive testing in hardware-in-the-loop simulators, followed by carefully graduated operational trials on uncrewed missions before any human lives are placed in an AI's hands.

But the trajectory is clear. With the increasing cadence of satellite launches, the growing commercial imperative for on-orbit servicing, and the long-horizon ambitions of human deep-space exploration, the question is not whether spacecraft will one day hand proximity operations over to autonomous AI systems — it is when, and how well those systems will perform when the moment comes. Research like Stanford's OWM represents an important and encouraging step along that path.

Learn More