Every ambitious idea in orbital industry — debris capture, satellite servicing, in-space manufacturing, asteroid prospecting — runs into the same wall: a human with a joystick cannot do the job. Light takes about a quarter of a second to make the round trip to geostationary orbit and back. To Mars, the one-way delay ranges from roughly 4 to 24 minutes depending on where the planets sit. A robotic arm reaching for a tumbling satellite, or a rover threading a boulder field, cannot wait for permission from Houston. The machine has to perceive, decide, and act on its own.
That is why we treat artificial intelligence and robotics not as one mission area among five, but as the enabling layer under all of them. It is also the area closest to what Space Dust Industries actually does today. While our orbital concepts remain vision, our AI and robotics work is real and running now: an on-premises multi-GPU research cluster, a fully offline edge AI stack, active reinforcement learning research, and a ground robotics project in development. We are building the brains before the spacecraft, because that is the correct order.
Autonomy is already flying
Spacecraft autonomy is not speculative — it is quietly becoming the norm, and the recent record is worth knowing.
NASA’s Perseverance rover drives itself across Mars using AutoNav, a self-driving system that builds 3D terrain maps from stereo images and plans safe paths while the rover is moving. It has covered nearly 700 meters in a single stretch without a human reviewing the route, and it lets Perseverance drive at close to its full mechanical speed — the rover completed a marathon’s worth of driving in less than half the time its predecessor Opportunity needed. The Ingenuity helicopter, designed as a technology demo for five flights, flew 72 times over almost three years in an atmosphere less than one percent as dense as Earth’s, navigating each flight entirely on its own using a downward-facing camera and onboard vision processing — teleoperating a helicopter across a multi-minute light delay was never an option.
The most dramatic demonstration came in September 2022, when NASA’s DART spacecraft struck the asteroid moonlet Dimorphos at over 6 kilometers per second. For the final four hours, a system called SMART Nav took over completely: it picked the target out of the camera feed, distinguished the small moonlet from the larger asteroid beside it, and steered itself to impact. No human could have closed that loop in time.
Closer to home, robots already work inside the International Space Station. NASA’s three Astrobee free-flyers — cube-shaped robots that navigate the station using vision-based localization — perform surveys, logistics tasks, and serve as a testbed for autonomy research, following the earlier Robonaut 2 humanoid experiment. And in geostationary orbit, Northrop Grumman’s Mission Extension Vehicles proved commercial on-orbit servicing is real: MEV-1 docked with the Intelsat 901 satellite in 2020, extended its life for five years, and in April 2025 completed the first-ever undocking between two commercial spacecraft in GEO before moving on toward its next client. DARPA’s robotic servicing program pushed the same frontier with dexterous robotic arms designed to inspect and repair satellites that were never built to be touched.
Deep reinforcement learning trains control policies through millions of simulated trials — an approach we run on our own GPU cluster today.
Learning to fly: reinforcement learning for spacecraft control
Classical spacecraft control is built on precise models and pre-planned sequences. That works beautifully when the world cooperates. It breaks down when the target is a dead satellite tumbling unpredictably, a landing site turns out rockier than the maps showed, or a thruster underperforms mid-maneuver.
Deep reinforcement learning offers a different path. Instead of hand-coding every contingency, an agent learns a control policy by trial and error across millions of simulated episodes — approaching a tumbling object, docking under sensor noise, guiding a powered descent — and is rewarded for outcomes rather than for following a script. This is an active and serious research field: work at NASA, universities, and industry labs has demonstrated learned policies for six-degree-of-freedom docking, planetary landing guidance, and constellation management in high-fidelity simulation, and learned components are beginning to appear alongside classical guidance in flight software research.
The hard problem is the one the field calls sim-to-real transfer: a policy that is superhuman in simulation must survive contact with reality, where dynamics, lighting, and sensor behavior never quite match the training environment. The standard tools — domain randomization, high-fidelity physics, hardware-in-the-loop testing, and runtime safety monitors that bound what the learned policy is allowed to do — are exactly the kind of patient, ground-based engineering that can be done without a launch. That is where our reinforcement learning research lives today: training and evaluating control policies on our own cluster, in simulation environments we control end to end, with a ground robotics platform in development as the physical proving ground. Verification and validation of learned systems — proving to a mission assurance board that a neural network will behave — is, in our view, as important a research problem as the learning itself, and we treat it that way.
Intelligence at the edge, where there is no cloud
There is no datacenter in orbit. A spacecraft’s computer must run every model it needs on board, within a strict power budget, through radiation that corrupts memory and degrades electronics, with no option to fall back to an API when the link drops. Space is the ultimate offline edge environment.
The industry is actively working this problem from both ends. Radiation-hardened processors like the RAD750 have flown for decades but offer a tiny fraction of the compute of a modern phone. The newer approach flies commercial hardware with software-level fault tolerance: HPE’s Spaceborne Computer-2 has operated commercial servers with GPU acceleration aboard the ISS since 2021, cutting some experiment downlink volumes dramatically by processing data in orbit instead of shipping raw files to Earth. Multiple missions and experiments have now flown Nvidia Jetson-class embedded GPUs on small satellites to run neural network inference in orbit, and next-generation rad-tolerant processors such as the High Performance Spaceflight Computing (HPSC) effort aim to deliver orders-of-magnitude more onboard compute for future NASA missions. The lesson from all of it: onboard AI turns a bandwidth problem into a compute problem, and compute is the side that is improving fastest.
This is the domain where SDI’s current work maps most directly onto spaceflight. We build and operate a fully offline AI stack — models selected, quantized, and served to run on constrained local hardware with no external dependency whatsoever. The engineering disciplines are the same ones flight software demands: making models small without making them wrong, guaranteeing behavior when inputs go out of distribution, and designing systems that degrade gracefully instead of failing silently. An AI system that genuinely needs no cloud is rare in today’s industry. In orbit, it is the only kind that works.
Multi-agent research asks how fleets of spacecraft can divide work, share findings, and tolerate the loss of individual members.
From one robot to many
The next leap after single-spacecraft autonomy is coordination. Large constellations already demand it — operators managing thousands of satellites rely on automated collision-avoidance and stationkeeping because no human team can hand-fly that many vehicles. Research programs at NASA, ESA, and universities are pushing further, toward true multi-agent systems: swarms of small spacecraft that map an asteroid from many angles at once, distributed telescopes that synthesize a single large aperture, and servicing fleets where inspection, capture, and refueling are handled by different specialized vehicles.
The interesting problems are decentralized by nature. Agents must reach consensus on a plan over intermittent, delayed links; reassign roles when a member fails; and share what they learn without flooding the network. Multi-agent reinforcement learning and classical distributed algorithms both contribute, and the frontier is combining them into systems that are provably safe. Our own multi-agent research runs on the ground first — coordinating simulated fleets, and increasingly the agentic AI systems we build for terrestrial work, where multiple autonomous software agents plan, delegate, and verify each other’s output. The architecture questions are strikingly similar whether the agents are processes in a cluster or spacecraft in a formation, which is precisely why we consider our current agentic systems work a direct investment in this mission area.
Planning, replanning, and navigating without help
Two more threads complete the autonomy stack we are developing.
The first is adaptive mission planning. A mission plan is a living object: instruments fail, opportunities appear, fuel margins shift. AI planners — a lineage that runs from NASA’s Remote Agent experiment on Deep Space 1 in 1999 through the onboard schedulers being developed for current rovers — can re-optimize a mission against thousands of constraints in seconds, weighing science return against risk and resources. Our concept work focuses on human-on-the-loop planning: the AI generates and continuously repairs plans, while people set objectives and constraints and retain authority over irreversible decisions. That division of labor mirrors how we already build agentic systems on the ground.
The second is autonomous navigation. Between AutoNav’s terrain-relative driving, SMART Nav’s terminal guidance, the crewed Orion vehicle’s optical navigation backup, and deep-space experiments in pulsar-based and celestial navigation, the toolkit for finding your way without constant contact from Earth is maturing fast. Our research interest is in fusing these modalities — optical, inertial, ranging — with learned components that recognize terrain and targets, so that a spacecraft arriving at an uncooperative object can build its own map and plan its own approach. The same perception problem, at smaller scale, is central to our 3D scanning work today: turning raw sensor data into geometric truth a machine can act on.
Our earlier concept studies also sketched the destinations this autonomy serves — orbital transfer and servicing stations that act as logistics hubs, autonomous prospecting spacecraft with AI-driven spectral analysis for asteroid surveys, and long-duration probes for environments too distant or hostile for teleoperation. These remain concepts, and we present them as such. What makes them worth sketching is that every one of them is gated on the same technology: robots that can be trusted to work alone.
The Road Ahead
We are candid about the sequence. Space Dust Industries has not flown hardware, and the spacecraft described here are vision. But this mission area is different from the other four in one respect: the core technology is not waiting on a launch. Autonomy is software, and software is built, trained, broken, and rebuilt on the ground — which is exactly what we are doing now, as a working AI company supporting NASA-connected work as a subcontractor. Our multi-GPU research cluster trains the models. Our offline edge stack proves they can run untethered on constrained hardware. Our reinforcement learning research attacks the control problem, our agentic systems work attacks the coordination problem, and our robotics project gives the software something physical to move. Every one of those threads is a direct, honest step toward machines that can work in a place where no one can reach over and take the controls. Intelligence first. Then orbit.