Space Robotics

Space is the domain where the two assumptions of everyday robotics fail at once. There is no repair: a rover on Mars cannot be serviced, and a satellite that fails in geosynchronous orbit is written off or, at best, visited once by a purpose-built spacecraft. And there is no live supervision: a signal between Earth and Mars takes minutes to arrive and minutes to return, so no operator closes a control loop in real time. Everything this wiki's manipulation modules treat as hard, generalization and dexterity, is secondary in space; everything they treat as background, reliability and verification, is the entire engineering problem. The result is a field that runs decades behind terrestrial robotics in raw capability and decades ahead of it in discipline, and both facts trace back to the same constraint.

This module covers the three places that discipline is visible. First the NASA/JPL rover program, which is the longest-running deployment of field autonomy anywhere, and the only one where autonomy is forced by physics rather than chosen for capability. Second in-situ resource utilizationin-situ resource utilization (ISRU)Producing mission consumables at the destination from local materials instead of launching them from Earth: oxygen from the Martian atmosphere, water or metals from lunar regolith, propellant from both. NASA frames it as astronauts living off the land, because every kilogram made on site is a kilogram that does not ride a launch vehicle, and for a Mars return the arithmetic is stark, since the propellant for the trip home would dominate an all-Earth-supplied mass budget. The first demonstration on another planet was MOXIE, which electrolyzed atmospheric carbon dioxide into oxygen on the Perseverance rover. The robotics side of ISRU is excavation, drilling, and material handling in vacuum and regolith, the same contact-rich manipulation problems as terrestrial robotics under harsher constraints., or ISRU, the program of turning local material into propellant and consumables, whose robotics half is excavation and drilling under the worst conditions manipulation has. Third on-orbit servicingon-orbit servicingUsing one spacecraft to work on another after launch: docking with an aging satellite to take over its propulsion, refueling it, replacing failed components, assembling large structures, or removing debris. The robotic content is rendezvous and proximity operations flown to centimeter-per-second tolerances, plus capture and manipulation of hardware that was usually never designed to be serviced. The capability has been demonstrated in stages since Japan’s ETS-VII and DARPA’s Orbital Express, and became a commercial service when Northrop Grumman’s MEV-1 docked with Intelsat 901 in 2020 to extend its life; debris inspection and removal is the unserved variant, where the client is uncooperative and tumbling.: rendezvous, capture, refueling, and debris removal, the one branch of space robotics that has become a commercial industry in the last five years.

Autonomy paced by the speed of light

A Mars rover operates on a cadence set by physics. The operations team plans a sol's activity during the previous sol, uplinks the sequence in a single communication window, and receives the results after the rover has already executed them. Under that cadence, every second of onboard decision-making is worth trading against bandwidth, which is why NASA/JPL's autonomy program is best read not as a capability project but as an optimization against scarce communication.

The numbers from Perseverance, published by its own autonomy team, show how far that optimization has gone. In the rover's first Mars year, its AutoNav system evaluated 88 percent of the 17.7 kilometers the vehicle traveled, and the previous record for autonomous distance evaluation was 2.4 kilometers, accumulated by Opportunity over fourteen years Verma 2023↓Autonomous robotics is driving Perseverance rover's progress on MarsVandi Verma, Mark W. Maimone, Daniel M. Gaines et al., Science Robotics 8(80), adi3099, 2023. AutoNav has driven 699.9 meters without a human reviewing the imagery, and 347.7 meters in a single sol, both planetary rover records Verma 2023↓Autonomous robotics is driving Perseverance rover's progress on MarsVandi Verma, Mark W. Maimone, Daniel M. Gaines et al., Science Robotics 8(80), adi3099, 2023. None of this is learned control in the sense of this wiki's policy modules; it is stereo vision, terrain assessment, and trajectory evaluation, engineered and verified a decade before flight. The one place a classifier sits inside the loop is targeting: the AEGIS system analyzes wide-angle imagery onboard and autonomously selects targets for the SuperCam instrument, so scientifically interesting targets get observed during or immediately after long drives without waiting for a ground communication cycle Verma 2023↓Autonomous robotics is driving Perseverance rover's progress on MarsVandi Verma, Mark W. Maimone, Daniel M. Gaines et al., Science Robotics 8(80), adi3099, 2023. AEGIS was first deployed on Curiosity in 2016, where onboard target selection for the ChemCam laser spectrometer measurably increased science data return Francis 2017↓AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team useR. Francis, T. Estlin, G. Doran et al., Science Robotics 2(7), eaan4582, 2017. That is the honest picture of learning in deep space: a bounded decision, inside a classical system, adopted because it spends bandwidth better than a human in the loop.

The same culture of designing for the case that cannot be retrieved explains the Ingenuity helicopter. On April 19, 2021, Ingenuity became the first aircraft to make a powered, controlled flight on another planet NASA Jet Propulsion Laboratory 2021↓NASA’s Ingenuity Mars Helicopter Succeeds in Historic First FlightNASA Jet Propulsion Laboratory, NASA/JPL news, 19 April 2021, in an atmosphere with about one percent of Earth's surface pressure, which is the reason nothing about the flight could be tested as-flown before launch. It was qualified as a technology demonstration: five flights over thirty days. It flew 72 times over almost three years before NASA concluded the mission in January 2024 NASA 2024↓After Three Years on Mars, NASA’s Ingenuity Helicopter Mission EndsNASA, NASA news release, 25 January 2024. A fourteen-fold margin over the design life is what verification-heavy engineering buys when the design point itself is conservative, and it is the same relationship between claim and evidence this wiki keeps asking of manipulation results, running in the opposite direction.

Living off the land: ISRU

ISRU is the engineering name for living off the land: producing mission consumables at the destination instead of launching them. The propellant needed to lift a crew off Mars and send it home would dominate an all-Earth-supplied mass budget, which is why the first demonstration of the technology on another planet was an oxygen generator. MOXIE, riding on Perseverance, electrolyzed carbon dioxide from the Martian atmosphere in a solid-oxide cell; across sixteen runs it produced 122 grams of oxygen, peaking at 12 grams per hour at 98 percent purity or better, about twice NASA's original production goals, before concluding operations in September 2023 NASA Jet Propulsion Laboratory 2023↓NASA’s Oxygen-Generating Experiment MOXIE Completes Mars MissionNASA Jet Propulsion Laboratory, NASA/JPL news, 6 September 2023.

MOXIE is an instrument, and instruments do not shovel. Scaled ISRU is a robotics problem: drilling, excavating, and moving granular material in vacuum, at cryogenic temperatures, with abrasive dust, over maintenance-free years. The first American attempt to exercise that hardware on the Moon is a case study in how partial the data comes back. NASA's PRIME-1 suite flew to the lunar south pole on Intuitive Machines' IM-2 lander in March 2025. The lander, Athena, came to rest on its side in a crater about 400 meters from its target site, and the mission ran roughly ten hours against the ten days planned NASA 2025↓NASA’s Lunar Drill Technology Passes Tests on the MoonNASA, NASA article, 29 April 2025. Within that window the hardware itself worked: TRIDENT, a one-metre rotary-percussive drill built by Honeybee Robotics, executed its commanded motions into the regolith, and the MSOLO mass spectrometer ran its scans, though the gases it detected were all of human origin, propellant and traces of Earth water, rather than lunar volatiles NASA 2025↓NASA’s Lunar Drill Technology Passes Tests on the MoonNASA, NASA article, 29 April 2025. NASA's own summary is precise about what was and was not demonstrated, and it is worth reading as the ISRU program's real progress bar: the mechanisms survived the environment and obeyed commands, while the science return was lost to a landing accident upstream of them NASA 2025↓NASA’s Lunar Drill Technology Passes Tests on the MoonNASA, NASA article, 29 April 2025. Compare that with the evaluation-crisis module's argument about bench results: in space, the environment gets a vote on every trial, and the first trial of a new capability is often also the last one available for years.

Orbital robotics: servicing and removal

The second tradition of space robotics operates closer to home, where the machines can at least be reached. On-orbit servicing is a rendezvous and manipulation problem: arrive at a target moving kilometers per second, match its motion to centimeters per second, then either dock with hardware that was never designed to be docked with or grasp hardware that was never designed to be grasped.

The table reads as one lineage. Japan's ETS-VII demonstrated autonomous rendezvous and docking between two spacecraft in 1998, along with the flight experiments for a two-metre, six-degree-of-freedom manipulator mounted on one of them Kawano 2001↓Result of Autonomous Rendezvous Docking Experiment of Engineering Test Satellite-VIIIsao Kawano, Masaaki Mokuno, Toru Kasai et al., Journal of Spacecraft and Rockets 38(1), 105-111, 2001Yoshida 2001↓ETS-VII Flight Experiments For Space Robot Dynamics and ControlKazuya Yoshida, Experimental Robotics VII, Lecture Notes in Control and Information Sciences, 209-218, 2001. DARPA's Orbital Express put the full servicing sequence together in 2007: over a four-month flight demonstration, the ASTRO servicer autonomously captured its NextSat client, transferred propellant to it, and robotically exchanged a battery with it Friend 2008↓Orbital Express program summary and mission overviewRobert B. Friend, Proc. SPIE 6958, Sensors and Systems for Space Applications II, 2008. The International Space Station runs the operational version of the same machinery every week: Canadarm2, a 17-metre arm that assembled the station and still performs the cosmic catches, grappling arriving cargo vehicles and berthing them Canadian Space Agency 2024↓About Canadarm2Canadian Space Agency, Canadian Space Agency, 2024, while Dextre, a two-armed robot riding the same arm, replaces exterior equipment including the station's 100-kilogram batteries, all of it teleoperationteleoperationDirect remote control of a robot by a human operator, used in robot learning mostly to record expert demonstrations. The ALOHA rig made the setup cheap and precise: the operator back-drives a leader arm that is a joint-for-joint twin of the follower the policy will run on, so the recorded motion is already in the robot's own action space. Mobile ALOHA extended the same leader-follower scheme to a mobile base for whole-body tasks. planned and flown by a ground crew Canadian Space Agency 2024↓About DextreCanadian Space Agency, Canadian Space Agency, 2024. What the station proves, at twenty-plus years of continuous operation, is that durable space manipulation is possible when the machine is maintained, reachable, and supervised, which is exactly what a satellite in a graveyard orbit is not.

The commercial era started in 2020, when Northrop Grumman's MEV-1 performed the first docking between two commercial spacecraft in orbit, capturing the Intelsat 901 broadcast satellite in the graveyard orbit above geosynchronous altitude and returning it to service Northrop Grumman 2025↓Northrop Grumman Achieves First-Ever Undocking Between Two Commercial Spacecraft in Geosynchronous OrbitNorthrop Grumman, Northrop Grumman news, 9 April 2025. Five years later, with the client's life extension complete, MEV-1 undocked, the first undocking between commercial spacecraft in that orbit, and moved on to its next servicing mission Northrop Grumman 2025↓Northrop Grumman Achieves First-Ever Undocking Between Two Commercial Spacecraft in Geosynchronous OrbitNorthrop Grumman, Northrop Grumman news, 9 April 2025. Life extension by docking is deliberately conservative robotics: the client satellite is cooperative, three-axis stabilized, and expects the visit. The frontier case is the client that cooperates with nobody. Astroscale's ADRAS-J, flying for JAXA's debris-removal program, rendezvoused with a discarded rocket upper stage during 2024, inspected it with three fly-around surveys, and on its final approach closed to about 15 metres, the closest any commercial spacecraft has come to a piece of debris through rendezvous and proximity operations Astroscale 2024↓Astroscale’s ADRAS-J Achieves Historic 15-Meter Approach to Space DebrisAstroscale Japan, Astroscale announcement, 11 December 2024. That approach ended in an autonomous abort, the onboard collision-avoidance system reacting to an unexpected attitude anomaly on the dead stage, which is the demonstration working as designed: the hard part of debris removal is the trust required to close the last 15 metres at all Astroscale 2024↓Astroscale’s ADRAS-J Achieves Historic 15-Meter Approach to Space DebrisAstroscale Japan, Astroscale announcement, 11 December 2024.

The honest counterweight to this progress is NASA's own OSAM-1, a mission meant to demonstrate refueling a satellite that had not been prepared for service. In March 2024, after an independent review, NASA discontinued the project, citing continued technical, cost, and schedule challenges and a broader community shift away from refueling unprepared spacecraft, with no committed partner remaining NASA 2024↓Update on Status of NASA's OSAM-1 ProjectNASA, NASA, 1 March 2024. A flagship servicing demonstration dying of cost growth, in the same decade the commercial servicers began operating, is not a contradiction. It is the economics of the reliability bar: servicing pays when the client is expensive and the servicer is reusable, and every unserviced design that launches resets the market backward.

What robot learning takes from it

Three transfers, none of them about hardware.

First, the latency argument from the drones module, completed at solar-system scale. The perception-latency study cited there measures how milliseconds of pipeline delay cap a drone's safe speed; light-time delay is the same mechanism with the constant set by physics instead of engineering, and its product is not a slower robot but an unsupervised one Verma 2023↓Autonomous robotics is driving Perseverance rover's progress on MarsVandi Verma, Mark W. Maimone, Daniel M. Gaines et al., Science Robotics 8(80), adi3099, 2023. Any robot whose control loop cannot close through a human, a Mars rover during a drive, a servicing spacecraft in final approach, must either freeze its policy before deployment or carry the whole decision onboard. That is the deep reason space autonomy is classical: a verified stereo pipeline can be argued about line by line before launch, and a policy trained on ten thousand demonstrations cannot, yet.

Second, the reliability bar, which this domain shares with the surgical robotics module and pushes further. The surgical field deploys at Level 0 autonomy because each increment must be individually cleared; space robotics deploys bounded autonomy only after test-like-you-fly verification, and its failures are instructive in both directions: Ingenuity exceeded its qualification fourteen-fold NASA 2024↓After Three Years on Mars, NASA’s Ingenuity Helicopter Mission EndsNASA, NASA news release, 25 January 2024, while OSAM-1, a billion-dollar-class program with the field's best review culture, was cancelled before it flew NASA 2024↓Update on Status of NASA's OSAM-1 ProjectNASA, NASA, 1 March 2024. The reliability-gap module's compounding arithmetic applies to a rover that must survive thousands of sols, but the deeper lesson for learned policies is evidentiary: flight heritage is the only currency this field accepts, and heritage is accumulated one bounded, supervised increment at a time, of which AEGIS's classifier inside a verified rover is the flight-proven example Francis 2017↓AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team useR. Francis, T. Estlin, G. Doran et al., Science Robotics 2(7), eaan4582, 2017.

Third, a calibration for claims. When a manipulation system is described as operating in unstructured environments, the honest comparator includes systems that operate where the environment was never mapped, the supervision never arrives in time, and the warranty is one-way. The Mars rovers and the commercial servicers hold that bar with classical, verified autonomy and a decade of qualification between design and flight. A learned policy will eventually hold it too, and the path there runs through exactly the niches AEGIS occupies: bounded decisions, inside verified systems, where the value of deciding onboard is measurable in watts and bits Verma 2023↓Autonomous robotics is driving Perseverance rover's progress on MarsVandi Verma, Mark W. Maimone, Daniel M. Gaines et al., Science Robotics 8(80), adi3099, 2023. Between here and a learned system on Mars, most of the remaining distance is the part of engineering that proves, before launch, what the system does in the case it was never shown.