CIMON

NASA. Public domain (NASA / US government work).

Overview

CIMON, the Crew Interactive MObile companioN, is a free-flying artificial intelligence crew assistant operated in the European Columbus module of the International Space Station. The project was commissioned by the German Space Agency at DLR to explore human-robot interaction on long-duration missions, and the hardware was developed by Airbus [1]. Working in confined and extreme environments over a long mission is a risk to crew performance and wellbeing, and CIMON is the technology demonstrator aimed at that risk [2]. Its purpose is to reduce astronaut workload and stress, give hands-free access to procedures and data, and serve as a research subject for human-AI teaming [1]. Unlike the task-oriented free-flyers alongside it on the station, its core function is to interact, converse and demonstrate a form of emotional intelligence. In operation it acts as a database, a computer and a camera, showing the equipment required for a task on its screen, presenting directions for conducting experiments, searching for objects and taking inventory [2].

CIMON-2, Int-Ball2 and Astrobee are the intra-vehicular free-flyers currently resident on the station [2].

Specifications

Parameter	CIMON-1	CIMON-2
Shape and size	32 cm sphere	32 cm sphere
Mass	approx. 5 kg	approx. 5 kg
Mobility	12 propelling fans, about 120 mN maximum in x	same
Navigation	stereo camera VSLAM with IMU fusion	same
Imaging	1280 x 1024 at 25 fps	same
Processor	Pokini F WIFI (AMD A4-6700T), two units	same
Software	Ubuntu with ROS	same
Power	Li-ion, crew replaceable, or wired input	same
Endurance on batteries	2 h	3 h
Data communication	wired connection, WLAN, Bluetooth	same
Payload interface	2 x USB, Bluetooth	same
Ground operations	ground station remote connection plus AI server	same

Source: [1].

Mission profile

Parameter	Value	Source
Host	International Space Station, Columbus module	[1]
CIMON-1 on orbit	2 July 2018 to 27 August 2019	[6]
CIMON-2 launch	5 December 2019, 18:29 CET, SpaceX CRS-19 from Cape Canaveral	[5], [6]
CIMON-2 planned duration	up to three years aboard the station	[5], [6]
Users and science subjects since 2018	more than five nations	[1]

CIMON-1 was developed by about 50 people over two years from August 2016, and CIMON-2 by 20 people in under a year [6]. Astrobee’s 2015 charter paper names Smart SPHERES and the Personal Satellite Assistant as its direct heritage [21]. Astrobee, by comparison, traces its own project start to the Human Exploration Telerobotics 2 project in October 2014 and reached its first ISS launch in 2019, a span that included a full propulsion redesign away from cold-gas thrusters [12], [16]. Airbus in Friedrichshafen and Bremen built the hardware, and the DLR Space Administration funded and oversaw the project with German Federal Ministry for Economic Affairs and Energy money, IBM supplied the artificial intelligence, Ludwig Maximilian University Munich was the scientific partner and Biotesc at the University of Lucerne provided ground support [6].

Mobility

CIMON moves and rotates in all directions on internal battery-powered fans, which provide both attitude control and translation within the module [1]. It has no external moving parts and the body is largely produced by 3D printing [4]. Published fan counts conflict: the program summary text gives 14 internal fans [1] while its own comparison table gives 12; the review of intra-vehicular free-flyers records 14 on CIMON and 12 on CIMON-2 [2]; DLR states twelve [4]. Thrust is deliberately limited to between 0.03 and 0.12 N across the three axes for safety [2], with a maximum of about 120 mN in the x direction [1]. The spherical form is the choice made across intra-vehicular free-flyers to limit the consequences of a collision with crew or cargo [2].

Astrobee took a different route to the same safety goal with a cube body: each of its two propulsion modules feeds a plenum from one centrifugal impeller and doles thrust out through six flapper-controlled nozzles per module, a design chosen specifically to keep exhaust velocity low, around 11 m/s against the roughly 250 m/s of the SPHERES cold-gas thrusters it replaced, so that a worst-case collision along the longest straight run in the ISS interior stays within what an impact-absorbing foam shell and a hardware speed cutoff can contain [8], [9]. CIMON instead limits thrust magnitude directly at the fan level and forgoes a shell, trading Astrobee’s soft-body collision margin for a smaller, unpadded sphere.

Power and energy

CIMON runs on lithium-ion batteries that the crew replaces, and it can also take a wired power input [1]. Endurance is 2 hours on CIMON-1 and 3 hours on CIMON-2, an increase of about 30 percent in the battery-powered autonomy [2], [6]. There is no docking station: CIMON requires crew for setup, unlike Astrobee and Int-Ball2, which reach a dock by themselves [1]. The Astrobee team made batteries available to the CIMON project in orbit until CIMON received its own items, which is the argument the joint operations teams make for a common pool of batteries, chargers, navigation markers and onboard computing across free-flyer projects. No battery capacity in watt-hours is published. The shift from consumables to rechargeable batteries across intra-vehicular free-flyers dates from the AERCam, Personal Satellite Assistant and SPHERES projects, and autonomous recharging is listed as one of the capabilities essential for any free-flyer tasked with maintaining an orbiting station [2].

Thermal

No thermal control design is published. CIMON carries temperature sensors [1] and operates in the pressurized cabin. All of these free-flyers use the existing infrastructure of the ISS for power supply, heat rejection and communications rather than providing their own, and none of the published Astrobee, CIMON or Int-Ball2 design and operations material describes a dedicated radiator or active heat rejection system distinct from the cabin environment [2], [12].

Compute and avionics

Onboard computing is two units of Pokini F WIFI single-board computers built around an AMD A4-6700T, running Ubuntu with ROS, backed by an AI server on the ground [1]. The split is architectural rather than incidental: onboard software handles autonomous navigation and flight control by visual odometry and motion planning, while the language understanding runs on Earth. Both Astrobee and Int-Ball2 place their vision loop on a separate processor or core for the same reason, and time-critical modules such as propulsion take the highest priority across all of these designs [2].

Int-Ball2, the other station free-flyer to add a docking station after CIMON’s launch, runs a single Nvidia Jetson TX2 under Ubuntu with ROS Melodic and Gazebo, with guest software isolated in Docker containers uplinked and started from the ground rather than run over a live ground-to-robot link, and its own visual SLAM cannot run at the same time as direct access to the raw navigation camera feed [10]. Its first on-orbit docking demonstration, flown 17 October 2023, closed autonomously on the magnetic docking station with no crew action, a capability CIMON does not have [11]. That is a narrower but self-contained onboard stack; CIMON keeps its own vision and motion loop onboard but sends every spoken exchange to Earth, so the two designs split the same problem, onboard perception versus onboard cognition, in opposite directions.

Autonomy

The artificial intelligence is IBM’s Watson platform running on Earth servers. Spoken commands are transmitted to the ground, processed by Watson, and the responses sent back to the robot [1]. CIMON is not self-learning; its knowledge base must be pre-loaded and trained by human operators. The cloud-dependent architecture was chosen for demonstration purposes and contrasts with systems designed for onboard autonomy. The consequence is operational: a continuous link is a core constraint on when CIMON can be used.

That constraint is the reason CIMON appears in the case for offline assistants. It is cited as an intelligent personal assistant limited by predefined, inflexible responses drawn from specific structured data sources, rule-based and requiring constant online connectivity [3]. One-way Mars communication delay reaches 24 minutes, which removes the ground-processing option entirely for a deep-space equivalent.

The other station free-flyers keep their autonomy loop onboard rather than splitting it across the link. Int-Ball2 flew a Proximal Policy Optimization-based docking controller derived from simulation training, reaching 97.2 percent stable docking against a 65.6 percent baseline once propeller drag-torque and spin-polarity effects were modeled in training, all without a continuous ground channel in the loop [13]. Astrobee similarly flew a reinforcement-learned attitude and translation controller trained in simulation and tested on a ground air-bearing table before being commanded to fly it on orbit [14]. Both demonstrate that the onboard compute on an ISS free-flyer is sufficient to host a learned control policy; CIMON’s architecture puts that same class of compute budget into hosting a spoken-language interface instead, and pays for it with the continuous link.

Navigation is vision-based. A dual 3D camera collects depth information and the relations between features to build a map with SLAM algorithms [2], fused with the IMU [1]. Directional microphones localize the sound source so the robot turns toward the speaker, and the front camera then establishes and maintains eye contact [2]. Ultrasound sensors handle collision avoidance [2], [4]. CIMON uses its own visual fiducials as an absolute localization reference, independent of the marker sets used by Astrobee and Int-Ball2 [1]. The nearest published localization record on the station is Astrobee’s: an extended Kalman filter fusing a 62.5 Hz IMU with sparse map landmarks, AR tags and handrail detections at roughly 2 to 5 Hz, which reached position errors of about 5 cm on a slow circle and 16.5 cm on a fast combined motion in ground tests, and used one map for three months in a busy lab [23]. No equivalent accuracy figure is published for CIMON’s stereo VSLAM.

Communications

CIMON has a wired connection, station WLAN and Bluetooth [1], the widest link set of the three current intra-vehicular free-flyers [2]. Because the language model runs on the ground, the WLAN path carries the interaction loop rather than telemetry alone, and the requirement for a continuous link is treated by the operations teams as the defining CIMON constraint [1]. The quality of that link inside the station is not uniform: Astrobee flight surveys in 2022 found that the free-space inverse-square path-loss model does not hold inside the US Lab, Node 2 and the JEM, with lateral position relative to the module centerline and hatch crossings dominating received signal strength [22]. The survey used an uncalibrated radio and antenna, so only relative differences are reliable [22]. A conversation that depends on the cloud therefore depends on where in the module CIMON is.

Payload and instruments

The sensor set is a crew camera, an infrared camera, an IMU, ultrasonic sensors and temperature sensors [1]. The optical suite as described by the program comprises stereo cameras for 3D vision and navigation, an infrared camera, two side cameras supporting augmented reality tasks and a high-resolution camera for documenting experiments. The manufacturer’s description of CIMON-2 lists a stereo camera for orientation, a high-resolution camera for facial recognition and two lateral cameras for photo and video documentation [6]; DLR describes two cameras for facial recognition plus five others for orientation and video documentation [5]. Imaging is 1280 x 1024 at 25 fps [1], against Int-Ball2’s 1920 x 1080 at 30 fps and 4056 x 3040 at 15 fps. The eyes on the front face are two of the cameras and are used for facial recognition [5].

The Astrobee program reached a comparable milestone on its own imaging suite: its perching arm and docking station were each commissioned and flying operational sessions within the first two years, versus CIMON’s absence of a dock across both flight units, which the joint review attributes to CIMON’s spherical shell leaving no surface area for a perching or docking mechanism without redesign [12].

The audio system uses multiple microphones to perform directional sound source localization, which is what lets the robot turn and face the astronaut who is speaking [1]. CIMON-2 carries eight directional microphones plus one additional directional microphone dedicated to speech recognition [6], [5], and a loudspeaker for speech and music [4], along with cameras and software for facial recognition, orientation and video documentation. Int-Ball2 carries a comparable stereo navigation camera plus a 4K imaging camera reaching 4056 x 3040 pixels at 15 fps [11], and its docking station specification gives the kind of external envelope figure, 169 x 169 x 92 mm, that CIMON has no equivalent to since it has no dock [11].

Two USB ports and Bluetooth are available as a payload interface [1], and the laser pointer and USB ports are hardware that has not yet been activated in flight.

Modes of operation

The crew interface is an LCD monitor showing an animated face that conveys operational status, a speaker, status LEDs, a laser pointer, on-off and microphone-mute buttons, and IMU-based physical contact detection [1]. The animated face is the status display, and the design intent is a more natural interaction than a text readout. CIMON acts as a database, a computer and a camera in the crew’s workflow rather than as a manipulator: of the free-flyers on the ISS only Astrobee carries an arm, and that is a perching arm [2]. Astrobee’s arm is used to grip station handrails for a stable extended vantage point rather than to manipulate objects, so none of the three current intra-vehicular free-flyers performs cargo or equipment handling [15].

Crew privacy is handled in hardware: CIMON provides a mechanical offline button [1]. Int-Ball2 indicates data acquisition with LEDs and Astrobee uses a set of visual cues, and in all three cases the crew are informed before any activity so they know video may be downlinked.

CIMON-2 can analyze emotion in language and show empathy when interacting with the crew [6], a capability none of the task-oriented survey and docking free-flyers on the station attempt [12], [11]. Teleoperation from the ground remains the fallback for fault detection and recovery on every intra-vehicular free-flyer regardless of the autonomy level reached [2]. Stewart et al. demonstrated the opposite direction is also possible: a reinforcement-learned controller trained entirely on the ground was uplinked and flown on Astrobee without further tuning, showing that at least some autonomy functions can migrate onboard after terrestrial training [14].

Ground operations

CIMON is operated over a ground station remote connection working alongside the AI server [1]. Ground verification used software simulations, ground model testing, an Elegant Bread-Board and hardware-in-the-loop simulation. Some performance variances are found only on orbit, which is why every one of these programs pairs simulation with hardware test articles. Building a prototype before the official design review accelerated development, which the joint operations teams list as a lesson common to all three ISS free-flyer programs. Early alignment with the Human Safety board supported a smooth development phase and control of all hazards.

The station’s free-flyer lineage runs through an earlier telerobotic demonstrator that none of the three current vehicles directly descends from in hardware. Smart SPHERES, a SPHERES satellite carrying a smartphone as its computer and camera, flew an inventory and wall survey on the ISS in December 2012 under ground teleoperation, and its documented failures, cold-gas propellant exhaustion partway through two of five survey runs, a ceiling impact that broke the position estimate, and blurred imagery from an asymmetric payload upsetting the spherical body assumption, are cited by both the Astrobee and CIMON programs as the requirement basis for a purpose-built free-flyer rather than an augmented station-keeping satellite [7].

Anomaly resolution across these programs runs tiered: reproduce the fault on identical ground hardware, attempt remote recovery by power cycle or software patch, then crew replacement of hardware on orbit, then down-mass for ground repair [1]. CIMON took the replacement route, substituting an upgraded robot rather than repairing the flown one, as the Int-Ball project also did; Astrobee’s Honey was instead returned to the ground for repair, flew again, then logged its single hardware problem in a corrupted SD card during its six years of operations [12]. Ground support for CIMON operations is provided by Biotesc at the University of Lucerne [6].

Earlier free-flyers and what carried over

CIMON is the third generation of free-flying robots built for crewed spacecraft, and each earlier family fixed a design choice it inherits. The SPHERES thesis characterized a cold-gas, battery-powered free-flying satellite in laboratory and KC-135 testing: about 10 minutes of continuous station keeping per propellant tank and at least 90 minutes of battery power, with translation authority not demonstrated against KC-135 turbulence [17]. Those limits are the reason the later station vehicles moved to fans and rechargeable batteries. The external inspection line began with AERCam Sprint, a 35 lb, 14 in vehicle flown on Shuttle mission STS-87, followed by Mini AERCam, a 7.5 in, roughly 10 lb sphere with twelve cold-gas thrusters, built as a ground technology demonstration and not flown [18]. Its autonomy layer used a rule-based command sequencer and a loss-of-communications response that sequences safe and all-stop commands, tested on an air-bearing table and in a hardware-in-the-loop simulation [19]. CIMON takes the sphere form and the supervised-autonomy posture from this line but moves the free-flyer inside the pressurized module and makes dialog its function. Robonaut, a 43-degree-of-freedom humanoid, took the opposite premise: the robot adapts to existing crew worksites and interfaces instead of the worksite being redesigned for the robot [20].

Close-proximity inspection autonomy for free-flyers has been studied further in simulation only: a 2025 architecture for a 12-thruster inspector near Lunar Gateway geometry found that linger mode costs about five times the translational impulse of flyby over a 41.32 m path, with perfect localization assumed and no hardware test [24]. It informs the CIMON successor question of onboard autonomy but has not flown.

Technologies developed

CIMON established that a verbal and visual interface is a usable additional tool for daily work in a crewed exploration environment, and most users gave positive feedback [1]. It was built to reduce crew workload and emotional stress, and it is the free-flyer whose declared product is the interaction itself rather than a camera view or an algorithm testbed [2]. Users and science subjects from more than five nations have interacted with the system since 2018, performing different and comparable tasks, which is the basis on which precise system requirements for a successor can be identified [1]. Astrobee’s Guest Science program is the scale comparison: over 2 terabytes of data delivered to 17 unique investigators across 22 projects through May 2025, run mostly without crew time on console [12].

The program’s own stated limitation is the constraint imposed by its setup and its dependencies on external services, which traces to the 2016 mission goal of fielding an early-available technology demonstrator [1]. The stated next step is a hardware update increasing autonomy and availability, including a docking station and permanent deployment, together with on-premises AI services based on large language models and agentic AI. Offline assistants built from knowledge graphs, retrieval-augmented generation and language models are being designed for exactly that case, to run without a link to Earth [3].

Software interoperability turned out to be the transferable result. In 2025 CIMON and Int-Ball2 exchanged data in real time in the ICHIBAN activity, which was possible because both run ROS as their middleware [1]. Modular software architecture on ROS was adopted independently by Astrobee, CIMON and Int-Ball2, and the joint teams identify shared middleware as the route to sharing sensory and computational resources between free-flyers in a resource-constrained environment.

Set against Astrobee’s operations record, CIMON’s crew-setup requirement is the visible cost of forgoing a dock: across six years on orbit Astrobee’s three units and docking station logged well over 165 real-time operations and more than 90 Guest Science sessions with the great majority of console time needing no crew present at all [12], a session cadence a manually-deployed, ground-dependent assistant cannot match. NASA’s own program materials describe the same docking and autonomous-recharge capability as the baseline expectation for a station free-flyer [15], which is the capability CIMON’s stated next-generation hardware update is intended to add.

What is not established

No performance measurement of either CIMON unit is published in the held sources. The manufacturer’s statement that CIMON-2 autonomy rose by around 30 percent over CIMON-1 carries no definition, baseline or unit [6]. Navigation accuracy, translation and rotation rates, station-keeping error, collision-avoidance standoff distance and speech recognition rate in ISS acoustic noise are not reported [6]. The 32 cm diameter and the fan count are design statements from the project page [4], and fan counts differ between sources [1], [2], [4]. Voice-commanded navigation to Biolab is reported as a single successful flight, so no reliability figure follows [6]. The CIMON-2 launch release gives component counts without specifications and its only operational data point, the 90-minute session, belongs to CIMON-1 [5]. The empathy and stress-reduction goals are stated aims; no study, sample or outcome is reported [6]. The offline assistant architecture that would replace the cloud link is a proposal with no implementation, experiment or quantitative result, and has not been tested on station hardware or with crew [3]. No battery capacity, thermal design or mass breakdown is published.

References