Astrobee

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

Overview

Astrobee is a fleet of three free-flying robots operating inside the pressurized volume of the International Space Station: Honey, Bumble and Queen [4], [13]. A free-flyer in a crewed cabin has to move in six degrees of freedom without a floor, localize against an interior that crew rearrange, avoid harming people and windows, and run for days without a person servicing it. Each requirement is harder in a cabin than outside it: exhaust and reaction mass are a contamination and safety problem, radio-frequency beacons are a certification problem, and every collision is a collision with something a crew member cares about [1]. Astrobee replaced SPHERES as the station’s intravehicular free-flyer facility. The design is modular and upgradeable, with propulsion by electric fan rather than stored propellant, built-in cameras, and a perching arm for grasping handrails.

Lineage and the design space

Intravehicular free-flyers have been built in several propulsion and navigation families, and the ones that flew in the cabin converged on the last.

The Personal Satellite Assistant set out the mission role of a free-flying intravehicular monitor and an autonomy framework built on planners from Deep Space One’s Remote Agent; only its first two prototype models were operational [12]. It never flew.

SPHERES flew cold-gas thrusters and localized against ultrasonic beacons. Its thesis-era hardware gave roughly 10 minutes of continuous station keeping per tank and 2 cm laboratory metrology precision [14]. Measured thrust at the flight maximum regulator setting was 0.1783 N against 0.2480 N predicted, and opening six thrusters at once cut each thruster’s thrust by more than 30 percent because the regulator could not hold pressure [15]. On the station, SPHERES flew formation-acquisition searches in March 2007, where multipath reflection of the ultrasound off module walls delayed relative estimator convergence to about 45 to 55 s [16]. Work on reconfigurable control found that pooling the ultrasound receivers of two attached satellites made the position estimate worse in hardware, not better [17]. SPHERES was useful as control research but kept the consumable and beacon constraints.

AERCam was the external-inspection branch: the Mini AERCam, a 7.5 in cold-gas sphere with GPS and MEMS-gyro relative navigation and twelve thrusters [18], with a CLIPS expert-system command sequencer and a trajectory-safety manager [19]. It targeted free flight outside the station and does not share Astrobee’s cabin constraints.

Smart SPHERES added a smartphone to a SPHERES satellite as a ground-commanded camera payload. In the 12 December 2012 survey of the Kibo module every ground command executed across two loss-of-signal periods, but two runs ended when the cold-gas propellant ran out and one ended when the satellite hit the ceiling hard enough to break its position estimate [9]. Astrobee’s charter absorbed those failure modes: propellant that is cabin air and a battery that recharges at a dock, and collision tolerance built into the mechanical design rather than into operational caution [11], [1].

Astrobee, JAXA’s Int-Ball2 and DLR/Airbus’s CIMON are the free-flyers operating in the station now. A joint review by members of all three teams finds that what separates platforms that fly often from those needing crew setup at every session is operational autonomy through a docking station: Astrobee and Int-Ball2 dock and recharge unattended, while SPHERES, the first Int-Ball and CIMON do not [10]. Int-Ball2 completed its first on-orbit demonstration of autonomous intravehicular flight on 17 October 2023, with vSLAM navigation and a magnetic docking station [25]. Communication redundancy also mattered in practice, including two Astrobee dock Ethernet failures recovered over Wi-Fi, and localization in a changing station interior is a shared failure mode across all three programs [10].

Specifications

Parameter	Value	Source
Envelope	12.5 in cube (about 30 cm)	[2], [1]
Mass	about 10 kg	[3]
Design life	5 years	[2]
Maximum velocity	0.5 m/s	[3]
Localization accuracy	1 cm	[3]
Payload bays	4, two on top and two on the bottom; up to 3 usable at once by guest science	[4]
Payload bay port	14.4 VDC nominal battery power and USB 2.0	[4]
Perching arm degrees of freedom	3, two joints plus gripper	[3]

Mission profile

Parameter	Value
Dock commissioned	11 February 2019, 164 sessions to 2025
Bumble commissioned	30 April 2019, 135 sessions to 2025
Honey commissioned	30 October 2019, 59 sessions to 2025
Perching arms commissioned	4 February 2021, 23 sessions to 2025
Queen commissioned	20 September 2021, 21 sessions to 2025
On-orbit real-time operations through 2025	more than 165, of which more than 90 were guest science test sessions
Console support	more than 1200 hours, about 220 of them involving crew
Operating volume	Japanese Experiment Module, Node 2, US Laboratory

Source: [4], and [3] for the operating volume.

Queen and Bumble operated independently in separate modules on 7 April 2022 [13]. Hardware problems over the period were few: one on Honey (a corrupted SD card), one on Queen and two on the dock, with Honey and Queen each returned to Earth once [4]. In summer 2025 NASA selected the commercial company Arkisys to continue operating the platform [4], [13].

Propulsion

Propulsion is electric and uses cabin air rather than stored propellant, which removes both resupply and any exhaust product that would contaminate the atmosphere. Because reaction mass is free in a pressurized cabin, the design minimizes exhaust velocity to maximize thrust power efficiency: roughly 0.1 psi plenum pressure and 11 m/s exhaust, against SPHERES cold gas at 25 psi and 250 m/s [1]. Two propulsion modules, right and left, each contain one centrifugal impeller feeding a single plenum that discharges through six variable-area nozzles, twelve in total [2].

The fan and nozzle sizing came from a trade over fan RPM and nozzle open area. For the higher-performing fans there is a significant benefit from increasing the area up to about 5 in2 (32 cm2) and little beyond it, so the largest nozzles targeted that value, an order of magnitude larger than in the initial concept drawings, and the final open area after occlusion by the flapper mechanism was 3.7 in2 (24 cm2) [4]. Requirements driving the trade were 0.6 N maximum thrust on at least one axis and a 65 dBA acoustic limit at maximum thrust, set by crew habitability [4].

Parameter	Value
Impellers	2, one per module
Nozzles	12, 6 per module
Maximum thrust per nozzle	0.3 N
Maximum total thrust	3.6 N
Nozzle exit velocity	11 m/s
Nozzle discharge coefficient	0.9
Flapper travel	64 deg
Flapper backlash	1.8 deg
Nozzle servo	MKS DS95

Source: [2].

Each nozzle exit area is set by a pair of gear-synchronized flappers driven by a hobby-grade radio-control servo. Modulating twelve exit areas against two constant-speed impellers gives holonomic control: force in any direction plus torque about any axis. Development outcomes recorded for reuse include abandoning 3D printed flight nozzles for machined aluminum after vendor part quality problems [2]. The aft impeller intake produces a net disturbance force of less than 5 percent of propulsion thrust [4]. Total thrust of 3.6 N against a 10 kg vehicle gives a maximum acceleration near 0.36 m/s2, which is orders of magnitude above the milli-g disturbance levels the station structure sees [7]. Those levels depend on frequency and crew activity: a sensor filtered to 5 Hz read 0.06 mg during crew sleep against 0.20 mg awake, and docking transients reach roughly 10 mg [30].

Localization and navigation

Localization is six-degree-of-freedom and vision-based, matching wall texture features against a prior sparse feature map, with no beacons or added infrastructure in the module [1]. Visual odometry allows continued navigation through regions where mapped features are absent [3]. Position accuracy is about 1 cm in the design documents [3]; the flight evaluation below measures larger errors.

Six cameras are carried [4]: NavCam for mapping and localization, SciCam, a 13 MP forward-facing RGB imager used to stream video of crew activities [1], and HazCam, SpeedCam, PerchCam and DockCam for obstacle detection, velocity estimation, arm operations and docking respectively [1], [3]. HazCam and PerchCam are time-of-flight depth sensors with a 224 by 171 pixel format [4].

The localizer has passed through three generations. The original estimator was an extended Kalman filter fusing a 62.5 Hz IMU with three vision sources at roughly 2 to 5 Hz: sparse-map landmarks, AR tags for docking and depth-image handrail detections for perching [21]. Ground tests against overhead-camera truth gave position RMSE from about 5 cm on a slow circle to 16.5 cm on a fast combined motion, and one map built in a busy lab remained usable for three months [21].

AstroLoc replaced it with a factor-graph localizer that folds visual odometry in as combined factors and marginalizes over a sliding window to hold runtime down, cutting average localization time from 1.13 s to 0.21 s against a baseline factor-graph implementation [8]. On 12 real ISS activities scored against mapping-derived ground truth, localization-mode position RMSE was 0.0491 m for AstroLoc against 0.0948 m for the previous localizer, and in visual-odometry-only mode AstroLoc had no lost events against two for the previous localizer and for the baseline [4]. Fewer map-building activities were then needed on orbit, about 40 percent fewer activities and 30 percent fewer images per activity [4].

AstroLoc2 decouples odometry from absolute localization into a loosely coupled sequential pipeline so that noise in the ISS map no longer degrades velocity and IMU bias estimates. On a public dataset of 10 ISS activities it reduced absolute trajectory error RMSE by 16 percent against AstroLoc, and it runs 41 and 87 percent faster than DSO and ORB-SLAM3 while being less accurate than both [23]. Whether AstroLoc2 has been uploaded to the station is not stated in the sources read.

Perching requires its own perception. A handrail detector using a single time-of-flight sensor and a three-stage geometric filter (gap distance, cross point, side width) rejects cables and clutter before line fitting and 6-DOF pose estimation [22].

Obstacle mapping is implemented with an octree occupancy map fed by HazCam depth, but it is not used during ISS operations, because of its computational cost, the slow speed of Astrobee relative to astronauts, and operational checks that clear paths before motion [4].

Compute

Flight computing is split across three processor boards by real-time requirement [1], [20]:

Board	Role
Low-Level Processor	hard real-time motion control
Mid-Level Processor	core flight software, navigation, obstacle detection
High-Level Processor	payload interface and guest science code

Guest science code runs on the high-level processor, isolated from the control loop, so a visiting researcher cannot destabilize the vehicle [1]. A single core of the flight computer is about ten times slower than a laptop-class Intel i9 core, which is the constraint behind the runtime emphasis in the localizer work [4].

Perching arm

A stowable arm mounted in a payload bay grips a handrail so the robot can shut down propulsion and hold station at low power [3]. While perched the arm joints double as a pan and tilt unit for the forward-facing cameras, the SciCam among them [3]. The flight arm has three actuated degrees of freedom, two joints plus the gripper, and stows completely inside the top aft payload bay. The gripper is under-actuated and compliant, closed passively by torsional springs in the finger joints and opened by a single motor reeling two tendons, so it holds a handrail through a long perch without drawing power [4]. Bench validation of the arm gave a mean handrail detachment force of 6.43 N over five pull trials, inside a 4.45 to 22.2 N band chosen so that crew cannot dislodge a perched robot by bumping it but can pull it free deliberately [6]. The end-effector module is swappable on orbit so that a guest scientist can fit a different gripper or add arm degrees of freedom [6].

Docking

The dock is 85 by 38 by 28 cm and provides two berths [3]. Alignment is visual, against fiducial markers on the dock, and a mechanical funnel and spring-loaded pogo pins accommodate about 1 cm of residual error [4]. Retention is magnetic: the requirement was that the magnets release under an astronaut pull above 22 N and hold under incidental contact below 4.5 N, so the robot’s own thrust cannot overcome them [4]. Self-docking to recharge is what permits multi-day operation with no crew involvement.

Safety

Safety is argued by being light, slow and soft. The vehicle mass is about 10 kg and the maximum operating speed set by ISS geometry is about 2.1 m/s [4]. The over-speed cutoff is 50 cm/s, chosen so that bumper impact peak force stays below the 556 N maximum astronaut pushoff force that the ISS window scratch panes are verified to tolerate. Analysis backed by bumper testing puts the worst-case impact on a window below 2700 N, enough to risk the scratch pane but not a pressure pane [4]. That figure is analysis and ground test, not flight experience.

Communications

While flying, Astrobees link to the station local area network over the ISS Wi-Fi network, and when docked they communicate over wired Ethernet through the dock, which raises bandwidth and serves as a backup link [4]. Payloads have used both the 2.4 GHz and 5 GHz bands [5]. A free-flyer is itself a usable instrument for characterizing that network: the Wi-Fi survey of the US Laboratory, Node 2 and the JEM found that the free-space inverse-square path-loss model does not apply inside the modules, with a fitted exponent of 0.084 along the module axis, 35 dB of attenuation through a hatch, and more than 20 dB of loss when the payload antenna sat orthogonal to the access point antenna [5].

Ground testing

The Granite Lab at NASA Ames supports propulsion and motion-control tests on a 1.2 by 1.2 m granite table, with Astrobee on a three-degree-of-freedom air bearing (x, y, yaw) fed by liquid CO2 cartridges lasting about 15 minutes [4]. The air bearing mass cannot be modeled precisely because the cartridge mass changes as the CO2 is consumed. The lab carries a high-fidelity dock unit, handrail mockups and backdrop posters of the ISS interior, and has been mapped with the flight sparse-mapping software [4]. It is a one-g simulation with three planar degrees of freedom, so pitch and roll need re-mounting the robot on its side.

Operations

Astrobee runs experiment sessions, environmental surveys and guest payloads on the standardized bays [4]. Nearly 80 percent of console time across the first six years needed no crew at all, and 17 investigators on 22 projects had been supported by May 2025, with over 2.0 TB of guest science data transferred [4]. Each payload bay carries a blind-mate connector supplying 14.4 V nominal battery power and USB 2.0 data, engaged by two tool-free quarter-turn fasteners, and up to three payloads can be hosted at once [1]. Investigators develop against a simulator and then run on flight hardware.

Guest science has used the platform for several kinds of work. Astrobee survey images from the Granite Lab and from the JEM were used in a two-robot change-detection scheme, where one robot builds the 3D model and another later scans for inconsistencies; on the ISS, unmodified surveys recovered objects such as a laptop and lights from batches of about seven NavCam images [24]. Astrobee has also served as a testbed for learned control: a Lie-group-symmetry reinforcement learning method for trajectory tracking was tested on a particle, an Astrobee and a quadrotor [28].

Research beyond the flown system

Several lines build on the free-flyer class without having flown on a station free-flyer.

• Safe inspection autonomy: an architecture for a small free-flyer inspecting an orbital outpost, using operator-specified inspection points, keep-in corridors, keep-out zones and a model predictive contouring controller, evaluated in MuJoCo against approximate Lunar Gateway geometry. Four simultaneous thruster failures drive the vehicle into the station in simulation [26].
• Learning-based docking: a reinforcement learning policy for Int-Ball2 docking in a model of the Kibo module reached 97.2 percent stable docking over more than 900 attempts per ablation when propeller drag torque and spin polarity were modeled, against 65.6 percent for a thrust-only baseline [27]. This is simulation only.
• Control barrier functions: a modular safety layer for a 13-DoF free-flying servicer capturing a tumbling target, enforcing approach corridors, collision zones and actuator bounds [29]. This is simulation only.

What is not established

Pre-flight design numbers do not match as-flown hardware and should not be read as performance data: an early platform description gives mass as about 6 kg and a hardware thrust limit of 0.72 N with a 0.5 m/s overspeed cutoff, against the roughly 10 kg, 0.6 N and 50 cm/s cutoff the bumper testing later justified [1], [4]. The 1 cm localization accuracy in the design documents is not supported by the flight evaluation, whose best mean position error is 4.9 cm [3], [4]. Three gaps remain in the operations record: localization error is measured against the authors’ own offline mapping pipeline rather than an independent metrology system, so the absolute error floor is not established; on-orbit station-keeping and pointing accuracy are described only qualitatively; and flight duration is stated inconsistently, 2.5 to 3 hours on one page against roughly 2 hours on four batteries elsewhere [4]. Bumper impact loads and the worst-case window force are analysis and ground test, not flight experience [4]. The comparison of the three ISS free-flyers rests on a review written by members of those teams, not on an independent evaluation [10].