Canadarm3

NASA/Alberto Bertolin, Bradley Reynolds. Public domain (NASA / US government work).

The problem

Gateway is a lunar-orbit station that is uncrewed for long periods and about 400,000 km from Earth, where the link is limited and delayed [6]. The arm that services it cannot rely on an on-board operator, as Canadarm2 does on the ISS, or on continuous ground supervision [4]. Canadarm3 is therefore specified as a goal-driven autonomous system able to maintain itself and make decisions with minimal human intervention [6].

The task list is broad: catching visiting spacecraft, relocating Gateway modules, maintaining, repairing and inspecting the station, assisting spacewalks and enabling science in lunar orbit [4]. Gateway has no truss and no mobile base, so the arm must also move itself. Canadarm3 is a system, not one arm: an eXploration Large Arm, an eXploration Dexterous Arm, specialized tools, a tool and orbital replacement unit caddy, ground control systems and AI-based control and mission planning software [4]. MDA Space in Brampton, Ontario holds the contract [8]. Delivery is stated as no earlier than 2029 [4], and the NASA ISAM catalog records a first use in 2028 [1].

What has flown

No Canadarm3 hardware has flown. The flown references are the arms it descends from and the free flyers that have run autonomy on orbit.

Shuttle arm and Canadarm2. The Shuttle arm (Canadarm) first deployed in 1981 [6] and was flown by Shuttle crew; it is 15 m long, 410 kg and 6 degrees of freedom [5]. Canadarm2 has serviced the ISS since 2001, is 17 m long, 1497 kg and 7 degrees of freedom, with 270 degrees of rotation in each direction per joint, and is controlled from the ground or by ISS crew [5], [26]. The NASA catalog gives Canadarm2 a reach of 17.6 m against the CSA 17 m [1], [5]. Canadarm2 needs the Mobile Base System and Mobile Transporter to traverse the truss [5].

Dextre. The Special Purpose Dexterous Manipulator is a 15 degree of freedom, two-armed system with a three-step orbital replacement unit exchange sequence [11]. The pre-launch paper argues that ground-commanded maintenance cuts crew time on external work; it reports no on-orbit performance and was written before commissioning [11]. This is the heritage for the dexterous arm.

JEMRMS. The Japanese Kibo arm pairs a 10 m, 780 kg main arm with a 2.2 m, 190 kg small fine arm; the figures are specification values, not measured loads [25].

Free flyers with autonomy. Astrobee has completed more than 165 real-time on-orbit operations and 90 guest science sessions through May 2025 [16]. Int-Ball2 made its first on-orbit flight on 17 October 2023 [18]. On 27 May 2025 a reinforcement-learned controller flew on the Astrobee Bumble in a session of about 7 minutes and 8 maneuvers, reported as the first learned control of a free flyer on the ISS [19]. The ISS intravehicular free flyers are compared in one joint review [17]. These are free-flying robots, not manipulators, so they inform software autonomy practice, not arm dynamics.

What is being tried now

The published work on the problems Canadarm3 has to solve sits in six families. None has flown on an arm with Canadarm3’s configuration.

Walking base and kinematics. Canadarm3 walks end over end between Gateway interfaces, with identical anchoring hands at both ends of the large arm [5], [4]. A free-flying arm has no privileged base, and any link can serve as the base body [12]; the Virtual Manipulator maps a spacecraft-mounted arm to a fixed-base equivalent [13]. Both are analysis with no hardware [12], [13]. For seven-joint station arms, a screw theory inverse kinematics solver for the Chinese space station Wentian arm fixes joint pairs and solves in closed form or by Newton iteration; it is numerical only [24]. No kinematic solver for Canadarm3 has been published.

Base disturbance and free-floating control. Selective decoupling of base momentum along mission-critical axes [33], dynamic-coupling-informed trajectory optimization [32] and real-time hardware-in-the-loop validation of a dual-arm controller [35] are all simulation. A control strategy that repositions the servicer’s center of mass with an external electric-propulsion module is likewise verified only in a multibody simulation [36]. These apply when the base moves; an anchored Gateway arm only approximates that case during berthing of large modules.

Capture of non-cooperative targets. A hierarchical fault-tolerant visual servoing architecture captured a tumbling satellite mockup in four hardware-in-the-loop runs, with a mean position error of 2.4 cm against a 4 cm capture envelope [20]. A controller for a chaser with a robotic arm and sloshing fuel was compared with a reference controller in simulation only [22]. A survey classifies visual servoing methods for servicing arms by recognition, approach and contact phase [21]. These are ground tests and simulations; the surveys of manipulation and capture cover the same ground [2], [3].

Dual-arm, multi-arm and assembly planning. A task-space constrained planner for a tightly coupled dual-arm manipulator [34], a nonlinear predictive handover controller for two arms [31], trajectory optimization for a four-arm walking robot on a target surface [30] and a collision avoidance formulation for autonomous on-orbit assembly [29] are all simulation. The assembly case is closest to Canadarm3’s module relocation task.

Learning and autonomy. Deep reinforcement learning for trajectory planning of teleoperated manipulators under communication delay compares mapping, prediction and state augmentation across four simulated environments [23]. Causal-curiosity reinforcement learning for object classification by a planetary manipulator was tested on a single simulated cube [37]. A lander-manipulator testbed at JPL runs autonomy experiments with a 7 degree of freedom arm on a Stewart platform [38]. An autonomy architecture for a free-flying inspector near an outpost uses keep-in corridors and keep-out zones and is simulation-only [39]. None of these is a Canadarm3 result.

Operations and communications. A DTN gateway on the ISS Ku-band path measured 15 Mbps effective downlink in ground-equivalent tests; flight deployment was in 2016 and the figure is a reference, not a Gateway design [15]. A 2005 overview of autonomous rendezvous, capture and assembly treated every demonstration it named as a plan [14]. Broader reviews of servicing history and architecture [27], the Artemis program plan [28] and the NASA state of play series [9], [10], [1] set the programmatic frame; the three state-of-play editions share verbatim text and count as one catalog.

The numbers

Specifications

Parameter	Value	Source
eXploration Large Arm reach	8.5 m	[5], [1]
eXploration Dexterous Arm reach	3 m	[1]
Degrees of freedom, each arm	7 (3 shoulder, 1 elbow, 3 wrist)	[5], [1]
Mass, large arm	715 kg, estimated	[5]
Boom exterior diameter	0.23 m, carbon fiber composite	[5]
Joint rotation	about 360 deg per joint	[5]
Unloaded tip speed	0.10 m/s	[5]
Loaded tip speed	not yet set	[5]
Cameras	4 color 4K, at the boom swivels and hands	[5]
Sensing	force-moment sensors, collision avoidance, 3D Vision Sensor	[5]
Control	autonomous, with ground or Gateway override	[5]

Comparison with the earlier Canadarms

Parameter	Canadarm	Canadarm2	Canadarm3	Source
Length	15 m	17 m	8.5 m	[5]
Mass	410 kg	1497 kg	715 kg estimated	[5]
Diameter	33 cm	35 cm	23 cm	[5]
Degrees of freedom	6	7	7	[5]
Joint rotation	elbow limited to 160 deg	270 deg each direction	about 360 deg per joint	[5]
Unloaded tip speed	60 cm/s	37 cm/s	10 cm/s	[5]
Loaded tip speed	6 cm/s	2 to 15 cm/s	to be determined	[5]
Cameras	2	4 color	4 color 4K	[5]
Control	Shuttle crew	ground or ISS crew	autonomous, with ground or Gateway override	[5]
Repair	on Earth	in space, replaceable sections	self-detaching sections repaired inside Gateway	[5], [2]
Operator	United States	Canada and United States	Canada	[5]

The CSA table gives four 4K color cameras [5]; a review of the Canadarm series gives six [2]. The NASA catalog separates long reach and short reach manipulators at 8 m, which puts the large arm on the long side and the dexterous arm on the short side [9], [1].

Contract history

Item	Value	Source
Phase A, system definition	complete	[7]
Phases B and C, 26 July 2021	$35.3 million, Gateway External Robotics Interfaces design	[7]
Phases C and D, 27 June 2024	$1 billion, final design plus construction, assembly, integration and test	[8]
Phase C and D period of performance	to March 2030	[8]
Canadian supply chain	more than 200 companies	[8]

The 2024 award covers the flight system: large arm, dexterous arm, specialized tools and ground segment, with commissioning support and training for on-orbit operations [8]. The Gateway External Robotics Interfaces supply power, data and video at the anchoring hands and were planned for delivery ahead of the arms [7], [4].

What is not established

No Canadarm3 performance value is as-built or measured. The 715 kg mass is marked an estimate and the loaded tip speed is to be determined [5]. The CSA and MDA sources are public-affairs pages without test data, drawings or requirement documents [4], [6], [8].

Nothing published covers power draw, thermal design, processor and radiation approach, communications rates, tool inventory, safing modes, positioning accuracy, flexibility of the shorter and thinner boom, or the autonomy architecture. The AI-based control and mission planning software is named as a deliverable but no architecture has been released [6]. Autonomous capture of an uncooperative target remains an open research problem in the surveys [3].

The 2028 first use in the ISAM catalog does not match the CSA date of no earlier than 2029 [1], [4]. The 2021 release tied delivery of the external interfaces to the 2024 HALO and PPE launch dates [7], and the 2020 Artemis plan carried the same dates [28]; neither held. The CSA has announced an intention to repurpose Canadarm3 investments toward the next phase of lunar exploration [4], and the consequence for the Gateway-hosted configuration is not stated in any published source.

References