Multi-agent

More than one vehicle changes what flight software has to decide. State that a single spacecraft holds locally becomes shared state, held on several processors that talk over links that are slow, narrow, noisy or absent, and some agent must decide what the team does next when the network splits. The questions are who allocates work, what is shared and at what cost in bandwidth, how maps and plans are merged, and what each vehicle does when it hears nothing. The classification of collective-behavior algorithms into ten mathematical classes finds that, apart from artificial potential functions, no single technique generalizes across both of the application areas it compares, and its scalability and bandwidth ratings are the authors’ judgment, not benchmark results [3].

Design space

Local rules with no coordination channel are the oldest family. Reynolds’s boids produce flock motion from separation, velocity matching and centering applied to nearby neighbors only, and the paper found that a global central force made a scattered flock converge on its centroid, which does not resemble a flock [1]. The work was a graphics simulation with no sensing error, actuator limit or communication model, so space swarm and formation work inherits the idea of locality and none of its numbers [1]. The same locality appears in the TERMES construction system, where an offline compiler turns a target structure into movement rules that three climbing robots follow using only local sensing, with no global map and no central controller [15]. Distributed sampling-based motion planning for spacecraft swarms is a space-specific variant: agents exchange node positions and sphere radii with neighbors and a distributed assignment step settles which agent takes which target. The reported evidence is two simulation figures with no metrics [14].

Broadcast-state task allocation without negotiation is the second family. ALLIANCE gives each robot a motivation for each task that grows with impatience and resets when a teammate is heard taking the task, so a teammate that stops broadcasting is treated as failed and its task is reallocated by the same mechanism. It ran on three heterogeneous robots in a laboratory with a radio limited to about one message per three seconds, and reported no completion times [2]. A robot taking over a failed teammate’s task there paused 5 to 8 s, a figure set by hand-chosen parameters [2].

Central planning with distributed execution is the third. The JPL rover-team line pairs a central planner that divides science goals among rovers with a CASPER planner on each rover that repairs its own plan when execution deviates [4]. The same prototype fed clustering of spectrometer readings into goal generation for three simulated rovers [5]. A later evaluation of goal interdependency used a geologist-derived reward model and randomly generated problems, with no rover or terrain in the loop [6]. ShAC coordinates shared activities during execution between two simulated Mars Exploration Rovers and the Odyssey orbiter over a simple master/slave protocol, on one scenario with three agents [7]. All four are simulation. The spacecraft analog is the TechSat 21 Autonomous Sciencecraft design, a cluster of SAR satellites with onboard planning; the program was canceled and never flew, so its timing figures are estimates [11].

Physical coupling is a fourth family, in which the carried object is the communication channel. CAMPOUT, a three-layer behavior-based architecture, was run with two Sample Return Rovers carrying an extended payload over natural terrain in the Arroyo Seco, with a single unreplicated traverse of 40 to 50 m on slopes up to 9 degrees and a final heading error under 1 degree [8]. A companion paper designed coordination to avoid inter-rover radio traffic by sensing force and yaw at the instrumented gimbals; its body contains no hardware experiment and its results are simulation [9]. A distributed model-predictive controller has since carried a 475 kg mock cargo with two 800 kg rovers on Earth analog terrain at 0.5 m/s, over three loop repeats of one route [10].

Formation flying is the orbital counterpart. SPHERES, designed at MIT and flown on the ISS, was validated first in laboratory and KC-135 tests, and its prototype hardware numbers do not describe the flown satellite [12]. In the March 2007 ISS session two satellites starting back to back with pointing errors of 150 to 167 degrees reacquired each other with limited-field-of-view ultrasound sensors; multipath off the station walls delayed convergence by 55 to 58 s [13].

Outside the space branches, the same coordination problems have been solved with fewer constraints. JPL’s CARACaS ran four full-size unmanned surface vessels with no human in the tasking loop, allocating Patrol, Track, Inspect and Trail tasks; the environment is two-dimensional and obstacle sparse and nothing in it bounds behavior under flight communication delay [16]. A planning formulation for underwater gliders that must surface periodically to bound dead-reckoning error treats synchronization as a decision, which is the shape of the planetary relay problem; the transfer is by analogy and the results are a case study [17].

What changed and what it enabled

Field teams showed how little supervision a communication-degraded team needs. In one Safety Research Mine run, Team CoSTAR’s operator issued 16 missions to two Vanguard robots and one Support robot, 4 of them communication node drops each taking 1 to 3 minutes, and the evidence is a single run [18]. Multi-robot SLAM scaled by computing only the loop closures predicted to reduce uncertainty most, tested on underground competition data and not on planetary terrain [19]. Learned and communication-aware methods are the newest strand: a flying scout that chooses where to transmit to a specialist by signal strength [29], and behavior-adaptive hazard mapping for heterogeneous robots without communication [30]. Both are simulation only, with modeled signal maps or assumed failure rates.

The CADRE technology demonstration, three rovers and a base station, is the first multi-agent autonomy stack built for flight on the Moon. Its architecture elects a leader, which plans for the team, while each agent runs its own motion planner and surface mobility stack [20]. A leader elected every ten seconds bounds the idle time after leader loss at 10 s, with planning state held on a designated survivor [21]. The election algorithm starts from Gallager-Humblet-Spira spanning-tree construction and ranks candidates by battery state of charge and CPU temperature; it was demonstrated with four software processes and not on the radio mesh [22]. Exploration is semi-centralized: the leader partitions unexplored area by k-means and each rover then explores its own region frontier by frontier, which removes inter-robot traffic. Analog testing covered 2724 m^2 at 1 g under flood lighting [23]. The shared database, MoonDB, rides a mesh radio with 1 Mbps aggregate bandwidth and motor interference that restricts transfer to time windows, and the paper reports no measured utilization or loss under that link [24]. Federated mapping would send neural map parameters instead of maps, but was validated on a glacier dataset and simulated Martian terrain, not CADRE data [25]. Assembly scheduling for the proposed Lunar Crater Radio Telescope is posed as a mixed integer program with time-varying travel times; the solver is stopped before certifying optimality [26]. A JPL infusion survey lists Mars PUFFER, DASHER, SWARMS and CADRE among the multi-agent efforts, with no performance data [31].

What has flown

Flown: CADRE launched on IM-3 for Reiner Gamma, and SPHERES formation acquisition ran aboard the ISS [20][13]. Flight results from CADRE are not in the held record; the sources above predate flight [21]. Demonstrated on Earth hardware: CAMPOUT cooperative transport [8], the shared-cargo controller [10], TERMES [15], and the CoSTAR heterogeneous team [18]. Simulation or proposal only: the CASPER rover-team planners [4][5][6], ShAC [7], TechSat 21 [11], federated mapping [25], and the communication-aware planners [29][30].

The relay problem is the multi-agent system already operating on Mars. Perseverance relay planning precomputes data volume for each orbiter overflight and filters against passes per sol and latency to the next decisional pass [27]. Mars surface-to-surface radio propagation is terrain-dominated, shown by ray tracing over Jezero Crater terrain with no surface measurement to check it against [28].

What is not established

No source in the held record reports CADRE flight performance: election latency, convergence time and message counts are not given for the leader election [22], and the database design is argued rather than measured under the EMI-limited link [24]. Most task-allocation work for rover teams has been run in simulation with no quantitative result, including the central-plus-onboard planners [4][5] and the coordination protocols [7]. The behavior-based and motivation-based architectures were tested on a few robots in benign settings, and no scaling study of swarm size exists in the planning papers [2][14]. The broadcast-state assumption has no test under the lossy, contended links a planetary team would see [2]. Partition behavior, meaning what a team does when the leader and members cannot hear each other, is argued in the CADRE design papers and not measured [21][22]. Results from maritime, underwater and underground teams [16][17][19] bound the algorithms and not the radiation, delay or processor limits of flight.