Pith. sign in

REVIEW 5 major objections 5 minor 53 references

MOSAIC claims that expressing every mission objective as a shared Point of Interest, managed centrally while robots execute them locally with layered autonomy, lets a single operator supervise a heterogeneous five-robot science team that ke

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 06:14 UTC pith:43CSZOO6

load-bearing objection A solid field-robotics engineering report whose abstract oversells the failure-robustness evidence: the robot that 'failed' was already out of the mission. the 5 major comments →

arxiv 2601.23038 v3 pith:43CSZOO6 submitted 2026-01-30 cs.RO

MOSAIC: Modular Scalable Autonomy for Intelligent Coordination of Heterogeneous Robotic Teams

David Oberacker , Julia Richter , Philip Arm , Marvin Grosse Besselmann , Lennart Puck , William Talbot , Maximilian Schik , Sabine Bellmann
show 5 more authors
Tristan Schnell Hendrik Kolvenbach R\"udiger Dillmann Marco Hutter Arne Roennau
This is my paper
classification cs.RO
keywords multi-robot systemsscalable autonomyfield roboticspoints of interesttask allocationheterogeneous teamsspace explorationsingle-operator supervision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that scalable autonomy for multi-robot scientific exploration is achievable by separating what the team must achieve from how each robot executes: every mission objective is expressed as a Point of Interest (POI) at a central mission level, and each robot autonomously claims and executes POIs it is capable of handling. This layering of autonomy — mission, robot task, and direct teleoperation — allows a single operator to supervise a heterogeneous team while keeping human attention for only the highest-level decisions and interventions. In a lunar-analog field test with five robots, the team completed 82.3% of assigned tasks with an Autonomy Ratio of 86%, and continued operating after one of the five robots failed completely. The working claim is that team-level redundancy, enabled by role specialization and dynamic POI allocation, is what gives resilience rather than any single capable platform.

Core claim

MOSAIC's central discovery is that a unified mission abstraction (POIs) combined with layered autonomy lets a small team of heterogeneous robots run a scientific exploration mission under one operator, and that the design absorbs individual robot failures without mission abort. The system keeps a global mission state on a mission-control station, while every robot runs its own POI planner and behavior tree that ranks and claims objectives by utility (POI type, distance, navigation cost, battery). Because roles such as scout and scientist are each covered by multiple robots, a failed robot's objectives are naturally reallocated to its teammates. The field evaluation backs this up: despite the

What carries the argument

The central object is the Point of Interest (POI) — a location in a shared earth frame, tagged with a type and metadata — which serves as the single mission-level unit of progress. Around it, the framework's key machinery is the three-layer autonomy stack (mission level: central POI management and operator interface; robot level: local POI planner plus behavior-tree execution; driver level: direct teleoperation) and an iterative, multi-step greedy planner that computes per-robot utilities from weighted features (type suitability, Euclidean and navigation distance, battery use), broadcasts planned utilities to teammates to avoid conflicts, and prunes distant candidates early. Role definition

Load-bearing premise

The framework assumes that one initial alignment of the robots' maps keeps them in a common coordinate frame for the entire mission; without continuous co-localization, drift will eventually make a POI mean different physical locations for different robots.

What would settle it

Run the system on a site larger than 60x40 m or for longer than 40 minutes without re-localizing robots, then compare a known landmark's coordinates in each robot's map: if the inter-robot error grows beyond the reach or sensor coverage of the assigned robot, the shared-POI abstraction, and with it the scalability claim, fails. The paper's own map-error measurement (0.145 m ± 0.252 m) gives a starting baseline to extrapolate from.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • With this framework, a single operator can run a five-robot exploration and measurement mission, with robots spending most of the mission in full autonomy and the operator intervening mainly for mission-level decisions.
  • The POI abstraction means a robot failure does not stop the mission: tasks are reallocated among remaining robots, as happened when the scout with automatic rock detection was lost.
  • The architecture's separation of mission-level objectives from robot-specific execution makes it possible to add or swap robot types without changing the mission representation.
  • Network design choices (best-effort topics, domain separation, throttled transforms) reduce wireless bandwidth to about 20 Mbit/s for a five-robot team over one Wi-Fi access point.
  • The reported KPIs (82.3% task success, 86% autonomy ratio) define a baseline for evaluating future multi-robot exploration systems and isolate operator workload as a key scaling bottleneck.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The scalability in the title is bounded by the 40-minute, 60x40 m validation: the shared-world consistency rests on a one-time ICP alignment, so longer missions or larger areas would need continuous cooperative localization before the same architecture can be claimed to scale.
  • Because resource identification (56.2%) is far below task completion (82.3%), the framework's autonomy is strongest at executing objectives, not at deciding what is scientifically interesting; a testable improvement would be to automate more of the target-detection and measurement-point refinement that now relies on the scientist-in-the-loop.
  • The operator workload of 78.2% is close to saturation, so scaling to ten robots would likely require the automated operator task queue and a dedicated teleoperator that the authors propose as future work; a worthwhile experiment is to measure how workload scales as team size grows under the same layered-autonomy design.
  • The reported lessons — e.g., favoring homogeneous software stacks, avoiding mixed-version middleware bridges, and maintaining scout-to-scientist ratio — can be read as a prescriptive recipe for future planetary missions, and each could be turned into a controlled comparison study.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper presents MOSAIC, a framework for supervised autonomy of heterogeneous multi-robot teams built around a mission-level Points-of-Interest abstraction, multiple autonomy levels (mission, task, driver), and a mixed ROS 1/ROS 2 communication architecture. The system was field-tested in a lunar-analog scenario with five robots — three scouts and two scientists — over a 40-minute mission. The authors report 82.3% task success, 86% Autonomy Ratio, 78.2% operator workload, and argue that the team remained effective despite the complete loss of one robot. The paper also contributes practical lessons on interoperability, networking, team composition, and operator workload.

Significance. If the headline claims were fully supported, MOSAIC would be a useful contribution to heterogeneous multi-robot field autonomy, particularly in space-analog settings. The strengths of the paper are genuine: a real field deployment with heterogeneous legged and wheeled platforms, directly reported trajectories and maps, concrete measurement examples, and an unusually candid discussion of the engineering difficulties encountered (ROS 1/2 bridging, DDS QoS, drift, operator workload). The feasibility of the integrated system is demonstrated. However, the evidence base is a single 40-minute run with no baseline or repetitions, and several KPI definitions are post-hoc or self-referential. The central robustness claim — that the framework absorbs the complete mid-mission failure of a robot — is not supported by the reported data, and the scalability claim is acknowledged by the authors to require co-localization work that is not yet implemented.

major comments (5)
  1. [Abstract, §VI.A, §VIII] The abstract and conclusion claim that the team continued despite 'the complete failure of one robot during the mission.' Section VI.A states that Dodo suffered water damage and was taken out of operation during the first two days, before the final campaign; the reported results refer to the final campaign using 'all functioning robots.' Figure 8 shows trajectories and activities for Spot, Dilly, Donkey, and Husky only, and Spot's software/networking issue at the start prevented its behavior tree but did not cause a complete loss — Spot was used as a teleoperated camera platform. Thus the key robustness demonstration is not present in the reported run. Either report a mission with a genuine mid-mission total failure or revise the claims to describe the actual scenario (a robot lost before the final run).
  2. [§VI.C, Table 1] The 'Task Success Ratio' of 82.3% is the success rate of scientific measurements performed by Donkey and Husky only (14 of 17 attempts: Donkey 5/5 ground and 3/5 rock; Husky 6/7 ground). It is not a team-level or mission-level success metric covering exploration, mapping, and POI handling. The abstract's phrase 'completed 82.3% of assigned tasks' overstates what was measured. With only 17 attempts, no confidence interval is reported. Please report per-robot and per-POI-type counts with uncertainty, and use a mission-level definition if that is the intended claim.
  3. [§VI.A, Table 1] The Robot Downtime and Autonomy Ratio results rely on excluding Husky's 80.8% apparent planning time as 'incorrect state labeling' rather than actual inactivity. This post-hoc exclusion changes total Robot Downtime from 52% to 37.8%, a large swing in a headline KPI. The paper does not provide independent evidence — e.g., log excerpts showing that Husky was executing POIs while its behavior tree failed to record them. If the state labeling is unreliable, then the logged state data underlying this KPI is itself questionable. Please justify the exclusion with concrete log evidence or report both values without asserting which is correct.
  4. [§IV.D, §VII] The shared mission abstraction depends on a consistent earth frame across robots, but co-localization is performed only once at startup via ICP alignment; Section IV.D states that no continuous co-localization is performed, and Section VII acknowledges that a reliable co-localization strategy is needed for longer deployments. The reported 60 m × 40 m, 40-minute experiment therefore cannot validate the 'scalable' claim, which is central to the paper's title and motivation. Please either add a quantitative drift analysis over the mission and a clear statement of the operational envelope, or explicitly restrict the scalability claim to future work.
  5. [§VI, Table 1] All headline numbers — Task Success Ratio, Autonomy Ratio, Quantitative Operator Workload, mapping rates, and Map Error — are single-run point estimates from one 40.3-minute mission with no baseline comparison, no repetitions, and no confidence intervals (except the Map Error standard deviation). For a validation claim about robustness and scalability, this is insufficient. The authors should at least state explicitly that this is a single case study, and ideally include repeated runs or, if that is not feasible, frame the results as a feasibility demonstration rather than a performance validation.
minor comments (5)
  1. [Eq. (1)] The formula for the depth-weighted utility is ambiguous as typeset: 'utility × duf^current_depth' appears as 'utility × duf current_depth'. Please clarify whether the depth_uncertainty_factor is raised to the current depth or multiplied by it, and define how the weights in the utility features are configured.
  2. [§VI.C] The text refers to 'Husky’s RFA scan,' but the instrument described in §IV.B.2 is an XRF (Olympus Vanta). Please correct the acronym.
  3. [Table 1 / §VI.B] The 'Ratio of Identified Resources' (56.2%) is derived from human identification because Dodo, the autonomous rock-detection robot, was unavailable; Section VI.C states that 'the ratio of autonomously identified rocks could not be determined.' This distinction should be made explicit in the table or its caption so the KPI is not misread as an autonomy metric.
  4. [Figure 8] The captions for panels (a) and (b) appear to be duplicated in the typeset manuscript. Please correct the caption for panel (b) to match the operator timeline it displays.
  5. [§IV.C.3] The POI planning depends on several operator-defined weights (type reward, distance, battery, and the depth_uncertainty_factor). No sensitivity analysis is provided, and the paper does not state how these weights were chosen. A brief discussion of sensitivity or a reference to the companion configuration would help readers judge the generality of the allocation behavior.

Circularity Check

0 steps flagged

No circular derivation; field-data evaluation has independent content, though the robustness example is not clearly supported and co-localization is acknowledged as open.

full rationale

The paper's only formal rule is Eq. (1), utility_weighted = utility × duf_current_depth, which discounts utility by planning depth; no reported output is defined from itself. The headline quantities (82.3% Task Success Ratio, 86% Autonomy Ratio, 78.2% Quantitative Operator Workload) are post-hoc counts from logged mission data (e.g., Donkey 14 attempts with 12 successful, Husky 9 attempts with 7 successful; RAD = IE/(IE+NT)), so they are measured outcomes, not forced predictions. The evaluation does use the authors' KPI framework [15] and the architecture builds on [13], but neither citation defines the outcome: the team could plausibly have scored worse. The genuine weaknesses are evidence issues rather than circularity. Section VI.A reports that Dodo "suffered water damage and had to be taken out of operation" during the first two days, before the final campaign, whereas the abstract's "complete failure of one robot during the mission" suggests a mid-mission failure in the reported run. Section IV.D states "After the initial alignment of the robot maps, no continuous co-localization is performed," and Section VII concedes "a reliable co-localization strategy is needed" for longer deployments; these limit scalability support but are not self-referential derivations. Score 2 reflects two non-load-bearing self-references (KPI definitions and base architecture), not equivalence-to-input or prediction-by-construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on hand-tuned planning parameters and several domain assumptions about localization, networking, and metric validity. The paper is transparent about co-localization and workload limitations, but the headline KPI numbers are obtained after post-hoc exclusions and without an external benchmark.

free parameters (5)
  • POI utility feature weights
    Operator-defined configuration weights type reward, Euclidean/navigation distance, and battery features in POI planning (Sec. IV.C.3); values are not reported, and they directly set which POIs each robot selects.
  • depth_uncertainty_factor (duf) = ≤1.0, value not stated
    Multiplies utility at each planning depth in Eq. (1); chosen to represent planning uncertainty, with no stated calibration procedure.
  • planning depth = not stated
    Number of look-ahead POI steps in the greedy planner; controls the trade-off between horizon and computation (Sec. IV.C.3).
  • POI type rewards
    Fixed positive/zero rewards per POI type per robot (Sec. IV.C.2–C.3); hand-set to encode robot capabilities and mission priorities.
  • Husky motion penalty adjustment = high penalties, values not stated
    Introduced before the final mission because of observed slippage; changed Husky's preference for nearby POIs (Sec. VI.C) and is mission-specific tuning.
axioms (5)
  • domain assumption No continuous co-localization; the initial ICP alignment of local maps remains valid for the mission duration.
    Sec. IV.D states that after initial alignment each robot uses its own SLAM and no online inter-robot registration is performed. If drift accumulates, POIs in the shared earth frame become inconsistent; the paper acknowledges this fails for longer/larger missions (Sec. VII).
  • domain assumption Single-access-point Wi-Fi with the QoS design supports the mission communication requirements.
    Sec. IV.G and VII describe the networking and ~20 Mbit/s bandwidth; the system depends on this physical-link assumption.
  • domain assumption The KPI metrics from [15] quantify operator workload and autonomy as intended.
    Sec. VI uses RAD/IE/NT-based Autonomy Ratio and Quantitative Operator Workload from a same-group companion paper; no independent validation of these metrics is provided.
  • ad hoc to paper Husky's high apparent planning time (80.8%) was an incorrect state-labeling bug, not real inactivity.
    Sec. VI.A: "we attribute this downtime largely to incorrect state labeling, likely caused by a missing claim confirmation in the behavior tree, rather than actual inactivity." This assumption justifies excluding Husky from the headline downtime KPI.
  • domain assumption Human-in-the-loop validation and operator-created POIs are part of the system, not external assistance.
    The mission relies on the operator placing/refining POIs and confirming rock candidates (Sec. V.C, IV.F). The 'autonomy ratio' therefore measures autonomy in execution, not full autonomy in objective generation.
invented entities (1)
  • Mission-level POI abstraction no independent evidence
    purpose: Unified objective representation decoupling mission goals from robot-specific execution (Sec. III.A.1).
    This is a design abstraction rather than an empirically discovered entity; its validity is evidenced only by the reported deployment. No independent falsifiable handle is provided outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 27294 in / 16668 out tokens · 175921 ms · 2026-08-03T06:14:56.431641+00:00 · methodology

0 comments
read the original abstract

Mobile robots have become indispensable for exploring hostile environments, such as in space or disaster relief scenarios, but often remain limited to teleoperation by a human operator. This restricts the deployment scale and requires near-continuous low-latency communication between the operator and the robot. We present MOSAIC: a scalable autonomy framework for multi-robot scientific exploration using a unified mission abstraction based on Points of Interest (POIs) and multiple layers of autonomy, enabling supervision by a single operator. The framework dynamically allocates exploration and measurement tasks based on each robot's capabilities, leveraging team-level redundancy and specialization to enable continuous operation. We validated the framework in a space-analog field experiment emulating a lunar prospecting scenario, involving a heterogeneous team of five robots and a single operator. Despite the complete failure of one robot during the mission, the team completed 82.3% of assigned tasks at an Autonomy Ratio of 86%, while the operator workload remained at only 78.2%. These results demonstrate that the proposed framework enables robust, scalable multi-robot scientific exploration with limited operator intervention. We further derive practical lessons learned in robot interoperability, networking architecture, team composition, and operator workload management to inform future multi-robot exploration missions.

Figures

Figures reproduced from arXiv: 2601.23038 by Arne Roennau, David Oberacker, Hendrik Kolvenbach, Julia Richter, Lennart Puck, Marco Hutter, Marvin Grosse Besselmann, Maximilian Schik, Philip Arm, R\"udiger Dillmann, Sabine Bellmann, Tristan Schnell, William Talbot.

Figure 1
Figure 1. Figure 1: Field experiments demonstrating multi-robot operation and locomotion capabilities in unstructured outdoor terrain. (a) Representative example of the test site with POI visualization overlaid on the map (?: unclaimed, robot icon: currently executing, lines: planned path) (b-d) Representative examples of the test site with multiple platforms operating concurrently. (e-g) Dilly traversing a steep, loose slope… view at source ↗
Figure 2
Figure 2. Figure 2: MOSAIC software architecture overview of both individual robots and the overall control system. The system-level components exist once in the system, robot-specific components exist on every robot in the team. Communication is achieved on a per-robot unique ROS 2 domain. Shared communication happens on the shared ROS 2 domain. irregularities, and other robots to ensure safe and feasible motion. The Mapping… view at source ↗
Figure 3
Figure 3. Figure 3: Custom-built mission control (a) and touchscreen science operations station (b) with key interface elements highlighted. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the heterogeneous robotic team used in this work. (a–c) Scout robots. (d) Overview of the full deployed team. (e–f) Scientist robots carrying manipulation and scientific payloads for close-range interaction and in-situ analysis. carrying specialized payloads. Their objective is to conduct detailed characterization and in-situ measurements once the scouts identify potential POIs. We employ two s… view at source ↗
Figure 5
Figure 5. Figure 5: Trenching maneuver with the right front leg executed by Dodo on loose terrain. (a-d) Consecutive frames illustrating the leg motions and body progression during the maneuver. dedicated controllers [12]: locomotion is realized through a perceptive reinforcement learning policy [30], while the arm is operated through a model predictive control framework [35]. Equivalently to the other ANYmals, the robot-leve… view at source ↗
Figure 6
Figure 6. Figure 6: The main autonomy behavior tree for each robot in ros_bt_py [38]. On each robot, the subtrees handling the different POI types vary depending on the robot’s capabilities. This tree runs for the entire duration of the mission. POI Planning The two introduced components are co￾ordinated by the POI planning component, which selects the most suitable POI for a robot by evaluating and comparing candidate POIs b… view at source ↗
Figure 7
Figure 7. Figure 7: Overview of the experimental test site and sampling targets. a) Target map derived from drone imagery, indicating measurement targets across the site. Rock targets are indicated by orange circles, while ground targets (colored sand patches) are marked in blue. b-c) Representative ground-level views of the test site with targets marked for correspondence with the aerial map. A. Mission Design and Rationale … view at source ↗
Figure 8
Figure 8. Figure 8: Mission overview, including trajectory overview (a) and activity timeline (b). ecuted multiple scientific measurements, while Husky per￾formed several ground measurements near the lander area. Due to high penalties assigned to prolonged motion — intro￾duced based on observed slippage in the days preceding the mission — Husky’s autonomous behavior favored POIs close to its current position. Some measurement… view at source ↗
Figure 9
Figure 9. Figure 9: Multi-robot mapping results. a) Aggregated point cloud (grey) with concave hull (orange) outlining the mapped area used for area computation (blue). b) Assembled point cloud colored by height. a substantial set of high-quality microscope measurements. These results, several of which are shown in [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative scientific inspection and measurement tasks performed during field experiments. (a) Dilly inspecting a rock target using a thermal camera. (b) Corresponding thermal image of the same target. (c-e) Donkey performing rock measurement with the RAMAN spectrometer (c) and the microscope (d-e). (f) Donkey performing a ground measurement with the microscope on a green sand path. (g) Husky performi… view at source ↗
Figure 11
Figure 11. Figure 11: Example microscope measurements. (A) Rock measurement: wide-angle camera image of a rock with the measurement location highlighted in blue. (B) Ground measurement: infrared image from the onboard stereo camera showing blue sand. For each measurement, microscope images (a–f) illustrate the same sample under different illumination: (a) red, (b) green, (c) blue, (d) white, (e) UV, and (f) infrared light. and… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 8 canonical work pages · 1 internal anchor

  1. [1]

    How are your robot friends doing? a design exploration of graphical techniques supporting awareness of robot team members in teleoperation,

    Seo, S. H., Young, J. E., and Irani, P., “How are your robot friends doing? a design exploration of graphical techniques supporting awareness of robot team members in teleoperation,”International Journal of Social Robotics, vol. 13, no. 4, pp. 725–749, Jul. 1, 2021, ISSN: 1875-4805. DOI:10.1007/s12369-020-00670-9

    work page doi:10.1007/s12369-020-00670-9 2021

  2. [2]

    Still not solved: A call for renewed focus on user-centered teleoperation interfaces,

    Rea, D. J. and Seo, S. H., “Still not solved: A call for renewed focus on user-centered teleoperation interfaces,”Frontiers in Robotics and AI, vol. 9, Mar. 29, 2022, ISSN: 2296-9144. DOI:10.3389/frobt.2022. 704225

    doi:10.3389/frobt.2022 2022

  3. [3]

    Inspiring field robotics advances through the design of the DARPA subterranean challenge,

    Orekhov, V. L., Maio, A. C., Daniel, R. P., and Chung, T. H., “Inspiring field robotics advances through the design of the DARPA subterranean challenge,” Field Robotics, vol. 3, pp. 560–604, Apr. 2023, ISSN: 2771-3989. DOI:10.55417/fr.2023018

    work page doi:10.55417/fr.2023018 2023

  4. [4]

    Toward multi user knowledge driven teleop- eration of a robotic team with scalable autonomy,

    Schmaus, P. et al., “Toward multi user knowledge driven teleop- eration of a robotic team with scalable autonomy,” in2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC) , Oct. 2023, pp. 867–873. DOI:10.1109/SMC53992.2023.10394365

    arXiv 2023

  5. [5]

    Curiosity: The mars science laboratory project,

    Cook, R. A., “Curiosity: The mars science laboratory project,” pre- sented at the International Astronautical Federation, Paris, France, Oct. 1, 2012

    2012

  6. [6]

    Overview and results from the mars 2020 perseverance rover’s first science campaign on the jezero crater floor,

    Sun, V. Z. et al., “Overview and results from the mars 2020 perseverance rover’s first science campaign on the jezero crater floor,” Journal of Geophysical Research: Planets , vol. 128, no. 6, e2022JE007613, 2023, ISSN: 2169-9100. DOI: 10.1029/ 2022JE007613

    2020

  7. [7]

    S., Sadler, B

    Prorok, A., Malencia, M., Carlone, L., Sukhatme, G. S., Sadler, B. M., and Kumar, V., Beyond robustness: A taxonomy of ap- proaches towards resilient multi-robot systems , Sep. 25, 2021. DOI: 10.48550/arXiv.2109.12343

    Pith/arXiv arXiv doi:10.48550/arxiv.2109.12343 2021

  8. [8]

    [Online]

    NASA / JPL, Mars rover mission comes to an end but science legacy lives on. [Online]. Available:https://www.jpl.nasa.gov/edu/ resources/teachable-moment/mars-rover-mission-comes-to-an- end-but-science-legacy-live-on/

  9. [9]

    The challenge 2021-2022,

    European Space Agency and European Space Resources Innovation Centre. “The challenge 2021-2022,” SPACE RESOURCES CHAL- LENGE, Accessed: Sep. 10, 2025. [Online]. Available:https://src. esa.int/the-challenge-2021-2022/

    2021

  10. [10]

    Cerberus in the darpa subterranean challenge,

    Tranzatto, M. et al., “Cerberus in the darpa subterranean challenge,” Science Robotics, vol. 7, no. 66, eabp9742, 2022. DOI:10.1126/ scirobotics.abp9742

    2022

  11. [11]

    Heterogeneous robot teams with unified per- ception and autonomy: How team CSIRO data61 tied for the top score at the DARPA subterranean challenge,

    Kottege, N. et al., “Heterogeneous robot teams with unified per- ception and autonomy: How team CSIRO data61 tied for the top score at the DARPA subterranean challenge,”IEEE Transactions on Field Robotics, vol. 2, pp. 100–130, 2025, ISSN: 2997-1101. DOI: 10.1109/TFR.2024.3522063

    arXiv 2025

  12. [12]

    Scientific exploration of challenging planetary ana- log environments with a team of legged robots,

    Arm, P. et al., “Scientific exploration of challenging planetary ana- log environments with a team of legged robots,”Science Robotics, vol. 8, no. 80, eade9548, Jul. 12, 2023. DOI:10.1126/scirobotics. ade9548 23 Oberacker et al. : MOSAIC: Modular Scalable Autonomy for Intelligent Coordination of Heterogeneous Robotic Teams

    doi:10.1126/scirobotics 2023

  13. [14]

    REALMS: Resilient exploration and lunar mapping system,

    Meer, D. van der et al., “REALMS: Resilient exploration and lunar mapping system,” Frontiers in Robotics and AI , vol. 10, Mar. 30, 2023, ISSN: 2296-9144. DOI:10.3389/frobt.2023.1127496

    arXiv 2023

  14. [15]

    A Practical Framework of Key Performance Indicators for Multi-Robot Lunar and Planetary Field Tests

    Richter, J. et al.,A practical framework of key performance indi- cators for multi-robot lunar and planetary field tests , 2026. DOI: 10.48550/ARXIV.2601.20529

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.20529 2026

  15. [16]

    Flexible supervised autonomy for exploration in subterranean environments,

    Biggie, H. et al., “Flexible supervised autonomy for exploration in subterranean environments,” Field Robotics, vol. 3, pp. 125–189,

  16. [17]

    Autonomous exploration development environment and the planning algorithms,

    Cao, C. et al., “Autonomous exploration development environment and the planning algorithms,” in2022 International Conference on Robotics and Automation (ICRA), IEEE, 2022, pp. 8921–8928. DOI: 10.1109/ICRA46639.2022.9812330

    arXiv 2022

  17. [18]

    Cooperative robotic exploration of a plan- etary skylight surface and lava cave,

    Domínguez, R. et al., “Cooperative robotic exploration of a plan- etary skylight surface and lava cave,”Science Robotics, vol. 10, no. 105, eadj9699, Aug. 13, 2025. DOI:10.1126/scirobotics.adj9699

    work page doi:10.1126/scirobotics.adj9699 2025

  18. [19]

    LUNARES: Lunar crater exploration with hetero- geneous multi robot systems,

    Cordes, F. et al., “LUNARES: Lunar crater exploration with hetero- geneous multi robot systems,”Intelligent Service Robotics , vol. 4, no. 1, pp. 61–89, 2011. DOI:10.1007/s11370-010-0081-4

    work page doi:10.1007/s11370-010-0081-4 2011

  19. [20]

    A mission architecture for a human-robot collaborative planetary exploration cascade,

    Jusner, S., Moser, S., Schaffler-Glößl, S., Halatschek, R., Eder, M., and Steinbauer-Wagner, G., “A mission architecture for a human-robot collaborative planetary exploration cascade,” in2024 International Conference on Space Robotics (iSpaRo) , IEEE, 2024, pp. 321–327. DOI:10.1109/iSpaRo60631.2024.10687418

    arXiv 2024

  20. [21]

    The ARCHES space-analogue demonstration mission: Towards heterogeneous teams of autonomous robots for collaborative scientific sampling in planetary exploration,

    Schuster, M. J. et al., “The ARCHES space-analogue demonstration mission: Towards heterogeneous teams of autonomous robots for collaborative scientific sampling in planetary exploration,” IEEE Robotics and Automation Letters , vol. 5, no. 4, pp. 5315–5322,

  21. [22]

    Fieldtestingofacooperativemulti-robotsample return mission in mars analogue environment,

    Sonsalla,R.etal.,“Fieldtestingofacooperativemulti-robotsample return mission in mars analogue environment,” inProceedings of the 14th symposium on advanced space technologies in robotics and automation (ASTRA), 2017

    2017

  22. [23]

    Heterogeneous ground and air platforms, homo- geneous sensing: Team CSIRO data61’s approach to the DARPA subterranean challenge,

    Hudson, N. et al., “Heterogeneous ground and air platforms, homo- geneous sensing: Team CSIRO data61’s approach to the DARPA subterranean challenge,”Field Robotics, vol. 2, pp. 595–636, 2022. DOI: 10.55417/fr.2022021

    work page doi:10.55417/fr.2022021 2022

  23. [24]

    Crazyswarm: A large nano-quadcopter swarm,

    Preiss, J. A., Honig, W., Sukhatme, G. S., and Ayanian, N., “Crazyswarm: A large nano-quadcopter swarm,” in 2017 IEEE International Conference on Robotics and Automation (ICRA) , May 2017, pp. 3299–3304. DOI:10.1109/ICRA.2017.7989376

    arXiv 2017

  24. [25]

    Nebula: Quest for robotic autonomy in challenging environments; team costar at the darpa subterranean challenge,

    Agha, A. et al., “Nebula: Quest for robotic autonomy in challenging environments; team costar at the darpa subterranean challenge,” arXiv preprint arXiv:2103.11470, 2021. DOI:10.48550/arXiv.2103. 11470

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.11470 2021

  25. [26]

    Robot operating system 2: Design, architecture, and uses in the wild,

    Macenski,S.,Foote,T.,Gerkey,B.,Lalancette,C.,andWoodall,W., “Robot operating system 2: Design, architecture, and uses in the wild,” Science Robotics, vol. 7, no. 66, eabm6074, May 11, 2022. DOI: 10.1126/scirobotics.abm6074

    doi:10.1126/scirobotics.abm6074 2022

  26. [27]

    ROS: An open-source robot operating system,

    Morgan Quigley et al., “ROS: An open-source robot operating system,” in 2009 IEEE International Conference on Robotics and Automation (ICRA), Kobe, Japan: IEEE, 2009

    2009

  27. [28]

    ANYmal – autonomous robotic inspection solu- tion

    ANYbotics AG. “ANYmal – autonomous robotic inspection solu- tion. ”[Online]. Available:https://www.anybotics.com/robotics/ anymal/

  28. [29]

    Spot®, the agile mobile robot

    Boston Dynamics, Inc. “Spot®, the agile mobile robot. ”[Online]. Available:https://bostondynamics.com/products/spot/

  29. [30]

    Learning robust perceptive locomotion for quadrupedal robots in the wild,

    Miki, T., Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., and Hutter, M., “Learning robust perceptive locomotion for quadrupedal robots in the wild,” Science Robotics, vol. 7, no. 62, eabk2822, Jan. 19, 2022. DOI:10.1126/scirobotics.abk2822

    doi:10.1126/scirobotics.abk2822 2022

  30. [31]

    Pedipulate: Enabling manipulation skills using a quadruped robot’s leg,

    Arm, P., Mittal, M., Kolvenbach, H., and Hutter, M., “Pedipulate: Enabling manipulation skills using a quadruped robot’s leg,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), May 2024, pp. 5717–5723. DOI:10.1109/ICRA57147.2024. 10611307

    arXiv 2024

  31. [32]

    [Online]

    Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y., and Girshick, R., Detectron2, 2019. [Online]. Available: https : / / github . com / facebookresearch/detectron2

    2019

  32. [33]

    Bdaiinstitute/spot_ros2, Nov. 3, 2025. Accessed: Nov. 4, 2025. [Online]. Available:https://github.com/bdaiinstitute/spot_ros2

    2025

  33. [34]

    DynaArm – ultra-lightweight and weatherproof robotic arm

    Duatic AG. “DynaArm – ultra-lightweight and weatherproof robotic arm. ”[Online]. Available:https://duatic.com/products/dynaarm/

  34. [35]

    et al., OCS2: An open source library for optimal control of switched systems , 2025

    Farshidian, F. et al., OCS2: An open source library for optimal control of switched systems , 2025. [Online]. Available: https:// github.com/leggedrobotics/ocs2

    2025

  35. [36]

    MIRA XTR,

    “MIRA XTR,” Accessed: Jan. 19, 2026. [Online]. Available:https: //www.metrohm.com/de_de/produkte/raman-spektroskopie/mira- xtr.html

    2026

  36. [37]

    VantaHandheld XRF Analyzer,

    “VantaHandheld XRF Analyzer,” Accessed: Nov. 4, 2025. [Online]. Available: https://main--eds-evident-website--evident-scientific. aem.live/de/products/xrf-analyzers/vanta

    2025

  37. [38]

    Distributed behavior trees for heterogeneous robot teams,

    Heppner, G., Berg, N., Oberacker, D., Spielbauer, N., Roennau, A., and Dillmann, R., “Distributed behavior trees for heterogeneous robot teams,” in 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE) , Aug. 2023, pp. 1–

    2023

  38. [39]

    VDB-mapping: A high resolution and real-time capable 3d mapping framework for versatile mobile robots,

    Grosse Besselmann, M., Puck, L., Steffen, L., Roennau, A., and Dillmann, R., “VDB-mapping: A high resolution and real-time capable 3d mapping framework for versatile mobile robots,” in 2021 IEEE 17th International Conference on Automation Science and Engineering (CASE) , Aug. 2021, pp. 448–454. DOI:10.1109/ CASE49439.2021.9551430

    arXiv 2021

  39. [40]

    31, 2025

    Leggedrobotics/open3d_slam, Oct. 31, 2025. Accessed: Nov. 4,

    2025

  40. [41]

    DOI:10.1109/CASE56687.2023.10260300

    arXiv 2023

  41. [42]

    3d global path planning for walking robots on sparse volumetric maps,

    Grosse Besselmann, M. et al., “3d global path planning for walking robots on sparse volumetric maps,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , Oct. 2024, pp. 5430–5437. DOI:10.1109/IROS58592.2024.10802767 24

    arXiv 2024

  42. [43]

    An efficient locally reactive controller for safe navigation in visual teach and repeat missions,

    Mattamala, M., Chebrolu, N., and Fallon, M., “An efficient locally reactive controller for safe navigation in visual teach and repeat missions,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 2353–2360, Apr. 2022, ISSN: 2377-3766. DOI:10.1109/LRA. 2022.3143196

    arXiv 2022

  43. [44]

    19, 2025

    Ori-drs/field_local_planner, Dec. 19, 2025. Accessed: Jan. 19,

    2025

  44. [45]

    Real-time loop clo- sure in 2d LIDAR SLAM,

    Hess, W., Kohler, D., Rapp, H., and Andor, D., “Real-time loop clo- sure in 2d LIDAR SLAM,” in2016 IEEE International Conference on Robotics and Automation (ICRA) , May 2016, pp. 1271–1278. DOI: 10.1109/ICRA.2016.7487258

    arXiv 2016

  45. [46]

    MEM: Multi-modal elevation mapping for robotics and learning,

    Erni, G., Frey, J., Miki, T., Mattamala, M., and Hutter, M., “MEM: Multi-modal elevation mapping for robotics and learning,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , Oct. 2023, pp. 11011–11018. DOI:10.1109/ IROS55552.2023.10342108

    arXiv 2023

  46. [47]

    Leggedrobotics/elevation_mapping_cupy, Nov. 4, 2025. Accessed: Nov. 4, 2025. [Online]. Available: https : / / github . com / leggedrobotics/elevation_mapping_cupy

    2025

  47. [48]

    NavPi: An adaptive local path-planning pipeline for 3d navigation in difficult terrain,

    Grosse Besselmann, M., Rönnau, A., and Dillmann, R., “NavPi: An adaptive local path-planning pipeline for 3d navigation in difficult terrain,” in Synergetic Cooperation between Robots and Humans , E. S. E. Youssef, M. O. Tokhi, M. F. Silva, and L. M. Rincon, Eds., Cham: Springer Nature Switzerland, 2024, pp. 224–235, ISBN: 978- 3-031-47272-5. DOI:10.1007/9...

    work page doi:10.1007/978-3-031-47272-5_19 2024

  48. [49]

    Exploringtheperformance of ROS2,

    Maruyama,Y.,Kato,S.,andAzumi,T.,“Exploringtheperformance of ROS2,” in Proceedings of the 13th International Conference on Embedded Software , ser. EMSOFT ’16, New York, NY, USA: AssociationforComputingMachinery,Oct.1,2016,pp.1–10,ISBN: 978-1-4503-4485-2. DOI:10.1145/2968478.2968502 DAVID OBERACKERreceived the M.Sc. degree in informatics from the Karlsruhe...

    arXiv 2016

  49. [50]

    Elevation mapping for locomotion and navigation using GPU,

    Miki, T., Wellhausen, L., Grandia, R., Jenelten, F., Homberger, T., and Hutter, M., “Elevation mapping for locomotion and navigation using GPU,” in2022 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS) , Oct. 2022, pp. 2273–2280. DOI: 10.1109/IROS47612.2022.9981507

    arXiv 2022

  50. [2020]

    DOI:10.1109/LRA.2020.3007468

    arXiv 2020

  51. [2023]

    DOI:10.55417/fr.2023004

    work page doi:10.55417/fr.2023004

  52. [2025]

    Available: https://github.com/leggedrobotics/ open3d_slam

    [Online]. Available: https://github.com/leggedrobotics/ open3d_slam

  53. [2026]

    Available:https://github.com/ori-drs/field_local_ planner

    [Online]. Available:https://github.com/ori-drs/field_local_ planner