Pith. sign in

REVIEW 4 major objections 7 minor 27 references

Spot-On: A Mixed Reality Interface for Multi-Robot Cooperation

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Spot-On claims that one operator can command two heterogeneous quadruped robots through a mixed-reality scene graph that updates as robots interact.

desk verdict Useful multi-robot MR system, but the headline effectiveness claim outruns the evidence: the user study ran with robots disconnected, so a referee should require task-success data. read the letter →

arxiv 2505.22539 v1 pith:MKSZJXIF submitted 2025-05-28 cs.HC cs.RO

Tim Engelbracht , Petar Lukovic , Tjark Behrens , Kai Lascheit , René Zurbrügg , Marc Pollefeys , Hermann Blum , Zuria Bauer This is my paper
classification cs.HCcs.RO
keywords mixedrealitymulti-robotcooperationscenegraphquadrupedrobotshuman-robotinteractionrobotmanipulationuserstudy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Spot-On is a mixed-reality control app that lets a single operator command two quadruped robots, one with a gripper arm and one with a basket, to cooperate on physical tasks in a shared room. The paper's central claim is that this interface gives an effective and intuitive way to run multi-robot collaboration in complex real scenes: the operator clicks objects in a 3D digital twin, and the robots carry out coordinated sequences such as opening a drawer, finding a requested object, and dropping it into the other robot's basket. Supporting evidence is a usability study with 16 participants, in which users rated most design and function aspects good or very good and completed a navigation task progressively faster. The framework matters because existing MR robot interfaces mostly handle a single robot or ignore relationships between objects, whereas Spot-On uses a live-updating scene graph as a shared, state-aware representation that both robots and the human keep in sync.

What carries the argument

The scene graph $G=(V,E)$ is the central object: each vertex $v_i=(\xi_i,c_i,P_i,\Phi_i,s_i)$ holds a pose, semantic class, point cloud, motion primitives, and optional on/off state, and edges encode spatial and functional relationships. The graph is built offline from an iPad LiDAR scan, segmented with Mask3D plus YOLOv8 for small elements like light switches, and serialized as JSON for lightweight transfer. During operation it acts as a shared brain: robots post object-state changes to per-robot servers, the MR device polls those changes at 10 Hz, and every robot and the human see the same updated world, which is what makes coordinated multi-robot tasks possible.

What would settle it

Run each collaborative task, Fetch and Drop, Search and Drop, and Operate and Check, multiple times in the same room and record completion rates and failure modes; if, for example, fewer than half of grasp-and-drop sequences complete or the scene graph fails to register a lamp-state change after a switch operation, the claim of an effective real-world framework would be contradicted.

Watch

Extended reading notes

Core claim

The paper claims that multi-robot collaboration can be driven through a dynamic scene graph displayed as an interactive 3D reconstruction, with robot actions triggered by clicking objects rather than by natural-language commands. Commands are decomposed into single-robot skills inherited from prior systems, such as drawer-axis estimation, light-switch affordance prediction, grasp computation, and GPT-4-based state checking. The scene graph is updated in real time as robots interact: when one robot operates a light switch, the other checks lamp states and the pair creates a new lamp-switch edge. The user study supports, in the authors' terms, that the system provides an effective and intuitive framework for MR-based multi-robot collaboration in complex, real-world scenarios.

Load-bearing premise

The framework's effectiveness rests on the inherited robotic skills—drawer-axis estimation, light-switch affordance prediction, grasp computation, and GPT-4 state checking—succeeding in the real scene, and the paper gives no quantitative success rates for those executions, only a demonstration video and interface-timing measurements.

Editorial extensions

If this is right

  • A single operator can command two heterogeneous robots with asymmetric capabilities through one compact interface, with gripping actions automatically delegated to the arm-equipped robot.
  • Collaborative multi-robot tasks such as Fetch and Drop, Search and Drop, and Operate and Check are triggered by a single button press in the scene view, hiding the underlying task decomposition.
  • The scene graph propagates state changes like drawer open/closed and lamp on/off to all agents, so robots become aware of each other's effects without extra per-robot coordination code.
  • Interaction by clicking reduces ambiguity compared with natural-language commands, and voice control is deliberately limited to app navigation to prevent accidental physical actions.
  • Object-state verification uses a GPT-4-based checker, allowing the system to confirm lamp and drawer states from images without hand-coded sensors.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the scene-graph-as-shared-brain design scales, there is no evident ceiling at two robots: the same server-and-query pattern could coordinate larger fleets of heterogeneous mobile manipulators, provided bandwidth and collision avoidance keep up.
  • The study's weakest point is that robots were static during user testing, so a natural next experiment is to measure real success rates of each collaborative task, for example running Fetch and Drop twenty times to see which inherited components fail most often.
  • The lamp-switch association learned through Operate and Check suggests the scene graph can grow affordances from interaction, which could evolve into fully unsupervised mapping of functional links in new buildings.
  • The clicking-over-language choice, together with confirmation prompts for voice actions, points to a general safety principle: voice and generative interfaces may drive high-level navigation, but low-level physical commands benefit from deliberate one-click confirmation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper presents Spot-On, a mixed-reality interface for commanding two Boston Dynamics Spot robots with asymmetric capabilities (one arm-equipped robot, one robot with a basket and camera). The system builds an object-centric scene graph from an iPad scan, segments objects with Mask3D and YOLOv8, and provides a HoloLens/Quest interface with clicking, hovering, voice navigation, and multi-robot task primitives (drawer opening, swing-door operation, light-switch operation, grasping, state checking). The authors define coordinated multi-robot tasks (Fetch & Drop, Search & Drop, Operate & Check) and evaluate the interface with a pre-study and a main user study of 16 participants, measuring UI-navigation times for three tasks and collecting subjective design grades. The central claim is that Spot-On provides an effective and intuitive framework for MR-based multi-robot collaboration in complex, real-world scenarios, supported by a video demonstration and the user study.

Significance. If the claims were fully supported, this would be a useful systems contribution: it combines a dynamic scene graph, mixed-reality interaction, and heterogeneous robot teaming in a way that goes beyond single-robot interfaces such as HoloSpot, and it explicitly addresses collaborative tasks such as fetching into a robot-carried basket and associating light switches with lamps. The paper's strengths are its concrete system architecture (Sec. 3.1-3.4), its integration of several prior perception and manipulation components, and the iterative user-driven design process reflected in the pre-study. However, the significance is currently limited by an evidentiary gap: the main user study ran with robots disconnected, so the 'effective multi-robot collaboration' part of the central claim rests only on a video and on unvalidated inherited primitives. The 'intuitive' component is partially supported by task-completion times, but no statistical tests are reported. As a result, the paper is more a feasibility-plus-usability report than a validated demonstration of effective multi-robot cooperation.

major comments (4)
  1. [Sec. 4.2.2 and Sec. 4.2.1] The main quantitative study measures only user-interface navigation times against a static scene: Sec. 4.2.2 states that 'robots were not connected to real robots during the user study.' Consequently, the abstract's central claim that Spot-On 'provides an effective and intuitive framework for MR-based multi-robot collaboration in complex, real-world scenarios' has no direct quantitative support for the 'effective' part. No success rates, trial counts, intervention counts, or failure logs are reported for the single-robot and multi-robot tasks in Sec. 3.4. Please either add a robot-execution evaluation (e.g., N repeated trials per task with success rates, operator interventions, and timing) or restrict the claim to interface usability plus a feasibility demonstration.
  2. [Sec. 4.2.1] The text states that completion time 'decreased significantly' for tasks 2 and 3 and draws conclusions about learning curves, experience, and tutorial impact, but no statistical tests or confidence intervals are reported. With n=16 and no paired comparisons, the observed differences could be due to chance; add appropriate tests (e.g., Wilcoxon signed-rank for paired task times, Mann-Whitney U for group comparisons) and report effect sizes.
  3. [Sec. 3.4] The multi-robot tasks are sequences of perception and manipulation primitives inherited from prior work: Spot-Compose drawer-axis estimation, SpotLight switch-affordance prediction, AnyGrasp grasping, and GPT-4-based state checking. This manuscript reports no success rates or failure analysis for any of these primitives in the demonstrated scene. Since a failure at any step would invalidate the 'effective' claim, please provide quantitative execution results for the component primitives as executed by Spot-On (or, if citing prior work, show that those results transfer to this setup).
  4. [Sec. 4.2.2, Table 3] There is an internal inconsistency: the text identifies 'robot visualization (Q3)' and 'app responsiveness (Q5)' as underperforming, but Table 3 lists Q3 as 'Did the app respond accurately to your selections?' and Q5 as 'How realistic were the robot visualizations?' Also, the abstract's 'good or very good rating in almost all cases' is not supported by the reported means of 3.69 (Q3) and 3.56 (Q5) if 'good/very good' corresponds to grades 4 or 5. Please correct the mapping and report the full distribution (e.g., counts per grade, standard deviations).
minor comments (7)
  1. [Throughout] There are several typos that should be corrected: 'sematic class' (Sec. 3.1), 'the the MR interface' (Sec. 3.1), 'mption primitive' (Sec. 3.4), 'inital' (Sec. 3.3), 'currnet' (Sec. 3.3), and 'V oice' in section headings.
  2. [Sec. 4.2.2] The phrase 'open box questions' should be 'open-ended questions'.
  3. [Sec. 4.2.1 and Fig. 7] The figure shows box plots but does not report group sizes for the expert vs. beginner and tutorial vs. no-tutorial splits; please state the number of participants in each subgroup.
  4. [Sec. 3.3 and references] The video demonstration link appears only as the text 'video'; provide a proper URL or supplementary-material identifier, and consider replacing the bare HoloLens URL in [18] with a standard citation.
  5. [Sec. 4.1] The phrase 'close to 90% of the times (88%)' is imprecise; report the exact percentage (88%) without the approximation.
  6. [Sec. 2] The related-work discussion of multi-robot MR interfaces is thin beyond HoloSpot; if no other external systems exist, state that explicitly, and if they do, add a brief comparison.
  7. [Sec. 3.2] The scene graph queries servers at 10 Hz, but no latency or update-correctness measurements are reported; a sentence on expected or measured end-to-end update latency would strengthen the 'shared brain' claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: usability data and a video demonstration support the claims independently; the robot-disconnected user study is an evidence gap, not a circular step.

full rationale

Spot-On is a systems and usability paper, not a derivation chain; no equation in the paper constructs an output from its own inputs. The abstract's claim of an 'effective and intuitive framework' rests on (i) navigation-time measurements and qualitative ratings from a 16-user study and (ii) a video demonstration of the multi-robot tasks. The explicitly stated limitation in Sec. 4.2.2—'robots were not connected to real robots during the user study' and 'robots being static in the scene'—is a genuine external-validity and evidence gap: the quantitative study cannot verify live multi-robot execution, and no success rates are reported for the Spot-Compose, SpotLight, AnyGrasp, or GPT-4 primitives. However, an evidence gap is not circularity: the paper never claims that the user study measured robot-task success, and the 'effective' claim is supported by a demonstration rather than by a fitted parameter or by the study itself. The self-citations to Spot-Compose [15] and SpotLight [5] are used as inherited software components, not as load-bearing justifications for the paper's specific usability claims; no quantity in the evaluation is defined in terms of another quantity it is said to predict, no uniqueness theorem is imported from the authors' prior work, and no known result is renamed. Thus no circular step meeting the quote-and-reduction standard is present; score 0.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the scene graph is a data structure borrowed from prior work. The listed free parameters are hand-chosen implementation thresholds that influence robustness but are not core scientific quantities. The axioms are the perception and skill components on which the multi-robot tasks depend, all inherited from prior work or adopted as standard assumptions.

free parameters (3)
  • obstacle clearance distance = 0.6 meters
    Hand-chosen threshold in body planning (Sec. 3.4); no sensitivity analysis is provided, and it affects whether the robot can safely approach objects.
  • candidate body position radius = 1 meter
    Hand-chosen radius around the object for generating candidate robot positions in body planning (Sec. 3.4); no sensitivity analysis is provided.
  • scene graph state query rate = 10 Hz
    Chosen polling rate for state changes on MR device (Sec. 3.2); affects freshness of interface updates but is not varied or evaluated.
assumptions (7)
  • domain assumption Mask3D provides accurate 3D semantic instance segmentation of the scanned scene
    Invoked in Sec. 3.1 to obtain scene graph nodes; the entire interaction depends on correct segmentation.
  • domain assumption Fine-tuned YOLOv8 reliably detects handles and light switches in the posed RGBD sequence
    Used in Sec. 3.1 to detect small functional elements and in Sec. 3.4 for task execution; failures would break drawer, door, and switch tasks.
  • domain assumption AnyGrasp computes executable grasps from the 3D segmentation masks
    Relied on in Sec. 3.4 for object grasping in Fetch & Drop and Search & Drop tasks.
  • domain assumption Spot-Compose drawer-axis estimation and SpotLight light-switch operation work in the cluttered test scene
    These prior frameworks are used without modification in Sec. 3.4; the paper provides no new success metrics for them in this setting.
  • domain assumption GPT-4 API correctly infers binary object states from captured images
    Used in Sec. 3.4 for drawer open/closed and lamp on/off state checks; incorrect inference would corrupt scene graph updates.
  • domain assumption The iPad lidar scan yields a point cloud accurate enough for the downstream tasks
    The offline scene graph in Sec. 3.1 is built from this scan; poor geometry would propagate to all robot interactions.
  • domain assumption The 16 participants in the main study are representative of the target user population
    The usability claims in Sec. 4.2 rest on this convenience sample, which is skewed toward ages 15-24 and includes no users over 34.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spot-On: A Mixed Reality Interface for Multi-Robot Cooperation." pith.science (2026). https://pith.science/paper/MKSZJXIF

@misc{pith2026250522539,
  author       = {Pith},
  title        = {Pith review of: Spot-On: A Mixed Reality Interface for Multi-Robot Cooperation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MKSZJXIF}},
  note         = {Machine review of arXiv:2505.22539}
}
read the original abstract

Recent progress in mixed reality (MR) and robotics is enabling increasingly sophisticated forms of human-robot collaboration. Building on these developments, we introduce a novel MR framework that allows multiple quadruped robots to operate in semantically diverse environments via a MR interface. Our system supports collaborative tasks involving drawers, swing doors, and higher-level infrastructure such as light switches. A comprehensive user study verifies both the design and usability of our app, with participants giving a "good" or "very good" rating in almost all cases. Overall, our approach provides an effective and intuitive framework for MR-based multi-robot collaboration in complex, real-world scenarios.

Figures

Figures reproduced from arXiv: 2505.22539 by the authors.

Figure 1
Figure 1. Spot-On. We introduce Spot-On, a Mixed-Reality ap￾plication that allows Multi-Robot Control and Collaboration This digital clone of the environment allows operators to command robots to interact with various elements - ranging from drawers and swing doors to movable or more func￾tional objects - with a high degree of autonomy. The se￾mantic states of and connections between objects are up￾dated within the scene grap… view at source ↗
Figure 2
Figure 2. Spot-On Overview. Our user friendly top-level mixed-reality interface allows for intuitive robotic control. Starting from the main menu, the user opens the scene view (I.). The user is presented with an interactable 3D reconstruction of the scene. By hovering objects in the scene and clicking them, objects can be selected (II.). Now, object specific actions are highlighted and can be selected at will (III.). Once th… view at source ↗
Figure 3
Figure 3. Networking. The networking infrastructure relies on one server per robot. The HoloLens posts commands for the robots to execute, while the robots post their states and changes they have made in the environment. We query these changes, so that the scene graph representation in the user interface stays up-to-date. 3.3. Mixed Reality App We introduce a mixed reality app to interact with both the environment and robots.… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: First, a home menu housing the app’s core func￾tionalities, and second, a 3D scene for interacting with dif￾ferent objects. Upon start-up, the app will display a splash screen displaying the Spot-On logo, followed by a widget asking the user whether they would like to …
Figure 6
Figure 6. Figure 6: Robots in action. Spot-On offers a wide range of robotic interactions: Opening drawers (top left) and swing doors (top right), search cabinets, grasping objects and fetching it to the other robot (bottom left), operating light switches and checking lamp states (bottom …
Figure 7
Figure 7. Figure 7: Time Graphs. The figure compares execution times for three tasks. Subfigures show 25th to 75th percentile boxes with median. Task Description Task 1 “Have the robots identify the connec￾tions between the lamps and the only light switch in the scene.” Task 2 “Have the r…
Figure 8
Figure 8. Figure 8: Color scheme comparison. Comparison of the two choices for main app color scheme [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Day/Night mode. Comparison between night and day mode of our app [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Close and Scene button. Comparison between two ways to transfer to scene view, close and scene button. ID Question Text Mean Grade Q1 How visually appealing was the interface? 4.31 Q2 How intuitive was the interaction with objects and buttons in the app? 4.25 Q3 Did t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    Lost & found: Updating dynamic 3d scene graphs from egocentric observations, 2024

    Tjark Behrens, Ren ´e Zurbr¨ugg, Marc Pollefeys, Zuria Bauer, and Hermann Blum. Lost & found: Updating dynamic 3d scene graphs from egocentric observations, 2024. 2

    work page 2024

  2. [2]

    Interactive Perception: Leveraging Action in Perception and Perception in Action

    Jeannette Bohg, Karol Hausman, Bharath Sankaran, Oliver Brock, Danica Kragic, Stefan Schaal, and Gaurav S. Sukhatme. Interactive perception: Leveraging action in per- ception and perception in action. CoRR, abs/1604.03670,

    work page Pith review arXiv

  3. [3]

    A 3d mixed reality interface for human-robot team- ing

    Jiaqi Chen, Boyang Sun, Marc Pollefeys, and Hermann Blum. A 3d mixed reality interface for human-robot team- ing. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11327–11333. IEEE, 2024. 1, 2

    work page 2024

  4. [4]

    SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes

    Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In CVPR, 2024. 2

    work page 2024

  5. [5]

    Spotlight: Robotic scene under- standing through interaction and affordance detection, 2024

    Tim Engelbracht, Ren ´e Zurbr¨ugg, Marc Pollefeys, Hermann Blum, and Zuria Bauer. Spotlight: Robotic scene under- standing through interaction and affordance detection, 2024. 2, 5

    work page 2024

  6. [6]

    Anygrasp: Robust and efficient grasp perception in spa- tial and temporal domains, 2023

    Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spa- tial and temporal domains, 2023. 2, 5

    work page 2023

  7. [7]

    HoloSpot: Intuitive Object Manipulation via Mixed Reality Drag-and-Drop

    Pablo Soler Garcia, Petar Lukovic, Lucie Reynaud, Andrea Sgobbi, Federica Bruni, Martin Brun, Marc Z ¨und, Riccardo Bollati, Marc Pollefeys, Hermann Blum, et al. Holospot: In- tuitive object manipulation via mixed reality drag-and-drop. arXiv preprint arXiv:2410.11110, 2024. 1, 2

    work page Pith review arXiv

  8. [8]

    Tenenbaum, Antonio Torralba, Florian Shkurti, and Liam Paull

    Qiao Gu, Alihusein Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Cor- ban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, Chuang Gan, Celso Miguel de Melo, Joshua B. Tenenbaum, Antonio Torralba, Florian Shkurti, and Liam Paull. Concept- graphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv, ...

    work page 2023

Show all 27 references
  1. [9]

    Sukhatme

    Karol Hausman, Scott Niekum, Sarah Osentoski, and Gau- rav S. Sukhatme. Active articulation model estimation through interactive perception. In 2015 IEEE International Conference on Robotics and Automation (ICRA) , pages 3305–3312, 2015. 2

    2015

  2. [10]

    Ditto in the house: Building articulation models of indoor scenes through interactive perception, 2023

    Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception, 2023. 2

    2023

  3. [11]

    Mrnab: Mixed reality-based robot navi- gation interface using optical-see-through mr-beacon, 2024

    Eduardo Iglesius, Masato Kobayashi, Yuki Uranishi, and Haruo Takemura. Mrnab: Mixed reality-based robot navi- gation interface using optical-see-through mr-beacon, 2024. 2

    2024

  4. [12]

    Ditto: Building digital twins of articulated objects from interaction,

    Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction,

  5. [13]

    Ultralytics YOLO, 2023

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics YOLO, 2023. 2, 3, 5

    2023

  6. [14]

    Sampling-based algo- rithms for optimal motion planning

    Sertac Karaman and Emilio Frazzoli. Sampling-based algo- rithms for optimal motion planning. The International Jour- nal of Robotics Research, 30(7):846–894, 2011. 6

    2011

  7. [15]

    Spot-compose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds, 2024

    Oliver Lemke, Zuria Bauer, Ren ´e Zurbr¨ugg, Marc Pollefeys, Francis Engelmann, and Hermann Blum. Spot-compose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds, 2024. 1, 2, 5

    2024

  8. [16]

    Demonstrating ok-robot: What really matters in integrating open-knowledge models for robotics

    Peiqi Liu, Yaswanth Orru, Jay Vakil, Chris Paxton, Nur Shafiullah, and Lerrel Pinto. Demonstrating ok-robot: What really matters in integrating open-knowledge models for robotics. In Robotics: Science and Systems XX . Robotics: Science and Systems Foundation, 2024. 2

    2024

  9. [17]

    Online interactive perception of articulated objects with multi-level recursive estimation based on task-specific priors

    Roberto Martin-Martin and Oliver Brock. Online interactive perception of articulated objects with multi-level recursive estimation based on task-specific priors. 2014. 2

    2014

  10. [18]

    Hololens

    Microsoft. Hololens. https://www.microsoft.com/ en-us/hololens. Accessed: 2025-01-04. 2

    2025

  11. [19]

    Simple open-vocabulary object detection with vi- sion transformers, 2022

    Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection with vi- si...

    2022

  12. [20]

    Structure from action: Learning interactions for artic- ulated object 3d structure discovery, 2023

    Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for artic- ulated object 3d structure discovery, 2023. 2

    2023

  13. [21]

    Kinematic structure estimation of arbitrary articulated rigid objects for event cameras

    Urbano Miguel Nunes and Yiannis Demiris. Kinematic structure estimation of arbitrary articulated rigid objects for event cameras. In 2022 International Conference on Robotics and Automation (ICRA) , page 508–514. IEEE Press, 2022. 2

    2022

  14. [22]

    Chatgpt api, 2025

    OpenAI. Chatgpt api, 2025. 5

    2025

  15. [23]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2

    2021

  16. [24]

    3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans, 2020

    Antoni Rosinol, Arjun Gupta, Marcus Abate, Jingnan Shi, and Luca Carlone. 3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans, 2020. 2

    2020

  17. [25]

    Mask3D: Mask Trans- former for 3D Semantic Instance Segmentation

    Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask Trans- former for 3D Semantic Instance Segmentation. 2023. 2, 3

    2023

  18. [26]

    Open- mask3d: Open-vocabulary 3d instance segmentation

    Ayc ¸a Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Open- mask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2

    2023 arXiv

  19. [27]

    Ar- ticulated object manipulation using online axis estimation with sam2-based tracking

    Xi Wang, Tianxing Chen, Qiaojun Yu, Tianling Xu, Zanxin Chen, Yiting Fu, Cewu Lu, Yao Mu, and Ping Luo. Ar- ticulated object manipulation using online axis estimation with sam2-based tracking. arXiv preprint arXiv:2409.16287,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.