Pith. sign in

REVIEW 3 major objections 6 minor 18 references

SGN-CIRL: Scene Graph-based Navigation with Curriculum, Imitation, and Reinforcement Learning

T0 review · 3 major / 6 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 The paper claims that feeding a Soft Actor-Critic navigation policy a CLIP-pooled open-vocabulary 3D scene graph, alongside imitation and curriculum learning, substantially improves mapless robot navigation success, especially in…

desk verdict The scene graph leaks the target's coordinates, so the paper's central claim about predicting target locations is untested; the engineering is real but the experiments need redesign. read the letter →

arxiv 2506.04505 v1 pith:SN7RWZRE submitted 2025-06-04 cs.RO cs.LG

Nikita Oskolkov , Huzhenyu Zhang , Dmitry Makarov , Dmitry Yudin , Aleksandr Panov This is my paper
classification cs.ROcs.LG
keywords maplessnavigation3DscenegraphreinforcementlearningimitationcurriculumSoftActor-Criticopen-vocabularyIsaacSim
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

The paper argues that a mobile robot navigating without a map can find a target object more reliably if its reinforcement learning policy is given a compact, learned encoding of a 3D scene graph in addition to camera images and a text command. The proposed SGN-CIRL framework combines Soft Actor-Critic reinforcement learning with imitation learning from Dijkstra-planned expert paths and curriculum learning that gradually increases the required initial distance and angle errors. In Isaac Sim experiments with an Aloha wheeled robot searching for a bowl, the scene-graph version outperforms the camera-only baseline, with the largest gains in the hardest configurations: up to 20 percentage points in structured scenes and 61 points when obstacles are randomly placed and the graph is supplied as ground truth. The paper identifies the graph encoder as the key mechanism, pooling CLIP text embeddings of objects with cosine-similarity weights relative to the target, thereby compressing semantic and spatial object information into a fixed-size policy input.

What carries the argument

The central mechanism is the CLIP-pooled scene-graph encoder. Each detected object is represented by a 518-dimensional vector: a 512-dimensional CLIP embedding of its text label plus the six numbers of its 3D bounding box, namely extent and position. The encoder computes the cosine similarity between each object embedding and the CLIP embedding of the target command, uses those similarities as weights, and pools the object vectors into a single 512-dimensional summary. This fixed-size summary is concatenated with image embeddings, the text command, and current velocities before being fed to the Soft Actor-Critic policy, so the agent's action selection can exploit spatial relationships between objects rather than relying only on pixels.

What would settle it

Train the SGN-CIRL (target) variant while withholding the bowl's bounding-box position from the graph, feeding only the other objects' boxes and text, and compare success on the same curriculum; if success drops to the graph-free baseline, the reported gains came from the supplied goal coordinate rather than from scene-graph reasoning.

Watch

Extended reading notes

Core claim

The central claim is that a learnable representation of an open-vocabulary 3D scene graph materially improves reinforcement-learning-based mapless navigation in partially observable indoor environments, beyond what is achievable with RGB-D images and a text command alone. The graph stores for each object its text description plus the extent and position of its 3D bounding box; the text is encoded with CLIP into 512 dimensions, concatenated with the six box dimensions, and the resulting per-object 518-dimensional rows are compressed by weighted pooling, where the weights are CLIP cosine similarities between each object's embedding and the target object's embedding. This compressed 512-dimensional scene vector is concatenated with current image embeddings, the text command, and velocities, and a Soft Actor-Critic policy outputs linear and angular velocities. The authors report that adding the graph raises average inference success from 71.5% to 78.6% in a simple scene, and from 82.4% to 89.8% in a complex two-table scene for the SGN-CIRL (scene) variant. They further report that in a randomly placed-obstacle scenario the graph-free agent nearly fails, while the graph-based agent reaches 82% success when given ground-truth graphs.

Load-bearing premise

The evaluation assumes that supplying the policy with a scene graph that already contains the target bowl's 3D bounding-box coordinates is still a test of predicting where the target is, rather than simply navigating to a known location.

Editorial extensions

If this is right

  • If the central claim is correct, scene graph representations can serve as an additional observation channel for mapless reinforcement learning navigation, improving success in hard configurations without changing the underlying control algorithm.
  • The reported gains concentrate in difficult cases, such as large initial distance errors and randomly placed obstacles, suggesting that graph-based spatial priors help exactly where camera-only reinforcement learning plateaus.
  • Combining imitation learning with curriculum learning alongside the graph stabilizes training and yields higher final success than reinforcement learning alone, indicating that the three ingredients are complementary.
  • Because the graph encoder uses CLIP text embeddings and open-vocabulary object descriptions, the same representation can be re-targeted by changing the textual command rather than retraining the perception module.
  • The pipeline is lightweight in simulation and, if reproduced on a physical robot, would offer a mapless alternative to map-based navigation that uses semantic knowledge about object arrangements.

Reading between the lines

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

  • The reported comparison does not fully isolate the graph's spatial-prediction value: in the SGN-CIRL (target) variant the graph supplies the bowl's 3D bounding-box position, so part of the gain may reflect navigation to a known coordinate rather than reasoning about where the target is likely to be.
  • A natural transfer test, not run in the paper, is zero-shot re-targeting to an object never seen during training using only a text prompt; the open-vocabulary design suggests this should work, but all experiments use a bowl as the sole target.
  • The 61-point jump in the random-obstacle experiment comes with manually constructed ground-truth graphs, so that gain conflates graph accuracy with graph benefit; re-running the protocol with an online scene-graph constructor would separate perception error from policy improvement.
  • If the scene-graph vector primarily acts as a spatial prior on target location, its benefit should shrink when the camera already views the target; this could be tested by stratifying success rates according to whether the bowl was visible at episode start.
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

3 major / 6 minor

Summary. The paper proposes SGN-CIRL, a mapless navigation framework that combines Soft Actor-Critic (SAC) with imitation learning, curriculum learning, and an open-vocabulary 3D scene graph encoded by CLIP-based pooling. Experiments in Isaac Sim compare scene-graph-based policies against a no-graph baseline in a simple scene, a complex scene with two walls and tables, and a scene with randomly placed obstacles. The authors report success rates and conclude that the 3D scene graph significantly improves navigation success, especially in difficult cases, and enables the agent to predict the target location in a partially observable environment. The code is open-sourced.

Significance. If the central claim were supported, the paper would offer a useful contribution: integrating open-vocabulary 3D scene graphs with RL, imitation learning, and curriculum learning for mapless navigation is a timely idea, and the open-source release is a strength. However, the experimental design currently feeds the target's 3D bounding-box position directly into the policy, so the claimed target-location prediction is untested and the success-rate gains are confounded. The significance of the work is therefore conditional on a substantially revised evaluation that removes this leakage and addresses the poor performance of the automatic scene-graph pipeline.

major comments (3)
  1. [III-A, IV-D] The scene graph input stores each object's 3D bounding-box position, and the SGN-CIRL (target) variant explicitly uses the bowl's 'position and extent of its 3d bounding box' as policy input. Because the SGN-CIRL (scene) variant pools per-object features with CLIP-similarity weights and the bowl node is present in the graph, its coordinates also enter the pooled representation. The agent is therefore given the target's location directly, so the abstract's claim that the agent can 'predict the location of the target object' is not tested, and the improvements over the no-graph baseline may simply reflect navigation to a known coordinate. The authors should remove the target's bounding box from the graph input and, ideally, explicitly evaluate the quality of the predicted target location.
  2. [IV-E, Table III] In the randomly placed obstacles experiment, the automatic scene-graph pipeline achieves only 21% success, whereas manually provided 'ground truth' graphs achieve 82%. The abstract and conclusion claim that using a 3D scene graph significantly improves success, but the practical graph-construction pipeline does not deliver the headline result; the improvement is demonstrated only with oracle graphs that would not be available in deployment. This discrepancy needs to be addressed directly, and the automatic-graph result should be treated as the primary evaluation of the proposed system.
  3. [IV-C, IV-D, Tables II-III] All success rates are reported as single numbers with no indication of the number of evaluation episodes, no multiple seeds, no error bars, and no significance tests. Given the modest differences (e.g., 82.4% vs 89.8% in Section IV-D) and the loose success criterion of 1.1 m distance error, the claimed improvements are not established statistically. The authors should report means and standard deviations across seeds, perform a significance test, and also report success at stricter distance thresholds.
minor comments (6)
  1. [III-A, Reward Design] The reward values are typeset as '−2 10' and '−1 10'; these appear to be formatting errors and should be corrected to the intended numerical values.
  2. [IV-D] The sentence 'The difference in learning rate is shown in Fig. 9' should refer to the difficulty level or success rate, since Fig. 9 plots the achieved difficulty level over episodes, not the learning rate.
  3. [II-A] In the paragraph discussing reference [14], the sentence 'Results in NVIDIA Isaac Sim demonstrates that our trained agent significantly outperforms...' appears to describe the current authors' own work rather than the cited work; this should be rephrased to avoid misattribution.
  4. [Abstract] The GitHub URL is typeset as 'Aloha graph' with a space; the link should be corrected to the actual repository address.
  5. [Figures 8 and 10] The success-rate-vs-distance curves are presented without confidence bands or error bars; adding these would make the variability across episodes visible.
  6. [V, Abstract] The Discussion states that the scene graph 'slightly improves the results on average', which is weaker than the abstract's 'significantly increased'; the wording should be aligned between these sections.

Circularity Check

1 steps flagged · score 6.0 of 10

Target-location 'prediction' is supplied as input: the scene graph stores the target bowl's 3D bounding-box position, and the SGN-CIRL (target) variant feeds it directly to the policy.

  1. self definitional [Abstract; Section III-A; Section IV-D]
    "The 3D scene graph models spatial relationships between objects, enabling the agent to efficiently navigate in a partially observable environment and predict the location of the target object. ... For each object in the scene, we extract a text description, along with the size and position of its 3D bounding box, which serve as part of the input for the RL policy. ... In the second SGN-CIRL (target), we use the information of target - bowl from the 3D scene graph, which includes position and extent of its 3d bounding box and its text clip embedding."

    The 'prediction of the target location' is not derived from spatial relations; the target bowl's 3D bounding-box position is already an element of the scene graph tensor given to the policy (III-A), and the SGN-CIRL (target) variant explicitly consumes that position as input (IV-D). The scene variant pools per-object features with CLIP weights, so a bowl node's coordinates are also present. Consequently the reported success rates test navigation to a supplied coordinate, and the claimed partial-observability location prediction is untested by construction. The experiment cannot distinguish scene-graph reasoning from oracle-coordinate following.

full rationale

The paper's RL pipeline and curriculum/imitation components are self-contained and not circular; however, the central scientific claim that the 3D scene graph enables prediction of the target's location in a partially observable environment is undermined by the design: the graph's object nodes include the size and position of each 3D bounding box, and the target variant uses the bowl's bounding-box position directly as policy input. The abstract and conclusion frame this as predictive reasoning, but the experiments actually provide the answer as an input, so the claimed 'prediction' reduces to reading off a coordinate. The scene variant's CLIP-weighted pooling still contains the bowl node's coordinates when the target is in the graph, so it does not isolate spatial-relationship inference either. This is a concrete, quotable reduction by construction rather than a vague concern. Self-citations (e.g., BBQ, Ref. [11]) are used as a tool, not as load-bearing justification, so they do not add circularity. Because the central 'prediction' claim is untested but the method itself is not simply a restatement of its inputs, partial circularity is appropriate (score 6).

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method relies on hand-crafted reward weights, curriculum thresholds, and success criteria, plus the assumption that a scene graph containing the target's position is a valid partial-observability observation. No new physical entities are introduced.

free parameters (4)
  • Sparse reward weights = -0.2, -0.1, 2, -5
    Hand-chosen reward values for subtask completion, collision, and time-out; these shape the learned policy but are not derived from data.
  • Curriculum success threshold = 85% over last 30 episodes
    Hand-chosen threshold controls when angular error is increased and distance error is reset; affects training dynamics.
  • Imitation episode fraction = 30%
    Hand-chosen fraction of episodes in which the control module provides actions instead of the policy.
  • Success criteria = 13 degrees and 1.1 meters
    Hand-chosen thresholds for considering an episode successful; 1.1 meters is a loose goal radius for reaching a bowl.
assumptions (4)
  • domain assumption A pre-built static 3D scene graph with object positions, sizes, and text labels is available before navigation and includes the target object.
    Invoked in Section III-A; the graph is constructed from the simulated scene, and the target bowl's bounding box is part of the graph input.
  • domain assumption CLIP embeddings of object text descriptions capture sufficient semantic similarity to identify the target by cosine similarity.
    Invoked in Section III-A; the graph encoder weights objects by CLIP cosine similarity to the target phrase.
  • domain assumption Dijkstra paths over the admissible grid are valid expert demonstrations for imitation learning.
    Invoked in Section III-B; Pure Pursuit follows these paths, and the actions are recorded in the replay buffer as demonstrations.
  • ad hoc to paper The navigation task is well-defined by the sparse reward and the success thresholds.
    The 1.1 meter success radius and 13 degree angle tolerance are chosen by the authors and are loose enough to inflate success rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SGN-CIRL: Scene Graph-based Navigation with Curriculum, Imitation, and Reinforcement Learning." pith.science (2026). https://pith.science/paper/SN7RWZRE

@misc{pith2026250604505,
  author       = {Pith},
  title        = {Pith review of: SGN-CIRL: Scene Graph-based Navigation with Curriculum, Imitation, and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SN7RWZRE}},
  note         = {Machine review of arXiv:2506.04505}
}
read the original abstract

The 3D scene graph models spatial relationships between objects, enabling the agent to efficiently navigate in a partially observable environment and predict the location of the target object.This paper proposes an original framework named SGN-CIRL (3D Scene Graph-Based Reinforcement Learning Navigation) for mapless reinforcement learning-based robot navigation with learnable representation of open-vocabulary 3D scene graph. To accelerate and stabilize the training of reinforcement learning-based algorithms, the framework also employs imitation learning and curriculum learning. The first one enables the agent to learn from demonstrations, while the second one structures the training process by gradually increasing task complexity from simple to more advanced scenarios. Numerical experiments conducted in the Isaac Sim environment showed that using a 3D scene graph for reinforcement learning significantly increased the success rate in difficult navigation cases. The code is open-sourced and available at: https://github.com/Xisonik/Aloha\_graph.

Figures

Figures reproduced from arXiv: 2506.04505 by the authors.

Figure 1
Figure 1. Simplified diagram of the developed SGN-CIRL approach. We [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Scheme of the proposed SGN-CIRL approach for mapless image-based robot navigation in a simulation environment. Its main feature is the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An example of path construction on a grid of admissible positions [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Experiment Scene 1 (simple) and Scene 2 (complex) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Aloha robot moves to the target - bowl on the table. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Graph of the achieved difficulty level as a function of the number of [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Graph of success rate as a function of initial distance error during [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 11
Figure 11. Figure 11: environment with randomly placed obstacles [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages

  1. [1]

    Robot Navigation with Map-Based Deep Reinforcement Learning

    G. Chen, L. Pan, Y . Chen, P. Xu, Z. Wang, P. Wu, J. Ji, and X. Chen, “Robot navigation with map-based deep reinforcement learning. arxiv 2020,” arXiv preprint arXiv:2002.04349 , 2020

    work page Pith review arXiv 2020

  2. [2]

    Autonomous exploration and mapping for mobile robots via cumulative curriculum reinforcement learning,

    Z. Li, J. Xin, and N. Li, “Autonomous exploration and mapping for mobile robots via cumulative curriculum reinforcement learning,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 7495–7500

    work page 2023

  3. [3]

    Mobile robot navigation using deep reinforcement learning,

    M.-F. R. Lee and S. H. Yusuf, “Mobile robot navigation using deep reinforcement learning,” Processes, vol. 10, no. 12, p. 2748, 2022

    work page 2022

  4. [4]

    Rl-dovs: Reinforcement learning for autonomous robot navigation in dynamic environments,

    A. K. Mackay, L. Riazuelo, and L. Montano, “Rl-dovs: Reinforcement learning for autonomous robot navigation in dynamic environments,” Sensors, vol. 22, no. 10, p. 3847, 2022

    work page 2022

  5. [5]

    Scene graph contrastive learning for embodied navigation,

    K. P. Singh, J. Salvador, L. Weihs, and A. Kembhavi, “Scene graph contrastive learning for embodied navigation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 10 884–10 894

    work page 2023

  6. [6]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” in International Conference on Learning Representations , 2018

    work page 2018

  7. [7]

    Sg-nav: Online 3d scene graph prompting for llm-based zero-shot object navigation,

    H. Yin, X. Xu, Z. Wu, J. Zhou, and J. Lu, “Sg-nav: Online 3d scene graph prompting for llm-based zero-shot object navigation,” Advances in Neural Information Processing Systems , vol. 37, pp. 5285–5307, 2025

    work page 2025

  8. [8]

    Hierarchical open-vocabulary 3d scene graphs for language-grounded robot navigation,

    A. Werby, C. Huang, M. B ¨uchner, A. Valada, and W. Burgard, “Hierarchical open-vocabulary 3d scene graphs for language-grounded robot navigation,” in First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024 , 2024

    work page 2024

Show all 18 references
  1. [9]

    Dynamic open-vocabulary 3d scene graphs for long-term language- guided mobile manipulation,

    Z. Yan, S. Li, Z. Wang, L. Wu, H. Wang, J. Zhu, L. Chen, and J. Liu, “Dynamic open-vocabulary 3d scene graphs for long-term language- guided mobile manipulation,” arXiv preprint arXiv:2410.11989 , 2024

    2024 arXiv

  2. [10]

    Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning,

    Q. Gu, A. Kuwajerwala, S. Morin, K. M. Jatavallabhula, B. Sen, A. Agarwal, C. Rivera, W. Paul, K. Ellis, R. Chellappa et al. , “Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) ...

    2024

  3. [11]

    Beyond bare queries: Open- vocabulary object grounding with 3d scene graph,

    S. Linok, T. Zemskova, S. Ladanova, R. Titkov, D. Yudin, M. Monastyrny, and A. Valenkov, “Beyond bare queries: Open- vocabulary object grounding with 3d scene graph,” 2024. [Online]. Available: https://arxiv.org/abs/2406.07113

    2024 arXiv

  4. [12]

    Learning object relation graph and tentative policy for visual navigation,

    H. Du, X. Yu, and L. Zheng, “Learning object relation graph and tentative policy for visual navigation,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VII 16 . Springer, 2020, pp. 19–34

    2020

  5. [13]

    Visual semantic navigation using scene priors,

    W. Yang, X. Wang, A. Farhadi, A. Gupta, and R. Mottaghi, “Visual semantic navigation using scene priors,” arXiv preprint arXiv:1810.06543, 2018

    2018 arXiv

  6. [14]

    Using deep reinforcement learning with automatic curriculum learn- ing for mapless navigation in intralogistics,

    H. Xue, B. Hein, M. Bakr, G. Schildbach, B. Abel, and E. Rueckert, “Using deep reinforcement learning with automatic curriculum learn- ing for mapless navigation in intralogistics,” Applied Sciences, vol. 12, no. 6, p. 3153, 2022

    2022

  7. [15]

    Reinforced imitation: Sample efficient deep reinforcement learning for mapless navigation by leveraging prior demonstrations,

    M. Pfeiffer, S. Shukla, M. Turchetta, C. Cadena, A. Krause, R. Sieg- wart, and J. Nieto, “Reinforced imitation: Sample efficient deep reinforcement learning for mapless navigation by leveraging prior demonstrations,” IEEE Robotics and Automation Letters , vol. 3, no. 4, pp. 44...

    2018

  8. [16]

    Autonomous navi- gation of mobile robots in unknown environments using off-policy reinforcement learning with curriculum learning,

    Y . Yin, Z. Chen, G. Liu, J. Yin, and J. Guo, “Autonomous navi- gation of mobile robots in unknown environments using off-policy reinforcement learning with curriculum learning,” Expert Systems with Applications, vol. 247, p. 123202, 2024

    2024

  9. [17]

    Implementation of the pure pursuit path tracking algorithm,

    R. C. Coulter, “Implementation of the pure pursuit path tracking algorithm,” 1992

    1992

  10. [18]

    Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

    Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” in Conference on Robot Learning (CoRL) , 2024

    2024

Pith tools

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