Pith. sign in

REVIEW 4 major objections 5 minor 28 references

GORDON: Graph-based Object-centric Rewards for Decomposition of Long-Horizon Manipulation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a dense reward learned from unsegmented, action-free videos of object-centric scene graphs can both train policies and reveal stage boundaries in long-horizon manipulation tasks.

desk verdict A well-integrated object-centric reward method with a promising but under-validated subtask discovery mechanism. read the letter →

arxiv 2608.03753 v1 pith:37O4FG42 submitted 2026-08-04 cs.RO

Andrea Protopapa , Davide Buoso , Francesca Pistilli , Georgia Chalvatzaki , Giuseppe Averta This is my paper · ORCID
classification cs.RO
keywords robotmanipulationrewardlearninggraphneuralnetworksobject-centricrepresentationlong-horizonreinforcementautomaticsubtaskdiscoveryvideodemonstrationstemporalcycleconsistency
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

Long-horizon robot manipulation is hard for reinforcement learning because a single reward for the whole task gives too little guidance. This paper claims that a reward learned from unsegmented, action-free videos can supply that guidance and more: it can also show where the task breaks into stages. The method, GORDON, turns each video frame into a graph of detected objects and their spatial relations, trains a graph neural network to embed these graphs so temporally similar states are close and object structure is preserved, and defines reward as negative distance to the demonstrated goal embedding. An activity-aware pooling step upweights objects whose state changes and masks the robot, so the reward tracks object transitions rather than arm motion or visual appearance. In long-horizon tasks the paper reports that the reward's temporal profile rises in steps aligned with subtask endings, and that using those steps to segment demonstrations and train sequential subpolicies reaches 74.4% average full-task success, about 35 percentage points above the best learned baseline and 25 above a manually decomposed oracle reward.

What carries the argument

The central mechanism is the step-like temporal profile of the full-task reward, produced by a GNN encoder plus activity-aware weighted pooling (Eq. 2). The pooling rule upweights objects that have ever become active and masks robot nodes, so latent distances to the demonstrated goal respond mainly to object-configuration changes; the persistence of activity keeps completed stages represented. Automatic subtask discovery reads significant changes in the reward gradient to identify transition frames, then uses the resulting segments to train subtask-specific encoders, rewards, and policies composed by a sequential executor.

What would settle it

On a task with known subtask annotations, retrain the reward model with the keep-active-forever persistence term removed (or with activity allowed to decay), then compute the overlap between inferred reward-gradient boundaries and the annotated subtask endings. If overlap collapses, the stage structure is an artifact of the persistence rule; if it survives, the subtask-discovery claim stands. A complementary check: on a task where all object movement happens early but later stages are purely relational, see whether the reward still produces steps at the relational stage boundaries.

Watch

Extended reading notes

Core claim

The paper's central claim is that a self-supervised, object-centric reward learned from unsegmented videos captures both goal proximity and the ordered object-state transitions that define long-horizon progress. GORDON encodes scenes as graphs of detected objects and spatial relations, trains a GNN with temporal cycle-consistency and reconstruction, masks the robot via activity-aware pooling, and uses negative latent distance to the goal as reward. In long-horizon tasks the reward's temporal profile rises in steps aligned with subtask endings, and these boundaries let the method split demonstrations and compose subtask-specific policies. The reported result is 74.4% average long-horizon full

Load-bearing premise

The load-bearing premise is that the sharp rises in the learned reward mark genuine semantic subtask boundaries; since the activity rule keeps every object that ever begins moving permanently upweighted, a step is created each time an object first moves, and the paper does not quantify the gradient threshold or show quantitative agreement with ground-truth stage annotations beyond a few curves.

Editorial extensions

If this is right

  • Action-free, unsegmented videos become a sufficient supervision source: no expert actions, no subtask labels, and no language or VLM decompositions are required.
  • The same reward signal serves both dense shaping for RL and a segmentation cue, so long-horizon policy learning can be decomposed into per-stage subpolicies automatically.
  • Robot masking plus activity weighting makes the reward robust to appearance shifts and moderate localization noise, as shown in the ablations.
  • Full-task rewards that suffice for early stages are not enough for later stages; the stage structure itself carries information needed for cumulative success.
  • The ManiSkill3 reimplementation of RLBench-style tasks preserves task semantics while cutting training steps roughly sixfold, making the long-horizon evaluation reproducible.

Reading between the lines

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

  • If the step boundaries are semantic rather than artifacts, the same reward profile could be used at deployment for online stage detection and subpolicy switching, whereas the paper currently detects boundaries offline from demonstrations.
  • Removing the 'once active, always active' persistence rule in Eq. (2) is a direct experiment that would tell whether the steps come from object-state semantics or from the rule itself; the paper ablates the loss objectives but not this persistence mechanism.
  • The method depends on detector-provided object boxes, which suggests a natural next test: replacing simulator annotations with open-vocabulary detection or segmentation models to see how much reward quality degrades under noisy, unlabeled, or occluded object sets.
  • The reward profile could serve as weak supervision for building a reusable skill library for imitation or planning, independent of RL, since segment boundaries define reusable manipulation phases.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes GORDON, a graph-based object-centric reward learning method that learns dense rewards from action-free video demonstrations. Each frame is converted into a graph of detected objects and spatial relations; a GNN is trained with temporal cycle-consistency and reconstruction losses; the reward is a negative latent-space distance to a demonstrated goal embedding, plus a classifier-gated terminal bonus. Activity-aware weighted pooling (Eq. 2) upweights non-robot objects that move, while masking robot nodes. For long-horizon tasks, the temporal profile of the learned reward is used to discover subtask boundaries, train subtask-specific reward models and policies, and compose them sequentially. Experiments cover one MAGICAL task and six ManiSkill3 reimplementations of RLBench tasks, reporting 74.4% average full-task success across four long-horizon tasks, +35.3 p.p. over RoboHorizon published numbers and +25.4 p.p. over a decomposed environmental-reward oracle. Ablations on PutShoesInBox analyze reward discrimination, training objectives, and the terminal bonus.

Significance. If validated, the paper would make a useful contribution: learning dense rewards from unsegmented videos while masking robot motion is a practical need, and automatic subtask discovery from the reward profile would extend prior work on video-only reward learning. The framework is clearly motivated, and the ablations in Tables III–V usefully isolate robot masking, the TCC+reconstruction objectives, and the learned terminal bonus. The ManiSkill3 reimplementation for faster, parallel RL training is also a practical plus, though validated only on one task. However, the central claims currently rest on two unverified pillars: the semantic validity of the discovered boundaries and the fairness of the headline comparison against RoboHorizon. Both are addressable, but they need direct evidence before the claims can be accepted.

major comments (4)
  1. [Sec. III-A, Eq. (2)-(3); Sec. III-C; Fig. 6] The step-like reward profile may be substantially manufactured by the persistence rule. Eq. (2) sets the pooling weight to 1 for inactive objects and to alpha=2 for active objects, and the text states that 'once an object becomes active, we keep it active for the rest of the episode.' With tau=0.005 (0.5% of the bounding-box diagonal), any small motion or incidental contact permanently upweights an object, injecting a discrete step into the pooled representation and hence into r_full each time an object first moves. The paper does not ablate this rule (e.g., alpha=1 or no persistence), does not report the reward-gradient threshold used for boundary detection, and gives only qualitative overlap in Fig. 6 for three tasks. Since the discovered boundaries define the segmentation used to train subtask-specific policies, the 'automatic subtask discovery' claim requires quantitative evidence th
  2. [Sec. IV-A, Table I; Sec. IV-B] The headline improvement of +35.3 p.p. over the 'strongest learned or language-guided baseline' is computed against RoboHorizon numbers taken from Ref. [21], which were obtained on the original RLBench simulator, whereas GORDON is evaluated on a ManiSkill3 reimplementation. The paper acknowledges this as a 'semantically matched benchmark comparison,' but only the short-horizon TakeUmbrellaOut task is used to validate implementation correspondence; none of the four long-horizon tasks is matched. This cross-simulator comparison is not a controlled baseline and should not be the primary quantitative claim. Either re-run RoboHorizon in the same ManiSkill3 environments or restrict the headline improvement claim to baselines evaluated in the same experimental setting.
  3. [Sec. IV-A, Tables I-II] Several long-horizon comparisons show very large standard deviations over only five seeds. Examples: PutShoesOutBox Ours 56.2 +/- 33.1 vs. Env. Rew. (decomp.) 41.2 +/- 39.9; PutItemInDrawer S1->S2 Ours 95.0 +/- 11.2 vs. 62.5 +/- 39.9. No significance tests, per-seed paired comparisons, or confidence intervals are reported, so the average +25.4 p.p. gain over the decomposed environmental reward may be dominated by one or two seeds. Please provide per-seed results and a paired test (or more seeds) for the main full-task success comparisons.
  4. [Sec. III-C] The subtask discovery procedure is underspecified. 'Significant changes in the reward gradient' is never defined algorithmically; there is no threshold, no smoothing or peak-picking procedure, and no pseudocode. Moreover, the paper never reports the number of discovered subtasks M per task/demonstration, nor whether M matches the 2-3 ground-truth stages of PutItemInDrawer, PutShoesInBox, PutShoesOutBox, and EmptyContainer. Because automatic decomposition is the paper's central novelty, this information is necessary to judge whether the method genuinely discovers the intended structure.
minor comments (5)
  1. [Table I] The table formatting for the XIRL and GraphIRL rows is broken (numbers run together with method names); please fix spacing and alignment.
  2. [Fig. 6] The inferred boundary locations are presented only as dashed lines. A numeric comparison (e.g., mean absolute distance to GT subtask ends, or a table of detected vs. annotated boundaries) would make the qualitative claim reproducible.
  3. [Sec. IV] Object detections are obtained from simulator annotations. The paper should state more explicitly that the method currently assumes access to reliable object detection and discuss how this would be obtained on a real robot, since this affects the practical claim of learning from 'video demonstrations.'
  4. [Sec. IV] Implementation details such as GNN hidden dimensions, number of layers, optimizer, learning rates, batch size, and SAC hyperparameters are not reported. A project page without code or a completeness checklist limits reproducibility.
  5. [Abstract / Sec. IV-A] The abstract says 'approximately +35 p.p.' while the body reports '+35.3 p.p.'; use a single consistent number.

Circularity Check

1 steps flagged · score 6.0 of 10

Subtask discovery reduces to activity-onset steps manufactured by the persistence pooling rule.

  1. self definitional [Section III-A (Eq. 2-3, weighted pooling) and Section III-C (subtask discovery)]
    "In practice, we smooth activity estimates over time to reduce sensitivity to detector jitter, and once an object becomes active, we keep it active for the rest of the episode to preserve information about objects involved in completed stages. This persistence reflects the intuition that objects involved in earlier stages often remain relevant for representing completed progress. ... Candidate transition frames are identified from significant changes in the reward gradient."

    The pooling weight in Eq. (2) jumps from 1 to α=2 at the first displacement of any non-robot object and is then held at 2 by the persistence rule. Because the reward (Eq. 7) is computed from the pooled embedding (Eq. 1), every activity-onset event produces a discrete step in the reward profile regardless of whether it corresponds to a genuine semantic stage boundary. The subtask-discovery procedure then thresholds 'significant changes in the reward gradient,' so the discovered boundaries are, by construction, the times at which objects first become active. This is an input design choice, not an emergent property of the learned representation. The alignment with ground-truth boundaries in Fig. 6 is therefore partly explained by the fact that subtask changes in these tasks coincide with firs

full rationale

The paper's short-horizon reward learning, graph representation, and RL experiments are internally consistent and not circular. However, the central long-horizon claim—that the learned reward profile 'reveals stage-wise object-state transitions, enabling automatic subtask discovery without manual segmentation'—is partially manufactured by construction. The persistence rule in Eq. (2)-(3) deliberately creates a permanent step in the pooling weight whenever an object first moves, which injects a step into the reward. The subtask discovery then detects reward-gradient changes, so the discovered boundaries essentially reduce to activity-onset times. This does not mean the method is useless, but it undermines the claim that the stage structure is discovered from unsegmented demonstrations rather than being a consequence of a hand-coded step-inducing mechanism. No self-citation chain is involved. Score 6 reflects partial circularity of a central prediction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities; the 'graph representation', 'activity-aware pooling', and 'subtask rewards' are modeling constructs rather than invented entities. The central claim rests on three types of unproven assumptions: availability of simulator-grade object annotations, the correlation between TCC-aligned latent distance and task progress, and the semantic validity of reward-gradient boundaries.

free parameters (6)
  • alpha (activity weight) = 2.0
    Weight for active non-robot objects in Eq. (2); set by hand, controls how much moved objects dominate the pooled embedding and thus shapes the reward profile.
  • activity threshold tau = 0.005
    Threshold in Eq. (3) for flagging an object as active; hand-set; governs which objects are upweighted, influencing the step structure of the reward.
  • temporal window k = 40
    Frame offset for displacement in Eq. (3); chosen by hand; affects activity estimates.
  • classifier threshold tau_c = selected per task on validation split
    Gate for the terminal bonus in Eq. (8); tuned per task, so it is a fitted value when assessing success detection.
  • loss weights (lambda_rec, lambda_box, lambda_giou, lambda_cls, beta) = 0.5, 0.9, 0.1, 1.0, 5
    Hand-selected balancing coefficients for reconstruction loss and terminal bonus (Eqs. 4, 5, 8).
  • positive goal window q = 3%
    Frames in the final q% of each demo are labeled positive for the goal classifier; set by hand.
assumptions (5)
  • domain assumption Ground-truth object detections (2D/3D boxes, class labels, robot labels) are available from simulator annotations
    Section IV states detections are obtained from simulator annotations and used to construct the reward representation. Real-world applicability depends on detector accuracy; the noise-robustness test perturbs boxes but still uses sim-derived boxes.
  • domain assumption Temporal cycle consistency organizes embeddings so that latent distance to the goal embedding correlates with semantic task progress
    The dense reward (Eq. 7) assumes negative L2 distance in TCC-aligned space is a progress measure; the paper validates this indirectly through RL success but does not prove it.
  • domain assumption Reward-gradient change points correspond to semantically meaningful subtask boundaries
    Section III.C: candidate transition frames are identified from significant changes in the reward gradient; the paper checks alignment qualitatively in Fig. 6 but does not define the gradient threshold or report quantitative boundary agreement.
  • domain assumption Averaging final-frame embeddings gives a valid goal representation
    Eq. (6) defines z_full_g as the mean of final-frame embeddings across demos; assumes this mean is a stable goal target, though goal states may be multi-modal.
  • domain assumption Subpolicies trained from perturbed initial states compose without error accumulation
    Section III.C says each subpolicy is trained from perturbed initial states around plausible configurations of the preceding phase, and the sequential executor switches based on task progress; this assumes feasible handover states.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GORDON: Graph-based Object-centric Rewards for Decomposition of Long-Horizon Manipulation." pith.science (2026). https://pith.science/paper/37O4FG42

@misc{pith2026260803753,
  author       = {Pith},
  title        = {Pith review of: GORDON: Graph-based Object-centric Rewards for Decomposition of Long-Horizon Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/37O4FG42}},
  note         = {Machine review of arXiv:2608.03753}
}
read the original abstract

Learning long-horizon manipulation skills with reinforcement learning remains challenging due to the complexity of reward design, the limited guidance of sparse rewards, and the high cost of manual subtask annotation. Visual demonstrations can provide supervision for reward learning, but rewards learned from raw pixels can be brittle and sensitive to visual variation, background appearance, and robot motion. In this work, we propose GORDON, a graph-based object-centric reward learning framework that learns dense rewards from action-free video demonstrations. Each visual scene is represented as a graph of detected objects and spatial relations, and a graph neural network is trained in a self-supervised manner to embed these graphs into a task-aligned latent space. To align the representation with semantic task progress, we introduce an activity-aware weighted pooling mechanism that emphasizes task-relevant objects while masking robot-dominated motion. The dense reward is then computed as distances in the learned latent space of the current state to demonstrated goal configurations, providing a measure of task progress. In long-horizon tasks, the temporal profile of this reward reveals stage-wise object-state transitions, enabling automatic subtask discovery without manual segmentation. The discovered segments are then used to train subtask-specific rewards and specialized policies that are composed sequentially. Experiments on seven manipulation tasks on MAGICAL and ManiSkill3 benchmarks show that our object-centric reward improves reinforcement learning in short-horizon settings and enables successful policy learning in complex long-horizon tasks through automatic decomposition, achieving an average success rate of 74.4% across the long-horizon tasks (on average approximately +35 p.p. vs. best learned baseline and approximately +25 p.p. vs. oracle).

Figures

Figures reproduced from arXiv: 2608.03753 by the authors.

Figure 1
Figure 1. GORDON, our object-centric reward learning framework. Raw-image [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. Demo videos are converted into object-centric graphs and encoded with a GNN trained using temporal cycle [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Automatic subtask discovery and sequential policy training for long [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Short-horizon tasks require relatively compact sequences of object [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Learned full-task reward profiles on three representative long-horizon tasks. Stars denote simulator-annotated subtask endings, while dashed lines denote [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [21]

    Robohorizon: An llm-assisted multi-view world model for long-horizon robotic manipulation,

    Z. Chen, J. Huo, Y . Chen, and Y . Gao, “Robohorizon: An llm-assisted multi-view world model for long-horizon robotic manipulation,”arXiv preprint arXiv:2501.06605, 2025. 9

    arXiv 2025

  2. [1]

    Policy invariance under reward transformations: Theory and application to reward shaping,

    A. Y . Ng, D. Harada, and S. Russell, “Policy invariance under reward transformations: Theory and application to reward shaping,” inInterna- tional Conference on Machine Learning (ICML), 1999

    work page 1999

  3. [2]

    Recent advances in robot learning from demonstration,

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,”Annual review of control, robotics, and autonomous systems, 2020

    2020

  4. [3]

    Learning by watching: A review of video-based learning approaches for robot manipulation,

    C. Eze and C. Crick, “Learning by watching: A review of video-based learning approaches for robot manipulation,”IEEE Access, 2025

    work page 2025

  5. [4]

    Xirl: Cross-embodiment inverse reinforcement learning,

    K. Zakka, A. Zeng, P. Florence, J. Tompson, J. Bohg, and D. Dwibedi, “Xirl: Cross-embodiment inverse reinforcement learning,” inConference on Robot Learning, 2022

    work page 2022

  6. [5]

    Hierarchical rein- forcement learning: A comprehensive survey,

    S. Pateria, B. Subagdja, A.-h. Tan, and C. Quek, “Hierarchical rein- forcement learning: A comprehensive survey,”ACM Computing Surveys (CSUR), 2021

    work page 2021

  7. [6]

    Graph inverse reinforcement learning from diverse videos,

    S. Kumar, J. Zamora, N. Hansen, R. Jangir, and X. Wang, “Graph inverse reinforcement learning from diverse videos,” inConference on Robot Learning, 2023

    work page 2023

  8. [7]

    Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai,

    S. Taoet al., “Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai,”Robotics: Science and Systems (RSS), 2025

    work page 2025

Show all 28 references
  1. [8]

    Time-contrastive networks: Self-supervised learning from video,

    P. Sermanetet al., “Time-contrastive networks: Self-supervised learning from video,” in2018 IEEE international conference on robotics and automation (ICRA), 2018

    2018

  2. [9]

    Temporal cycle-consistency learning,

    D. Dwibedi, Y . Aytar, J. Tompson, P. Sermanet, and A. Zisserman, “Temporal cycle-consistency learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019

    2019

  3. [10]

    Learning reward functions for robotic manipulation by observing humans,

    M. Alakuijala, G. Dulac-Arnold, J. Mairal, J. Ponce, and C. Schmid, “Learning reward functions for robotic manipulation by observing humans,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  4. [11]

    Vip: Towards universal visual reward and representation via value- implicit pre-training,

    Y . J. Ma, S. Sodhani, D. Jayaraman, O. Bastani, V . Kumar, and A. Zhang, “Vip: Towards universal visual reward and representation via value- implicit pre-training,”arXiv preprint arXiv:2210.00030, 2022

    2022 arXiv

  5. [12]

    Liv: Language-image representations and rewards for robotic control,

    Y . J. Ma, V . Kumar, A. Zhang, O. Bastani, and D. Jayaraman, “Liv: Language-image representations and rewards for robotic control,” in International Conference on Machine Learning, 2023

    2023

  6. [13]

    Shadow: Leveraging segmentation masks for cross-embodiment policy transfer,

    M. Lepert, R. Doshi, and J. Bohg, “Shadow: Leveraging segmentation masks for cross-embodiment policy transfer,” inProceedings of The 8th Conference on Robot Learning, 2025

    2025

  7. [14]

    Augmented reality for robots (arro): Pointing visuomotor policies towards visual robustness,

    R. Mirjalili, T. J ¨ulg, F. Walter, and W. Burgard, “Augmented reality for robots (arro): Pointing visuomotor policies towards visual robustness,” IEEE Robotics and Automation Letters, 2026

    2026

  8. [15]

    Relay pol- icy learning: Solving long-horizon tasks via imitation and reinforcement learning,

    A. Gupta, V . Kumar, C. Lynch, S. Levine, and K. Hausman, “Relay pol- icy learning: Solving long-horizon tasks via imitation and reinforcement learning,” inProceedings of the Conference on Robot Learning, 2020

    2020

  9. [16]

    Taco: Learning task decomposition via temporal alignment for control,

    K. Shiarlis, M. Wulfmeier, S. Salter, S. Whiteson, and I. Posner, “Taco: Learning task decomposition via temporal alignment for control,” in International Conference on Machine Learning, 2018

    2018

  10. [17]

    Sequential dexterity: Chain- ing dexterous policies for long-horizon manipulation,

    Y . Chen, C. Wang, L. Fei-Fei, and K. Liu, “Sequential dexterity: Chain- ing dexterous policies for long-horizon manipulation,” inProceedings of The 7th Conference on Robot Learning, 2023

    2023

  11. [18]

    Universal visual decomposer: Long-horizon manipu- lation made easy,

    Z. Zhanget al., “Universal visual decomposer: Long-horizon manipu- lation made easy,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024

    2024

  12. [19]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahnet al., “Do as i can, not as i say: Grounding language in robotic affordances,”arXiv preprint arXiv:2204.01691, 2022

    2022 arXiv

  13. [20]

    RoboGen: Towards unleashing infinite data for auto- mated robot learning via generative simulation,

    Y . Wanget al., “RoboGen: Towards unleashing infinite data for auto- mated robot learning via generative simulation,” inProceedings of the 41st International Conference on Machine Learning, 2024

    2024

  14. [22]

    Deco: Task decomposition and skill composition for zero-shot generalization in long-horizon 3d manipulation,

    Z. Chenet al., “Deco: Task decomposition and skill composition for zero-shot generalization in long-horizon 3d manipulation,”IEEE Robotics and Automation Letters, 2026

    2026

  15. [23]

    Subtask-aware visual reward learning from segmented demonstrations,

    C. Kimet al., “Subtask-aware visual reward learning from segmented demonstrations,” inICLR, 2025

    2025

  16. [24]

    Egtr: Extracting graph from transformer for scene graph generation,

    J. Im, J. Nam, N. Park, H. Lee, and S. Park, “Egtr: Extracting graph from transformer for scene graph generation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024

    2024

  17. [25]

    The magical benchmark for robust imitation,

    S. Toyer, R. Shah, A. Critch, and S. Russell, “The magical benchmark for robust imitation,”NeurIPS, 2020

    2020

  18. [26]

    Rlbench: The robot learning benchmark & learning environment,

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison, “Rlbench: The robot learning benchmark & learning environment,”IEEE Robotics and Automation Letters, 2020

    2020

  19. [27]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” inInternational conference on machine learning, 2018

    2018

  20. [28]

    Graph transformer networks,

    S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim, “Graph transformer networks,”Advances in neural information processing systems, 2019

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.