Pith. sign in

REVIEW 3 major objections 6 minor 66 references

Bilevel Learning for Bilevel Planning

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read IVNTR learns neural predicates from demonstration transitions and uses them as relational state abstractions in a bilevel planner, achieving 77% average success on unseen robot planning tasks where existing methods stay below 35%.

desk verdict IVNTR is a genuine step forward in predicate invention—replacing predefined classifiers with neural ones trained from effect-vector supervision—but its load-bearing sparsity assumption is acknowledged yet never empirically checked, so the 77% headline is a conditional result. read the letter →

arxiv 2502.08697 v3 pith:7JF4JDKY submitted 2025-02-12 cs.RO

classification cs.RO
keywords bilevelplanningpredicateinventionneuralpredicatesimitationlearningrelationalstateabstractioncompositionalgeneralizationtaskandmotionneuro-symbolic
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

IVNTR is a method for teaching a robot the high-level concepts—called predicates—that a bilevel planner needs, and for learning them directly from demonstration transitions instead of from hand-designed feature detectors. A bilevel planner alternates between a symbolic level, where operators and predicates abstract the state, and a continuous level, where samplers choose concrete actions. The paper's claim is that the missing dynamic predicates can be invented by alternating symbolic learning of predicate effects with neural training of predicate classifiers, each supplying supervision for the other. If the claim holds, a robot can learn relational concepts such as 'an object is reachable from this platform' as neural networks and then compose them zero-shot into plans for unseen objects, unseen states, and longer horizons. On six simulated domains, IVNTR reports 77% average success on held-out tasks compared with under 35% for prior methods, and the same learned predicates drive a real quadruped mobile manipulator in long-horizon tests.

What carries the argument

The load-bearing mechanism is the effect vector, a compact discrete summary of how a candidate predicate would appear in every operator of the domain. The ground effect vector derived from it marks each ground predicate whose arguments are a subset of the action's objects with the same +1/−1/0 entry, while all other ground predicates are assumed unchanged, and this creates a differentiable training signal: a Jensen–Shannon divergence term keeps untouched predicates' predictions constant, and binary cross-entropy pushes changed predicates to the indicated truth values. The validation loss, decomposed action by action, is folded back into node values of an effect-tree so that a UCT-style expansion prioritizes sparse effect vectors likely to eliminate contradictions. This closed loop—symbolic effects labeling neural classifiers, neural losses ranking symbolic candidates—is what removes the need for pre-defined predicate classifiers.

What would settle it

Run IVNTR on a domain that includes a side-effect action—say, a grasping action that also knocks over an object not named in its arguments—so a ground predicate about the displaced object flips while the ground effect vector marks it unchanged. In a chained demonstration, the same intermediate state will then be labeled True in one transition and False in the next, the classifier's validation loss should stay high, and success on held-out tasks should collapse.

Watch

Extended reading notes

Core claim

The central claim is that predicate invention for bilevel planning can be turned into a 'learn-then-select' process with neural classifiers, and that IVNTR is the first such process to work directly from demonstrations. For each typed predicate signature, IVNTR searches a tree of lifted effect vectors, where each vector assigns +1, −1, or 0 to every action to indicate whether the predicate is an add effect, a delete effect, or untouched by that action. Each candidate vector turns a demonstration transition into supervised learning data: ground predicates whose object tuples are contained in the action's objects are labeled by the vector, and all other ground predicates are required to stay unchanged, so a neural network can be trained on raw object-centric features—poses, point clouds, or images—without any state annotation. The trained classifier's per-action validation loss then updates the search, steering future candidates toward effect vectors that are consistent with the demonstrations. The surviving neural predicates are subselected to minimize a planning objective and integrated with learned operators and samplers, yielding a complete bilevel planner.

Load-bearing premise

The load-bearing premise is effect sparsity: a ground action may change the truth values only of predicates whose objects all appear among that action's arguments, so every other predicate is guaranteed to keep its value across the transition.

Editorial extensions

If this is right

  • If the central claim holds, a new robotics domain can go from raw demonstration trajectories plus a handful of known goal and static predicates to a working bilevel planner, without an engineer hand-defining the dynamic predicates.
  • Because the classifiers are relational and the planner is symbolic, the system should generalize zero-shot to more objects and longer plan horizons than any training task; the paper reports exactly this on test distributions with more platforms, more blocks, and more targets than training.
  • The invented predicates do double duty: they define operator preconditions and effects, and they also act as success indicators that filter bad low-level sampler proposals; the paper's ablation shows dropping this guidance degrades success by up to 98.4%.
  • The method widens predicate invention to high-dimensional states—SE(3) poses, point clouds, and RGB images—where grammar-based predicate pools fail, making bilevel planning applicable to richer sensory inputs.
  • Learned neural predicates can beat human-specified ground-truth predicates: in the Blocks and Climb-Measure ablations, IVNTR's discovered predicates yield a lower planning objective than oracle effect vectors.

Reading between the lines

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

  • The effect-vector labeling scheme is a general weak-supervision mechanism: any deterministic transition system with known action signatures can generate per-transition labels without annotating states, so the same alternating loop could be reused to invent object-centric concepts in non-robotic neuro-symbolic learning problems.
  • The method's practical ceiling is set by effect sparsity; a natural next test is to inject side-effect actions into an existing domain and measure how quickly the contradiction in labels degrades planning, which the paper does not run.
  • Neural predicates trade interpretability for expressiveness; a post-hoc layer that names or explains the learned classifiers would help practitioners trust the invented abstractions, but the paper does not address this.
  • Because the final predicate selection still optimizes a non-differentiable planning objective by hill-climbing, a future version could try to differentiate through the selection stage, but the paper's separation is what keeps neural training tractable.
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 IVNTR, a bilevel learning framework that invents neural predicates for bilevel planning directly from demonstrations. The method alternates between symbolic learning of lifted effect vectors and neural learning of predicate classifiers: effect vectors provide transition-level supervision through the losses Lzero and Lone (Eqs. 1-5), validation losses guide a UCT-style tree search over effect vectors, and a planning-objective score J selects a compact predicate subset. The approach is evaluated in six simulated robot planning domains with SE(2), SE(3), and point-cloud state spaces, and on a real Boston Dynamics Spot robot for two mobile manipulation domains; the reported average success on unseen tasks is 77% versus below 35% for the comparison methods.

Significance. If the stated results are valid and reproducible, IVNTR would be a significant advance: it is, to my knowledge, the first predicate-invention method for bilevel planning that learns neural classifiers from demonstrations without hand-engineered classifiers or program-synthesis candidates. The paper deserves credit for its explicit loss formulation, the alternation between symbolic effect learning and neural classifier learning, the breadth of experimental domains, the real-robot deployment, and the informative ablations (ground-truth effect vectors, sampling without predicates, search-efficiency comparisons, and RGB-image states). The central weakness is that the supervision pipeline is valid only under an effect-sparsity assumption that is stated in Appendix B and never empirically validated in the six evaluation domains, so the current evidence mainly supports the method on domains that happen to satisfy the assumption rather than the claimed generality.

major comments (3)
  1. [Section IV-A, Definition 4, Eq. (3), and Appendix B] The whole supervision pipeline is load-bearing on Assumption 1 (Effect Sparsity). Definition 4 assigns effect 0 to every ground predicate whose objects are not contained in the ground action's object set, and Eq. (3) (Lzero) trains the classifier to keep such atoms unchanged. Appendix B explicitly acknowledges that this assumption fails when operators and action signatures have different variable sets or when actions have side effects, and it cites domains from prior work where the assumption breaks. The six experimental domains are never checked against this condition. For example, MoveToReach(?r,?p) and Gaze(?r,?t) change the robot pose relative to every target, so a predicate P(?r,?t) can change even when t is not in the action's object set; in that case Lzero labels a changing atom as unchanged and the mislabeled signal propagates through classifier training, validation-loss ranking, and predicate selection. Please add a quantitative validation of Assumption 1 in each domain (e.g., the fraction of transition pairs and atoms for which some predicate outside the action's object set changes), and either restrict the generality claims to domains satisfying the assumption or extend the method to detect and correct such violations.
  2. [Section V, Tables I-IV] The headline generalization claims rest on averaged success rates from five random seeds (three seeds for the real-robot results in Table II) without standard deviations, confidence intervals, or per-seed values. For instance, Table I reports a single number for each T_test success rate, Table III reports planning-objective values with no variance, and Table IV reports success rates and drop percentages with no error bars. It is therefore impossible to assess whether the improvements over baselines are stable or dominated by a single seed. Please report per-seed results or error bars/confidence intervals for all tables, and specify the number of repeated trials for each real-robot task.
  3. [Reproducibility / code availability] The manuscript repeatedly refers to 'our source code' in Appendices C, E, F, and H, but the paper and project website do not provide a code or data release link. Because the method involves many implementation choices that are not fully specified in the text (predicate-variable correspondence annotations, pruning thresholds, sampler architectures, and search hyperparameters), the empirical results are not independently reproducible as written. Please provide a public code repository with demonstration-generation scripts and hyperparameter settings, or a sufficiently detailed appendix so that the experiments can be reproduced without recourse to hidden source code.
minor comments (6)
  1. [Eq. (6) and Appendix D] The threshold tau in Eq. (6) and the pruning threshold tau in Appendix D are never given numerical values; please report the values used in each domain and provide a sensitivity analysis or at least one ablation on these thresholds.
  2. [Section IV-B and Appendix D] The UCT exploration constant, the maximum predicate arity, the effect-vector iteration limit, and the pruning threshold are not specified; please state the exact values used in the experiments.
  3. [Table I] The column labeled 'drop percentage' is not defined in the caption or text; please clarify how it is computed, since it seems to measure the relative decrease from T_train to T_test but the formula is not stated.
  4. [Figure 4] The phrase 'Due to the unreasonable effect supervisions' should be 'unreasonable effect supervision' or rephrased; the surrounding explanation of the intermediate state being labeled both True and False would also be clearer if it explicitly pointed to the conflicting labels in the two transition pairs.
  5. [Table V] The abbreviations 'MAOff' and 'MAOn' are used in the table without being defined in the caption or nearby text; please spell them out as MoveAwayOff and MoveAwayOn.
  6. [Definition 4 and Appendix B] The notation is inconsistent between Definition 4, which uses Opsi_p subset O_Ci, and Appendix B, which uses O_psi subset O_C; please align the notation for the boundary case where the predicate's object set equals the action's object set.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the effect-vector supervision loop is self-supervised, but final predicate selection and held-out task evaluation provide external grounding.

full rationale

IVNTR's training loop alternates symbolic effect-vector proposals with neural classifier fitting, and the same loss is used for training and validation (Eqs. 5-6). This is a mutual-consistency / self-supervised procedure, not a case where a predicted quantity is defined as the fitted input. The loss is not vacuous: because the classifier sees only the state, contradictory labels for the same intermediate state appearing in different transitions (Figure 4) produce high validation error, so the loss carries real information from the demonstrations. The load-bearing Assumption 1 (Appendix B) is an unverified precondition: if actions have side effects on objects outside their argument lists, Definition 4's zero entries and Lzero mislabel transitions. That is a correctness risk and limits the claimed generality, but it is not circularity; the paper explicitly flags it in Section VII and Appendix B. The final predicate set is selected by the planning objective J comparing learned plan skeletons against demonstration skeletons, and the headline numbers are success rates on held-out test tasks, both external to the training labels. Self-citations to prior predicate-invention work [17] and sparsity-assumption work [19] are foundational references, not load-bearing uniqueness theorems or ansatz smuggling. No step in the derivation reduces by definition to its own input.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central method leans on a stack of domain assumptions: known goal and static predicates, known transition function, an oracle-generated demonstration set, and the Effect Sparsity Assumption that grounds all neural supervision. The first three are stated in Section II and in the limitations; the last is introduced in Appendix B and is load-bearing because every training label is derived from it.

free parameters (4)
  • Validation-loss threshold tau (Eq. 6)
    Predicates are accepted into the candidate pool only if their validation loss is below tau; the value of tau is not reported and directly controls pool size.
  • Pruning threshold tau in Appendix D
    Used to prune effect vectors whose non-zero entries already yield high loss; value not reported.
  • Maximum predicate arity and effect-vector iteration limit
    The paper enumerates typed variable compositions "with maximum input arity" and stops tree expansion at a max iteration, but no values are given.
  • UCT exploration constant in tree search
    The tree expansion uses UCT scores but the exploration constant is not specified.
assumptions (6)
  • standard math Neural networks can be trained by gradient descent on the differentiable loss L of Eq. (1)
    The method assumes standard deep learning optimization works for the predicate classifiers; no formal guarantee is provided.
  • domain assumption The transition function f is known and deterministic (Section II)
    The demonstration states are recovered by applying f to initial states, and transitions are used to create supervision.
  • domain assumption Goal predicates Psi_G and static predicates Psi_sta are known a priori (Section II, Section VII)
    The paper explicitly lists this as a limitation; only dynamic predicates are invented.
  • domain assumption Effect Sparsity Assumption 1 (Appendix B)
    For any action C, predicates whose objects are not a subset of the action's objects must be unchanged. This assumption is required for the ground effect vectors to be valid training labels.
  • domain assumption Each action has one unique operator and predicate variables correspond to action variables (Appendix B and C)
    If the option set differs from the operator set, sparsity breaks; the paper notes this in Appendix B.
  • domain assumption Demonstrations D are task-solution pairs generated by a manually designed oracle bilevel planner (Section V-A)
    The training signal depends on valid, relatively efficient plans; the paper does not test learning from human demonstrations.
invented entities (1)
  • Invented neural predicates (e.g., P11(?r), P21(?r,?t))
    purpose: Relational state abstractions used as preconditions, add/delete effects, and sampler success indicators in bilevel planning.
    The predicates are evaluated only within the paper's own held-out tasks; no external benchmark confirms the semantic content of individual learned predicates, and the paper notes they are hard to interpret.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bilevel Learning for Bilevel Planning." pith.science (2026). https://pith.science/paper/7JF4JDKY

@misc{pith2026250208697,
  author       = {Pith},
  title        = {Pith review of: Bilevel Learning for Bilevel Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7JF4JDKY}},
  note         = {Machine review of arXiv:2502.08697}
}
read the original abstract

A robot that learns from demonstrations should not just imitate what it sees -- it should understand the high-level concepts that are being demonstrated and generalize them to new tasks. Bilevel planning is a hierarchical model-based approach where predicates (relational state abstractions) can be leveraged to achieve compositional generalization. However, previous bilevel planning approaches depend on predicates that are either hand-engineered or restricted to very simple forms, limiting their scalability to sophisticated, high-dimensional state spaces. To address this limitation, we present IVNTR, the first bilevel planning approach capable of learning neural predicates directly from demonstrations. Our key innovation is a neuro-symbolic bilevel learning framework that mirrors the structure of bilevel planning. In IVNTR, symbolic learning of the predicate "effects" and neural learning of the predicate "functions" alternate, with each providing guidance for the other. We evaluate IVNTR in six diverse robot planning domains, demonstrating its effectiveness in abstracting various continuous and high-dimensional states. While most existing approaches struggle to generalize (with <35% success rate), our IVNTR achieves an average of 77% success rate on unseen tasks. Additionally, we showcase IVNTR on a mobile manipulator, where it learns to perform real-world mobile manipulation tasks and generalizes to unseen test scenarios that feature new objects, new states, and longer task horizons. Our findings underscore the promise of learning and planning with abstractions as a path towards high-level generalization.

Figures

Figures reproduced from arXiv: 2502.08697 by the authors.

Figure 1
Figure 1. (Top) Our bilevel learning framework invents [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The Climb-Transport domain is presented as a running example. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Overview of IVNTR during training. Given transition pairs in the continuous space, IVNTR invents neural predicates with different [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Detailed neural learning process for predicate [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Detailed symbolic learning process for predicate group [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the five domains (excluding Climb-Transport) we have studied in this work. These domains feature various state [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison between IVNTR with other search strategies in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Typical failure cases in the real robot tests of the Climb-Transport domain. The most common failure is that the target is dropped [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Visualization of the classification results on sampled poses [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Score optimization process during the predicate selection [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: An example demonstration with 3 blocks, where the states are represented in RGB images. In each step, we have used an object detection algorithm to obtain object centric images as the input states to our IVNTR algorithm. Method/Predicate P1 P2 P3 P4 P5 Avg. IVNTR (Our…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 54 canonical work pages

  1. [1]

    Mimicgen: A data generation system for scalable robot learning using human demonstrations

    Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596 , 2023

  2. [2]

    Dif- fusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Dif- fusion policy: Visuomotor policy learning via action diffusion. In Robotics: Science and Systems (RSS) , 2023

  3. [3]

    Learning fine-grained bimanual manipulation with low-cost hardware

    Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. In Robotics: Science and Systems (RSS), 2023

  4. [4]

    Equivariant diffusion policy

    Dian Wang, Stephen Hart, David Surovik, Tarik Keleste- mur, Haojie Huang, Haibo Zhao, Mark Yeatman, Jiuguang Wang, Robin Walters, and Robert Platt. Equivariant diffusion policy. arXiv preprint arXiv:2407.01812 , 2024

  5. [5]

    Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning

    Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. arXiv preprint arXiv:2407.01479 , 2024

  6. [6]

    What Planning Problems Can A Relational Neural Network Solve? In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), volume 36, 2024

    Jiayuan Mao, Tom ´as Lozano-P ´erez, Josh Tenenbaum, and Leslie Kaelbling. What Planning Problems Can A Relational Neural Network Solve? In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), volume 36, 2024

  7. [7]

    LogiCity: Advancing Neuro-Symbolic AI with Abstract Urban Simulation

    Bowen Li, Zhaoyu Li, Qiwei Du, Jinqi Luo, Wenshan Wang, Yaqi Xie, Simon Stepputtis, Chen Wang, Katia P Sycara, Pradeep Kumar Ravikumar, et al. LogiCity: Advancing Neuro-Symbolic AI with Abstract Urban Simulation. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , pages 69840– 69864, 2024

  8. [8]

    Towards a unified theory of state abstraction for mdps

    Lihong Li, Thomas J Walsh, and Michael L Littman. Towards a unified theory of state abstraction for mdps. In AI&M, 2006

Show all 66 references
  1. [9]

    State abstractions for lifelong reinforce- ment learning

    David Abel, Dilip Arumugam, Lucas Lehnert, and Michael Littman. State abstractions for lifelong reinforce- ment learning. In International Conference on Machine Learning. PMLR, 2018

  2. [10]

    From skills to symbols: Learning symbolic representations for abstract high-level planning

    George Konidaris, Leslie Pack Kaelbling, and Tom ´as Lozano-P´erez. From skills to symbols: Learning symbolic representations for abstract high-level planning. Journal of Artificial Intelligence Research , 16:215–289, 2018

  3. [11]

    Learning Grounded Action Abstrac- tions From Language

    Lionel Wong, Jiayuan Mao, Pratyusha Sharma, Zachary S Siegel, Jiahai Feng, Noa Korneev, Joshua B Tenenbaum, and Jacob Andreas. Learning Grounded Action Abstrac- tions From Language. In Proceedings of the International Conference on Learning Representations (ICLR) , 2024

  4. [12]

    Discovering State And Action Abstractions For Generalized Task And Motion Planning

    Aidan Curtis, Tom Silver, Joshua B Tenenbaum, Tom ´as Lozano-P´erez, and Leslie Kaelbling. Discovering State And Action Abstractions For Generalized Task And Motion Planning. In Proceedings of The AAAI Conference On Artificial Intelligence (AAAI) , number 5, pages 5377– 5384, 2022

  5. [13]

    Guiding Long-Horizon Task and Motion Planning with Vision Language Models, 2024

    Zhutian Yang, Caelan Garrett, Dieter Fox, Tom ´as Lozano- P´erez, and Leslie Pack Kaelbling. Guiding Long-Horizon Task and Motion Planning with Vision Language Models, 2024

  6. [14]

    From Reals to Logic and Back: Inventing Symbolic V ocabularies, Actions and Models for Planning from Raw Data

    Naman Shah, Jayesh Nagpal, Pulkit Verma, and Siddharth Srivastava. From Reals to Logic and Back: Inventing Symbolic V ocabularies, Actions and Models for Planning from Raw Data. arXiv preprint arXiv:2402.11871 , 2024

  7. [15]

    Learning Symbolic Operators for Task and Motion Planning

    Tom Silver, Rohan Chitnis, Joshua Tenenbaum, Leslie Pack Kaelbling, and Tom ´as Lozano-P ´erez. Learning Symbolic Operators for Task and Motion Planning. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 3182–3189, 2021

  8. [16]

    Tenenbaum, Tom´as Lozano-P´erez, and Leslie Pack Kaelbling

    Tom Silver, Ashay Athalye, Joshua B. Tenenbaum, Tom´as Lozano-P´erez, and Leslie Pack Kaelbling. Learning Neuro-Symbolic Skills for Bilevel Planning. In Pro- ceedings of the Conference on Robot Learning (CoRL) , 2022

  9. [17]

    Predicate Invention for Bilevel Planning

    Tom Silver, Rohan Chitnis, Nishanth Kumar, Willie McClinton, Tom´as Lozano-P ´erez, Leslie Kaelbling, and Joshua B Tenenbaum. Predicate Invention for Bilevel Planning. In Proceedings of The AAAI Conference on Artificial Intelligence (AAAI) , volume 37, pages 12120– 12129, 2023

  10. [18]

    GLIB: Efficient Exploration for Relational Model-Based Rein- forcement Learning via Goal-Literal Babbling

    Rohan Chitnis, Tom Silver, Joshua B Tenenbaum, Leslie Pack Kaelbling, and Tom ´as Lozano-P ´erez. GLIB: Efficient Exploration for Relational Model-Based Rein- forcement Learning via Goal-Literal Babbling. In Pro- ceedings of The AAAI Conference on Artificial Intelligence (AAAI...

  11. [19]

    Practice Makes Perfect: Planning To Learn Skill Parameter Policies

    Nishanth Kumar, Tom Silver, Willie McClinton, Linfeng Zhao, Stephen Proulx, Tom ´as Lozano-P ´erez, Leslie Pack Kaelbling, and Jennifer Barry. Practice Makes Perfect: Planning To Learn Skill Parameter Policies. In Proceed- ings of the Robotics: Science And Systems (RSS) , 2024

  12. [20]

    Learning Efficient Abstract Planning Models That Choose What to Predict

    Nishanth Kumar, Willie McClinton, Rohan Chitnis, Tom Silver, Tom´as Lozano-P ´erez, and Leslie Pack Kaelbling. Learning Efficient Abstract Planning Models That Choose What to Predict. In Proceedings of the Conference on Robot Learning (CoRL) , 2023

  13. [21]

    VisualPredicator: Learning Abstract World Models With Neuro-Symbolic Predicates For Robot Planning, 2024

    Yichao Liang, Nishanth Kumar, Hao Tang, Adrian Weller, Joshua B Tenenbaum, Tom Silver, Jo ˜ao F Henriques, and Kevin Ellis. VisualPredicator: Learning Abstract World Models With Neuro-Symbolic Predicates For Robot Planning, 2024

  14. [22]

    Active learning for teaching a robot grounded relational symbols

    Johannes Kulick, Marc Toussaint, Tobias Lang, and Manuel Lopes. Active learning for teaching a robot grounded relational symbols. In IJCAI, pages 1451–1457. Citeseer, 2013

  15. [23]

    From skills to symbols: Learning symbolic representations for abstract high-level planning

    George Konidaris, Leslie Pack Kaelbling, and Tom ´as Lozano-P´erez. From skills to symbols: Learning symbolic representations for abstract high-level planning. Journal of Artificial Intelligence Research (JAIR) , 2018

  16. [24]

    Embodied Active Learning of Relational State Abstractions for Bilevel Planning

    Amber Li and Tom Silver. Embodied Active Learning of Relational State Abstractions for Bilevel Planning. In Proceedings of the Conference on Lifelong Learning Agents (CoLLAs), pages 358–375, 2023

  17. [25]

    InterPreT: Interactive Predicate Learning from Language Feedback for Generalizable Task Planning

    Muzhi Han, Yifeng Zhu, Song-Chun Zhu, Ying Nian Wu, and Yuke Zhu. InterPreT: Interactive Predicate Learning from Language Feedback for Generalizable Task Planning. arXiv preprint arXiv:2405.19758 , 2024

  18. [26]

    Grounding Predi- cates through Actions

    Toki Migimatsu and Jeannette Bohg. Grounding Predi- cates through Actions. In Proceedings of the International Conference on Robotics and Automation (ICRA) , pages 3498–3504. IEEE, 2022

  19. [27]

    Classical Planning in Deep Latent Space: Bridging the Subsymbolic-Symbolic Boundary

    Masataro Asai and Alex Fukunaga. Classical Planning in Deep Latent Space: Bridging the Subsymbolic-Symbolic Boundary. In Proceedings of The AAAI Conference on Artificial Intelligence (AAAI) , volume 32, 2018

  20. [28]

    Unsupervised Grounding of Plannable First-Order Logic Representation From Images

    Masataro Asai. Unsupervised Grounding of Plannable First-Order Logic Representation From Images. In Proceedings of the International Conference on Automated Planning and Scheduling (ICAPS), volume 29, pages 583– 591, 2019

  21. [29]

    Learning Neural- Dymbolic Descriptive Planning Models via Cube-Space Priors: the V oyage Home (to STRIPS)

    Masataro Asai and Christian Muise. Learning Neural- Dymbolic Descriptive Planning Models via Cube-Space Priors: the V oyage Home (to STRIPS). In Proceedings of the International Joint Conferences on Artificial Intel- ligence (IJCAI), pages 2676–2682, 2021

  22. [30]

    Bisimulation Makes Analogies in Goal-conditioned Reinforcement Learning

    Philippe Hansen-Estruch, Amy Zhang, Ashvin Nair, Patrick Yin, and Sergey Levine. Bisimulation Makes Analogies in Goal-conditioned Reinforcement Learning. In Proceedings of the International Conference on Ma- chine Learning (ICML) , pages 8407–8426, 2022

  23. [31]

    Predicate Invention from Pixels via Pretrained Vision- Language Models

    Ashay Athalye, Nishanth Kumar, Tom Silver, Yichao Liang, Tom´as Lozano-P ´erez, and Leslie Pack Kaelbling. Predicate Invention from Pixels via Pretrained Vision- Language Models. arXiv preprint arXiv:2501.00296 , 2024

  24. [32]

    Tenenbaum, Tom ´as Lozano-P´erez, and Leslie Pack Kaelbling

    Rohan Chitnis, Tom Silver, Joshua B. Tenenbaum, Tom ´as Lozano-P´erez, and Leslie Pack Kaelbling. Learning Neuro-Symbolic Relational Transition Models for Bilevel Planning. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages...

  25. [33]

    Integrated Task and Motion Planning

    Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tom´as Lozano-P ´erez. Integrated Task and Motion Planning. Annual Review of Control, Robotics, and Autonomous Systems, 4(1):265–293, 2021

  26. [34]

    The Fast Downward Planning System

    Malte Helmert. The Fast Downward Planning System. Journal of Artificial Intelligence Research , 26:191–246, 2006

  27. [35]

    Divergence Measures based on the Shannon Entropy

    Jianhua Lin. Divergence Measures based on the Shannon Entropy. IEEE Transactions on Information theory , 37(1):145–151, 1991

  28. [36]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980 , 2014

  29. [37]

    Learning Representations by Back-propagating Errors

    David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning Representations by Back-propagating Errors. nature, 323(6088):533–536, 1986

  30. [38]

    Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search

    R´emi Coulom. Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search. In Proceedings of the International Conference on Computers and Games (ICCG), pages 72–83, 2006

  31. [39]

    Mastering the Game of Go without Human Knowledge

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the Game of Go without Human Knowledge. nature, 550(7676):354–359, 2017

  32. [40]

    Relational Inductive Biases, Deep Learning, and Graph Networks

    Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational Inductive Biases, Deep Learning, and Graph Networks. arXiv preprint arXiv:1806.01261, 2018

  33. [41]

    Attention is All You Need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is All You Need. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , volume 30, 2017

  34. [42]

    Language Segment-Anything

    Luca Medeiros. Language Segment-Anything. https: //github.com/luca-medeiros/lang-segment-anything, 2024

  35. [43]

    Sam 2: Segment Anything in Images and Videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024

  36. [44]

    Hi- erarchical Task and Motion Planning in the Now

    Leslie Pack Kaelbling and Tom ´as Lozano-P ´erez. Hi- erarchical Task and Motion Planning in the Now. In Proceedings of the International Conference on Robotics and Automation (ICRA) , pages 1470–1477. IEEE, 2011

  37. [45]

    Relay Policy Learning: Solv- ing Long-Horizon Tasks via Imitation and Reinforcement Learning

    Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. Relay Policy Learning: Solv- ing Long-Horizon Tasks via Imitation and Reinforcement Learning. In Proceedings of the Conference on Robot Learning (CoRL), pages 1025–1037, 2020

  38. [46]

    Augment- ing Reinforcement Learning with Behavior Primitives for Diverse Manipulation Tasks

    Soroush Nasiriany, Huihan Liu, and Yuke Zhu. Augment- ing Reinforcement Learning with Behavior Primitives for Diverse Manipulation Tasks. In Proceedings of the International Conference on Robotics and Automation (ICRA), pages 7477–7484, 2022

  39. [47]

    Neural Logic Machines

    Honghua Dong, Jiayuan Mao, Tian Lin, Chong Wang, Lihong Li, and Denny Zhou. Neural Logic Machines. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1–10, 2019

  40. [48]

    Directed-Info GAIL: Learning Hierar- chical Policies from Unsegmented Demonstrations using Directed Information

    Mohit Sharma, Arjun Sharma, Nicholas Rhinehart, and Kris M Kitani. Directed-Info GAIL: Learning Hierar- chical Policies from Unsegmented Demonstrations using Directed Information. In Proceedings of the International Conference on Learning Representations (ICLR) , 2018

  41. [49]

    Compile: Compositional Imitation Learning and Execution

    Thomas Kipf, Yujia Li, Hanjun Dai, Vinicius Zambaldi, Alvaro Sanchez-Gonzalez, Edward Grefenstette, Pushmeet Kohli, and Peter Battaglia. Compile: Compositional Imitation Learning and Execution. In Proceedings of the International Conference on Machine Learning (ICML) , pages 3...

  42. [50]

    PDSketch: Integrated Domain Programming, Learning, and Planning

    Jiayuan Mao, Tom ´as Lozano-P ´erez, Josh Tenenbaum, and Leslie Kaelbling. PDSketch: Integrated Domain Programming, Learning, and Planning. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 36972–36984, 2022

  43. [51]

    BLADE: Learning Compositional Behaviors from Demonstration and Language

    Weiyu Liu, Neil Nie, Ruohan Zhang, Jiayuan Mao, and Jiajun Wu. BLADE: Learning Compositional Behaviors from Demonstration and Language. In Proceedings of the Conference on Robot Learning (CoRL) , 2024

  44. [52]

    Keypoint Abstraction using Large Models for Object-Relative Imitation Learning

    Xiaolin Fang, Bo-Ruei Huang, Jiayuan Mao, Jasmine Shone, Joshua B Tenenbaum, Tom ´as Lozano-P ´erez, and Leslie Pack Kaelbling. Keypoint Abstraction using Large Models for Object-Relative Imitation Learning. arXiv preprint arXiv:2410.23254, 2024

  45. [53]

    Gener- alized Planning in PDDL Domains with Pretrained Large Language Models

    Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B Tenenbaum, Leslie Kaelbling, and Michael Katz. Gener- alized Planning in PDDL Domains with Pretrained Large Language Models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 38, pages 20256– 20264, 2024

  46. [54]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , 2022

  47. [55]

    V oxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models

    Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. V oxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. In Proceedings of the Conference on Robot Learning (CoRL), pages 540–562, 2023

  48. [56]

    Look Before You Leap: Unveiling the Power of GPT-4v in Robotic Vision-Language Planning

    Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look Before You Leap: Unveiling the Power of GPT-4v in Robotic Vision-Language Planning. arXiv preprint arXiv:2311.17842, 2023

  49. [57]

    Open-World Task and Motion Planning via Vision-Language Model Inferred Constraints

    Nishanth Kumar, Fabio Ramos, Dieter Fox, and Cae- lan Reed Garrett. Open-World Task and Motion Planning via Vision-Language Model Inferred Constraints. In CoRL Workshop on Language and Robot Learning: Language as an Interface , 2024

  50. [58]

    Pddlstream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning

    Caelan Reed Garrett, Tom ´as Lozano-P ´erez, and Leslie Pack Kaelbling. Pddlstream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning. In Proceedings of the International Confer- ence on Automated Planning and Scheduling (ICAPS) , volume 30, ...

  51. [59]

    Howe, Craig A

    Drew McDermott, Malik Ghallab, Adele E. Howe, Craig A. Knoblock, Ashwin Ram, Manuela M. Veloso, Daniel S. Weld, and David E. Wilkins. PDDL-the Planning Domain Definition Language. 1998

  52. [60]

    Anytime Motion Plan- ning using the RRT

    Sertac Karaman, Matthew R Walter, Alejandro Perez, Emilio Frazzoli, and Seth Teller. Anytime Motion Plan- ning using the RRT. In Proceedings of the International Conference on Robotics and Automation (ICRA) , pages 1478–1483, 2011

  53. [61]

    Skill-based curiosity for intrinsically motivated reinforcement learning

    Nicolas Bougie and Ryutaro Ichise. Skill-based curiosity for intrinsically motivated reinforcement learning. Ma- chine Learning, 109, 2020

  54. [62]

    H˚akan LS Younes and Michael L Littman. PPDDL1. 0: An extension to PDDL for expressing planning domains with probabilistic effects. Techn. Rep. CMU-CS-04-162, 2:99, 2004

  55. [63]

    Hybrid Declarative-Imperative Representations for Hybrid Discrete-Continuous Decision- Making

    Jiayuan Mao, Joshua B Tenenbaum, Tom´as Lozano-P´erez, and Leslie Pack Kaelbling. Hybrid Declarative-Imperative Representations for Hybrid Discrete-Continuous Decision- Making. In Proceedings of the International Workshop on the Algorithmic Foundations of Robotics (WAFR)

  56. [64]

    Accelerating 3D Deep Learning with Py- Torch3D

    Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with Py- Torch3D. arXiv:2007.08501, 2020

  57. [65]

    Pointnet: Deep Learning on Point Sets for 3D Classification and Segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep Learning on Point Sets for 3D Classification and Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652–660, 2017

  58. [66]

    grounded on

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778, 2016. APPENDIX A. Complete Notation Table We have presented the com...

Pith tools

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