pith. sign in

arxiv: 2606.26588 · v1 · pith:55LCD7OJnew · submitted 2026-06-25 · 💻 cs.RO

Inference-Time Robot Behavior Steering through Physically-Aware Reconfiguration of Task-Structure

Yiyuan Pan , Hanjiang Hu , Shangtao Li , Xusheng Luo , Changliu Liu This is my paper

Pith reviewed 2026-06-26 05:30 UTC · model grok-4.3

classification 💻 cs.RO
keywords robot behavior steeringinference-time adaptationneural automatonsynchronous productvisuomotor policypreference followingtask structureresidual controller
0
0 comments X

The pith

ReStruct steers a trained robot policy at inference time by synchronously composing a preference automaton into its neural state-machine skeleton while freezing the low-level controller.

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

The paper shows that a visuomotor policy can be decomposed into an explicit high-level automaton skeleton plus a residual continuous controller, allowing new user preferences to be folded in at test time through a synchronous product operation on the automata. This updates the action priors supplied to the frozen controller so that the resulting behavior satisfies the preferences while remaining physically feasible. A sympathetic reader would care because existing steering methods either demand retraining or produce plans that ignore physics, whereas this approach works across object-centric goals and temporal-logic constraints without touching the original policy weights.

Core claim

ReStruct adopts the automaton to represent the preference and incorporates it into the skeleton through a synchronous product, thereby reconfiguring the task structure. With the controller kept frozen, the action priors provided by the skeleton are updated accordingly to enable physically-aware control under a modified task structure. Experiments in simulation and the real world show that the method steers object-centric specifications as well as temporal-logic constraints and exceeds both prior steering techniques and VLA models by up to 25 percent in combined task success and preference adherence.

What carries the argument

The neural automaton that decomposes the visuomotor policy into a high-level state-machine skeleton capturing task structure and a low-level residual controller; the synchronous product of this skeleton with a preference automaton reconfigures the supplied action priors.

If this is right

  • ReStruct handles both object-centric specifications and temporal-logic constraints without retraining.
  • The method improves task success and preference-following over prior steering approaches and over VLA models by up to 25 percent.
  • Only the high-level skeleton is edited; the low-level controller remains frozen throughout.
  • The approach works in both simulated environments and real-world robot deployments.

Where Pith is reading between the lines

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

  • The same decomposition might allow runtime insertion of safety constraints expressed as automata without touching the policy weights.
  • If the skeleton can be extracted from any visuomotor network, the technique could apply to policies trained by imitation or reinforcement learning alike.
  • Natural-language preferences could be compiled into automata upstream, turning the method into a language-to-behavior interface that still respects physics.

Load-bearing premise

The learned automaton skeleton accurately captures the original task structure so that its synchronous product with any new preference automaton still yields physically feasible action priors for the unchanged residual controller.

What would settle it

A trial in which the reconfigured skeleton produces action priors that cause the robot to violate physical constraints or collide, even though the combined automaton satisfies the stated preference.

Figures

Figures reproduced from arXiv: 2606.26588 by Changliu Liu, Hanjiang Hu, Shangtao Li, Xusheng Luo, Yiyuan Pan.

Figure 1
Figure 1. Figure 1: Behavior Steering Mechanism for Different Models. (Preference: ϕ1 “Put the can into the shelf”; ϕ2 “Put the can to the higher place”) (a) The multi-task policy can place the can into the trash bin and the mug onto the shelf, respectively. (b) The end-to-end policy requires the injection of expert knowledge. (c) The neuro-symbolic policy either fails by generating a logically valid but physically infeasible… view at source ↗
Figure 2
Figure 2. Figure 2: ENAP diagram. (ii) Action-prior Bridge: Each edge (q, σ) → q ′ stores a fixed action priors a t base drawn from demonstrations that exercise this transition. (iii) Low-level controller: A residual policy network πψ outputs a cor￾rection ∆at = πψ(qt, ot, at base). The final action is aˆt = a t base + ∆at. The inference process of ENAP follows a deterministic predict–act– update cycle ( [PITH_FULL_IMAGE:fig… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of ReStruct on PegInsert task. (Left) Base policy P1; (Middle) ReStruct transforms preference into a DFA; its synchronous product with M1 yields the steered skeleton M2. (Right) Trajectories are repopulated, producing P2, which inherits P1’s frozen low-level. Blue / Black / Red lines are high-level skeleton / low-level controller / user-side flows. form τ = {(ot, at)} T t=0. We are given a base po… view at source ↗
Figure 4
Figure 4. Figure 4: VLM-based Preference Ground￾ing. The VLM refines the alphabet and skele￾ton M1, then generates preference DFA. compiles it into a DFA A using toolbox Spot [19], and leaves the skeleton unchanged (M′ 1 = M1). (ii) If ϕ requires finer semantic granularity, the VLM refines the relevant clusters (e.g., pick → pick red, pick blue) to obtain an expanded al￾phabet Σ ′ . We then generate A over Σ ′ , and refine M1… view at source ↗
Figure 5
Figure 5. Figure 5: Task Suite. Causal skill 0∼4: rotate block, close drawer, open drawer, push right, push left; Unconstrained skill 0∼4: turn on led, turn on bulb, rotate block, push left, move slider. Challenges: PassTeaPot demands a trajectory-level preference, CanSorting’s ϕ requires compositional generalization, e.g., a new combination of red can with the blue bowl. ConeCover tests spatial reasoning with task memory, si… view at source ↗
Figure 7
Figure 7. Figure 7: Steering a Sequence from a Skill Graph of Unconstrained-ϕ1. (Left) ReStruct has narrower deviation tubes. (Right) ϕ1 steers the fully-connected PMM graph into a sequential chain. 4.3 Decisiveness via Statefulness (Q3) [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Episode-length distribution over successful ϕ1-steered PegInsert rollouts. ReStruct inherits statefulness from its state-machine￾based steering, providing a level of decisiveness ab￾sent in other symbolic structures. First, the node structure encodes task history as an internal memory. Moreover, on multimodal tasks such as PegInsert, ReStruct can commit to a single preferred mode, thereby reducing runtime … view at source ↗
Figure 8
Figure 8. Figure 8: Diagram of Skeleton Variations for StackCube (Fixed). The base skeleton and the DFA of preference ϕ1 are synthesized into a steered state machine, with an unrefined cluster table. cluster table (refined or not) with the state-machine topology, showing how preference ϕ1 compiles into a DFA and product-composes with the base skeleton to yield the steered state machine. For readability, the nodes in the Skele… view at source ↗
Figure 9
Figure 9. Figure 9: Diagram of Skeleton Variations for PegInsert. The base skeleton and the DFA of preference ϕ1 are synthesized into a steered state machine, with a refined cluster table [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Diagram of Skeleton Variations for SeqPushT. The base skeleton and the DFA of preference ϕ1 are synthesized into a steered state machine, with an unrefined cluster table. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Diagram of Skeleton Variations for ConeCover. The base skeleton and the DFA of preference ϕ1 are synthesized into a steered state machine, with an unrefined cluster table [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Diagram of Skeleton Variations for CanSorting. The base skeleton and the DFA of preference ϕ1 are synthesized into a steered state machine, with an unrefined cluster table. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Diagram of Skeleton Variations for PassTeaPot. The base skeleton and the DFA of preference ϕ1 are synthesized into a steered state machine, with a refined cluster table [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
read the original abstract

A central challenge in deploying learned robot policies is inference-time behavior steering: redirecting a policy at test time to satisfy user preferences not anticipated during training, without retraining. Existing methods fail in two modes: end-to-end methods require fine-tuning or expert-level guidance, while neuro-symbolic methods rely on predefined symbols whose edits can result in logically reasonable but physically infeasible plans. To address this challenge, we propose ReStruct, which builds upon a neural automaton policy that decomposes a visuomotor policy into a high-level state-machine skeleton capturing task structure and a low-level continuous controller represented as a residual policy. Specifically, ReStruct adopts the automaton to represent the preference and incorporates it into the skeleton through a synchronous product, thereby reconfiguring the task structure. With the controller kept frozen, the action priors provided by the skeleton are updated accordingly to enable physically-aware control under a modified task structure. Extensive experiments from simulation and real-world show that ReStruct steers a wide range of preferences, from object-centric specifications to temporal-logic constraints, and after steering surpasses existing methods, exceeding VLA models in both task success and preference-following by up to 25%.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces ReStruct, a neuro-symbolic approach for inference-time steering of learned robot policies. It decomposes a visuomotor policy into a neural automaton representing a high-level state-machine skeleton and a low-level residual controller. User preferences (object-centric or temporal-logic) are encoded as automata and incorporated via synchronous product with the skeleton; the resulting reconfigured action priors are fed to the frozen controller to produce physically-aware behavior without retraining. Experiments in simulation and real-world settings claim that ReStruct handles diverse preferences and outperforms baselines including VLA models by up to 25% in task success and preference adherence.

Significance. If the physical realizability of the reconfigured priors holds, the approach would meaningfully advance inference-time adaptation by combining the editability of automata with the performance of end-to-end policies, reducing reliance on retraining or expert guidance. The reported real-world experiments and breadth of preference types (object-centric to temporal logic) would strengthen its practical relevance if the feasibility guarantee is substantiated.

major comments (2)
  1. [Method (neural automaton decomposition and synchronous product)] The central claim that the synchronous product yields physically feasible action priors for the frozen residual controller (without retraining) is load-bearing, yet the manuscript provides no explicit loss term, training objective, or safeguard that enforces correspondence between automaton states and the controller's training distribution. This leaves open the possibility that product-generated transitions fall outside the controller's support.
  2. [Experiments section] Experimental results claim up to 25% gains over VLA models in both task success and preference following, but the abstract and method description omit details on experimental design, number of trials, baselines, statistical tests, or how physical feasibility was verified post-reconfiguration (e.g., via constraint violation rates or failure mode analysis).
minor comments (2)
  1. [Method] Notation for the automaton states, transitions, and the synchronous product operation should be formalized with explicit definitions or equations to improve reproducibility.
  2. [Abstract] The abstract states 'extensive experiments from simulation and real-world' but does not reference specific figures or tables showing quantitative metrics; cross-references would aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [Method (neural automaton decomposition and synchronous product)] The central claim that the synchronous product yields physically feasible action priors for the frozen residual controller (without retraining) is load-bearing, yet the manuscript provides no explicit loss term, training objective, or safeguard that enforces correspondence between automaton states and the controller's training distribution. This leaves open the possibility that product-generated transitions fall outside the controller's support.

    Authors: The neural automaton is obtained by decomposing the original visuomotor policy, so its states are by construction aligned with regions where the residual controller was trained. The synchronous product modifies only the high-level state machine while the controller remains frozen and receives action priors from the reconfigured skeleton; physical feasibility is therefore inherited from the original policy rather than enforced by an additional loss. We acknowledge that the manuscript does not spell out an explicit safeguard term. We will add a clarifying paragraph in the method section describing the decomposition procedure and the resulting distribution alignment. revision: partial

  2. Referee: [Experiments section] Experimental results claim up to 25% gains over VLA models in both task success and preference following, but the abstract and method description omit details on experimental design, number of trials, baselines, statistical tests, or how physical feasibility was verified post-reconfiguration (e.g., via constraint violation rates or failure mode analysis).

    Authors: The full manuscript reports simulation and real-robot trials, comparisons against VLA and other baselines, and success/preference metrics, with physical feasibility assessed via task completion and observed failure modes. However, the abstract and method sections are indeed terse on trial counts, statistical tests, and explicit feasibility verification. We will revise both sections to include the number of trials per condition, the statistical tests performed, and a short failure-mode analysis that reports constraint-violation rates. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and description contain no equations, derivations, or load-bearing steps that reduce the claimed performance (e.g., 25% gains) to fitted parameters, self-definitions, or self-citation chains by construction. ReStruct is presented as an application of existing automata concepts (neural automaton decomposition into skeleton + residual controller, followed by synchronous product) to inference-time steering, with the controller kept frozen. No self-referential reductions or ansatzes smuggled via citation are evident in the text. This is the expected honest non-finding for a methods paper whose central claim rests on empirical validation rather than a closed mathematical derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no free parameters, axioms, or invented entities are extractable from the provided text.

pith-pipeline@v0.9.1-grok · 5748 in / 1117 out tokens · 22030 ms · 2026-06-26T05:30:49.594360+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

61 extracted references · 8 linked inside Pith

  1. [1]

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard. Calvin: A benchmark for language- conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    2022

  2. [2]

    H. Zhou, M. Ding, W. Peng, M. Tomizuka, L. Shao, and C. Gan. Generalizable long-horizon manipulations with large language models.arXiv preprint arXiv:2310.02264, 2023

    arXiv 2023

  3. [3]

    Y . Wang, L. Wang, Y . Du, B. Sundaralingam, X. Yang, Y .-W. Chao, C. P ´erez-D’Arpino, D. Fox, and J. Shah. Inference-time policy steering through human interactions. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 15626–15633. IEEE, 2025

    2025

  4. [4]

    Singhal, Z

    R. Singhal, Z. Horvitz, R. Teehan, M. Ren, Z. Yu, K. McKeown, and R. Ranganath. A gen- eral framework for inference-time scaling and steering of diffusion models.arXiv preprint arXiv:2501.06848, 2025

    arXiv 2025

  5. [5]

    Black, N

    K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, M. Y . Galliker, et al.π0.5: a vision-language-action model with open-world general- ization. In9th Annual Conference on Robot Learning, 2025

    2025

  6. [6]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    2025

  8. [8]

    B. P. Bhuyan, A. Ramdane-Cherif, R. Tomar, and T. Singh. Neuro-symbolic artificial intelli- gence: a survey.Neural Computing and Applications, 36(21):12809–12844, 2024

    2024

  9. [10]

    Nakamoto, O

    M. Nakamoto, O. Mees, A. Kumar, and S. Levine. Steering your generalists: Improving robotic foundation models via value guidance.arXiv preprint arXiv:2410.13816, 2024

    arXiv 2024

  10. [11]

    H ¨aon, K

    B. H ¨aon, K. Stocking, I. Chuang, and C. Tomlin. Mechanistic interpretability for steering vision-language-action models.arXiv preprint arXiv:2509.00328, 2025

    arXiv 2025

  11. [12]

    J. Mao, J. B. Tenenbaum, and J. Wu. Neuro-symbolic concepts.arXiv preprint arXiv:2505.06191, 2025

    arXiv 2025

  12. [13]

    Y . Pan, X. Luo, H. Hu, P. Yu, and C. Liu. Emergent neural automaton policies: Learning symbolic structure from visuomotor trajectories.arXiv preprint arXiv:2603.25903, 2026

    Pith/arXiv arXiv 2026

  13. [14]

    M. Y . Vardi. An automata-theoretic approach to linear temporal logic. InLogics for concur- rency: structure versus automata, pages 238–266. Springer, 2005

    2005

  14. [15]

    De Giacomo and M

    G. De Giacomo and M. Y . Vardi. Linear temporal logic and linear dynamic logic on finite traces. 2013

    2013

  15. [16]

    M. T. Spaan. Partially observable markov decision processes. InReinforcement learning: State-of-the-art, pages 387–414. Springer, 2012

    2012

  16. [17]

    Kurniawati

    H. Kurniawati. Partially observable markov decision processes and robotics.Annual review of control, robotics, and autonomous systems, 5(1):253–277, 2022. 18

    2022

  17. [18]

    S. Xu, X. Luo, Y . Huang, L. Leng, R. Liu, and C. Liu. Nl2hltl2plan: Scaling up natural lan- guage understanding for multi-robots through hierarchical temporal logic task specifications. IEEE Robotics and Automation Letters, 2025

    2025

  18. [19]

    Duret-Lutz, A

    A. Duret-Lutz, A. Lewkowicz, A. Fauchille, T. Michaud, E. Renault, and L. Xu. Spot 2.0—a framework for ltl and-automata manipulation. InInternational Symposium on Automated Tech- nology for Verification and Analysis, pages 122–129. Springer, 2016

    2016

  19. [20]

    R. Wang, J. Mao, J. Hsu, H. Zhao, J. Wu, and Y . Gao. Programmatically grounded, composi- tionally generalizable robotic manipulation.arXiv preprint arXiv:2304.13826, 2023

    arXiv 2023

  20. [21]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  21. [22]

    D. A. Reynolds et al. Gaussian mixture models.Encyclopedia of biometrics, 741(659-663):3, 2009

    2009

  22. [23]

    T. Mu, Z. Ling, F. Xiang, D. Yang, X. Li, S. Tao, Z. Huang, Z. Jia, and H. Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations.arXiv preprint arXiv:2107.14483, 2021

    arXiv 2021

  23. [24]

    O. Mees, L. Hermann, and W. Burgard. What matters in language conditioned robotic imitation learning over unstructured data.IEEE Robotics and Automation Letters, 7(4):11205–11212, 2022

    2022

  24. [25]

    Zhang, Y

    E. Zhang, Y . Lu, W. Wang, and A. Zhang. Language control diffusion: Efficiently scaling through space, time, and tasks.arXiv preprint arXiv:2210.15629, 2022

    arXiv 2022

  25. [26]

    Reuss, ¨O

    M. Reuss, ¨O. E. Ya˘gmurlu, F. Wenzel, and R. Lioutikov. Multimodal diffusion transformer: Learning versatile behavior from multimodal goals.arXiv preprint arXiv:2407.05996, 2024

    arXiv 2024

  26. [27]

    Reuss, H

    M. Reuss, H. Zhou, M. R ¨uhle, ¨O. E. Ya˘gmurlu, F. Otto, and R. Lioutikov. Flower: Democratiz- ing generalist robot policies with efficient vision-language-action flow policies.arXiv preprint arXiv:2509.04996, 2025

    arXiv 2025

  27. [28]

    P. F. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences.Advances in neural information processing systems, 30, 2017

    2017

  28. [29]

    Y . J. Ma, W. Liang, G. Wang, D.-A. Huang, O. Bastani, D. Jayaraman, Y . Zhu, J. Fan, et al. Eureka: Human-level reward design via coding large language models. InInternational con- ference on learning Representations, volume 2024, pages 26516–26560, 2024

    2024

  29. [30]

    Y . Pan, Z. Liu, and H. Wang. Wonder wins ways: Curiosity-driven exploration through multi- agent contextual calibration.Advances in Neural Information Processing Systems, 38:116109– 116137, 2026

    2026

  30. [31]

    Y . Chen, D. K. Jha, M. Tomizuka, and D. Romeres. Fdpp: Fine-tune diffusion policy with hu- man preference. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 12010–12016. IEEE, 2025

    2025

  31. [32]

    Goyal, R

    P. Goyal, R. J. Mooney, and S. Niekum. Zero-shot task adaptation using natural language. arXiv preprint arXiv:2106.02972, 2021

    arXiv 2021

  32. [33]

    Stone, T

    A. Stone, T. Xiao, Y . Lu, K. Gopalakrishnan, K.-H. Lee, Q. Vuong, P. Wohlhart, S. Kirmani, B. Zitkovich, F. Xia, et al. Open-world object manipulation using pre-trained vision-language models.arXiv preprint arXiv:2303.00905, 2023. 19

    arXiv 2023

  33. [34]

    S. Holk, D. Marta, and I. Leite. Predilect: Preferences delineated with zero-shot language- based reasoning in reinforcement learning. InProceedings of the 2024 ACM/IEEE Interna- tional Conference on Human-Robot Interaction, pages 259–268, 2024

    2024

  34. [35]

    Thumm, C

    J. Thumm, C. Agia, M. Pavone, and M. Althoff. Text2interaction: Establishing safe and prefer- able human-robot interaction.arXiv preprint arXiv:2408.06105, 2024

    arXiv 2024

  35. [36]

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots.arXiv preprint arXiv:2402.10329, 2024

    Pith/arXiv arXiv 2024

  36. [37]

    Bharadhwaj, R

    H. Bharadhwaj, R. Mottaghi, A. Gupta, and S. Tulsiani. Track2act: Predicting point tracks from internet videos enables generalizable robot manipulation. InEuropean Conference on Computer Vision, pages 306–324. Springer, 2024

    2024

  37. [38]

    V . Liu, A. Adeniji, H. Zhan, S. Haldar, R. Bhirangi, P. Abbeel, and L. Pinto. Egozero: Robot learning from smart glasses.arXiv preprint arXiv:2505.20290, 2025

    arXiv 2025

  38. [39]

    Black, M

    K. Black, M. Nakamoto, P. Atreya, H. Walke, C. Finn, A. Kumar, and S. Levine. Zero- shot robotic manipulation with pre-trained image-editing diffusion models. InInternational Conference on Learning Representations, volume 2024, pages 33431–33452, 2024

    2024

  39. [40]

    Sundaresan, Q

    P. Sundaresan, Q. Vuong, J. Gu, P. Xu, T. Xiao, S. Kirmani, T. Yu, M. Stark, A. Jain, K. Haus- man, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. In8th Annual Conference on Robot Learning, 2024

    2024

  40. [41]

    R. T. Icarte, T. Klassen, R. Valenzano, and S. McIlraith. Using reward machines for high-level task specification and decomposition in reinforcement learning. InInternational Conference on Machine Learning, pages 2107–2116. PMLR, 2018

    2018

  41. [42]

    Y . Chen, J. Arkin, C. Dawson, Y . Zhang, N. Roy, and C. Fan. Autotamp: Autoregressive task and motion planning with llms as translators and checkers. In2024 IEEE International conference on robotics and automation (ICRA), pages 6695–6702. IEEE, 2024

    2024

  42. [43]

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakr- ishnan, K. Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022

    Pith/arXiv arXiv 2022

  43. [44]

    Liang, W

    J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng. Code as policies: Language model programs for embodied control. In2023 IEEE International conference on robotics and automation (ICRA), pages 9493–9500. IEEE, 2023

    2023

  44. [45]

    C. R. Garrett, T. Lozano-P´erez, and L. P. Kaelbling. Pddlstream: Integrating symbolic planners and blackbox samplers via optimistic adaptive planning. InProceedings of the international conference on automated planning and scheduling, volume 30, pages 440–448, 2020

    2020

  45. [46]

    Konidaris and A

    G. Konidaris and A. Barto. Skill discovery in continuous reinforcement learning domains using skill chaining.Advances in neural information processing systems, 22, 2009

    2009

  46. [47]

    Pertsch, Y

    K. Pertsch, Y . Lee, and J. Lim. Accelerating reinforcement learning with learned skill priors. InConference on robot learning, pages 188–204. PMLR, 2021

    2021

  47. [48]

    Da Silva, G

    B. Da Silva, G. Konidaris, and A. Barto. Learning parameterized skills.arXiv preprint arXiv:1206.6398, 2012

    Pith/arXiv arXiv 2012

  48. [49]

    Konidaris, L

    G. Konidaris, L. P. Kaelbling, and T. Lozano-Perez. From skills to symbols: Learning symbolic representations for abstract high-level planning.Journal of Artificial Intelligence Research, 61: 215–289, 2018. 20

    2018

  49. [50]

    Silver, A

    T. Silver, A. Athalye, J. B. Tenenbaum, T. Lozano-P´erez, and L. P. Kaelbling. Learning neuro- symbolic skills for bilevel planning.arXiv preprint arXiv:2206.10680, 2022

    arXiv 2022

  50. [51]

    J. Mao, C. Gan, P. Kohli, J. B. Tenenbaum, and J. Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision.arXiv preprint arXiv:1904.12584, 2019

    Pith/arXiv arXiv 1904

  51. [52]

    H. M. Pasula, L. S. Zettlemoyer, and L. P. Kaelbling. Learning symbolic models of stochastic domains.Journal of Artificial Intelligence Research, 29:309–352, 2007

    2007

  52. [53]

    Chitnis, T

    R. Chitnis, T. Silver, J. B. Tenenbaum, T. Lozano-Perez, and L. P. Kaelbling. Learning neuro- symbolic relational transition models for bilevel planning. In2022 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4166–4173. IEEE, 2022

    2022

  53. [54]

    M. Poli, S. Massaroli, L. Scimeca, S. Chun, S. J. Oh, A. Yamashita, H. Asama, J. Park, and A. Garg. Neural hybrid automata: Learning dynamics with multiple modes and stochastic transitions.Advances in Neural Information Processing Systems, 34:9977–9989, 2021

    2021

  54. [55]

    Y . Pan, Y . Xu, Z. Liu, and H. Wang. Planning from imagination: Episodic simulation and episodic memory for vision-and-language navigation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 6345–6353, 2025

    2025

  55. [56]

    Y . Xu, Y . Pan, and Z. Liu. Dream to recall: Imagination-guided experience retrieval for memory-persistent vision-and-language navigation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  56. [57]

    Du and S

    M. Du and S. Song. Dynaguide: Steering diffusion polices with active dynamic guidance. Advances in Neural Information Processing Systems, 38:44192–44221, 2026

    2026

  57. [58]

    Y . Liu, J. I. Hamid, A. Xie, Y . Lee, M. Du, and C. Finn. Bidirectional decoding: Improving action chunking via guided test-time sampling.International Conference on Learning Repre- sentations (ICLR), 2025. URLhttps://arxiv.org/abs/2408.17355

    arXiv 2025

  58. [59]

    S. Wu, X. Luo, J. Zhang, J. Xie, J. Song, H. T. Shen, and L. Gao. Policy contrastive decoding for robotic foundation models.arXiv preprint arXiv:2505.13255, 2025

    Pith/arXiv arXiv 2025

  59. [60]

    Wagenmaker, M

    A. Wagenmaker, M. Nakamoto, Y . Zhang, S. Park, W. Yagoub, A. Nagabandi, A. Gupta, and S. Levine. Steering your diffusion policy with latent space reinforcement learning.arXiv preprint arXiv:2506.15799, 2025

    Pith/arXiv arXiv 2025

  60. [61]

    Y . Wu, R. Tian, G. Swamy, and A. Bajcsy. From foresight to forethought: Vlm-in-the-loop policy steering via latent alignment.arXiv preprint arXiv:2502.01828, 2025

    arXiv 2025

  61. [62]

    Z. Li, J. Liu, Z. Dong, T. Teng, Q. Rouxel, D. Caldwell, and F. Chen. Towards deploying vla without fine-tuning: Plug-and-play inference-time vla policy steering via embodied evolution- ary diffusion.IEEE Robotics and Automation Letters, 11(5):6234–6241, 2026. 21

    2026