Pith. sign in

REVIEW 3 major objections 5 minor 8 references

An observation-only pipeline — temporal segmentation plus grammar induction — can extract reusable skills and compositional hierarchies from unlabelled video, beating action-supervised baselines and accelerating downstream RL.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

An observation-only pipeline that combines optimal-transport skill segmentation (ASOT) with Sequitur grammar induction produces reusable skill hierarchies that accelerate reinforcement learning in Craftax and Minecraft.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The segmentation results are the real contribution; the hierarchy metrics are a self-consistency check, so the 'semantically meaningful' claims should be read with caution. the 3 major comments →

arxiv 2601.23156 v2 pith:RGP45SHS submitted 2026-01-30 cs.LG cs.FL

Unsupervised Hierarchical Skill Discovery

Damion Harvey , Geraud Nangue Tasse , Benjamin Rosman , Branden Ingram , Steven James This is my paper
classification cs.LG cs.FL
keywords hierarchical skill discoveryskill segmentationtemporal action segmentationgrammar inductionSequiturunsupervised learningreinforcement learningoptions
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that hierarchical behaviour structure can be extracted from raw observations alone — no action labels, rewards, or human annotation. The proposed method, HiSD, segments pixel-level trajectories into discrete skills by temporally coherent optimal-transport clustering, collapses contiguous same-skill runs into symbols, and then runs a linear-time grammar-induction algorithm over all episodes to compress recurring subroutines into a multi-level hierarchy. The authors report that HiSD's hierarchies are more consistent and reusable than those of the baselines, even though those baselines were given the true sub-task ordering, and that the recovered hierarchy — not just the flat skills — supplies the inductive bias that lets a reinforcement-learning agent solve a sparse-reward crafting task in roughly 30,000 steps where primitive-action policies fail. A careful reader would care because the method's inputs are just observations, so if the claim holds it points a path to learning reusable task structure from the large amounts of unlabelled video that already exist.

Core claim

HiSD claims a two-stage, fully unsupervised pipeline can recover skills and their composition from unlabelled trajectories. Temporally regularised optimal transport labels each frame as one of K visually coherent skills; same-skill runs become atomic symbols. The linear-time grammar inducer Sequitur compresses repeated symbol pairs into non-terminals, so the induced grammar's leaves are skills and its internal nodes are recurring subroutines. The paper claims this grammar tracks the intended decomposition in controlled tasks, beats action-supervised baselines on most segmentation metrics (Craftax, Minecraft), and that the hierarchy, not flat skills, gives a masked-PPO agent the inductive bia

What carries the argument

The central mechanism is the pairing of ASOT — unsupervised temporal action segmentation posed as unbalanced optimal transport with a fused Gromov-Wasserstein temporal-regularity term, which labels frames while preventing label flicker — with a modified Sequitur grammar induction. Sequitur greedily turns repeated adjacent-symbol pairs into non-terminal rules while maintaining two invariants (digram uniqueness and rule utility), and its derivation tree is the hierarchy. The modification that carries the argument is the boundary token: episodes are concatenated with a unique marker that is forbidden from entering any production rule, so subroutines never span episode boundaries, and a non-term

Load-bearing premise

The load-bearing premise is that Sequitur's compression of clean skill labels is the correct task decomposition: the paper's 'ground truth' hierarchy is its own modified Sequitur run on clean labels, so if that grammar does not encode the true sub-goal structure, the hierarchy results demonstrate self-consistency, not semantic correctness.

What would settle it

Check whether HiSD's internal grammar nodes align with independently defined subgoal points — the first frame after each crafted item appears, or human-annotated milestone frames — better than chance-level boundaries. If the non-terminals do not coincide with externally anchored subgoals, the 'semantically meaningful hierarchy' claim fails. A second check: re-run the downstream RL experiment with behaviour-cloning-from-observation instead of action-labelled cloning; if the hierarchy's advantage disappears, the observation-only claim holds only for discovery, not learning.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Unlabelled demonstration video becomes a usable input for structure discovery: skills and hierarchies can be mined from observations alone, skipping the expensive step of collecting action labels or rewards.
  • Hierarchy matters for learning, not just description: in both evaluation domains the hierarchical option agent outperformed the flat skills-only agent, implying compositional structure helps temporal credit assignment.
  • The discovered skills are object-centric and reusable: visually distinct approaches to the same resource are grouped into a single skill, which is what makes cross-episode reuse and compact grammars possible.
  • The pipeline is cheap to run: the paper reports training and inference on consumer-grade GPUs with 6–10 GB of VRAM, in contrast to the roughly 24 GB the baselines required, lowering the barrier for scaling structure discovery.

Where Pith is reading between the lines

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

  • Editorial inference: the paper's hierarchy evaluation is anchored to a ground truth it defines as its own modified Sequitur run on clean labels (§5.4; §D.3 reports 'perfect recovery' of that object). The hierarchy metrics therefore establish self-consistency under label noise; validating the 'semantically meaningful' claim requires an independent anchor, such as human- or reward-defined subgoal st
  • Editorial inference: the Table 2 results in the noisiest Minecraft setting — one unique tree per episode and a 283-node average tree against a 22.7-node truth — suggest the pipeline's validity is strongest where segmentation is clean. A testable extension is to make the grammar stage probabilistic or to feed it segmentation confidence, which might resist cascading symbol errors.
  • Editorial inference: the paper's own downstream experiment still assumes action labels for behaviour cloning (§7). A fuller test of the observation-only thesis is to close that loop with behaviour-cloning-from-observation, so that both discovery and control run without actions.
  • Editorial inference: the boundary-token modification is a transferable recipe — keep grammar induction from merging across sequence boundaries — that should generalize to other concatenated-sequence compression tasks, such as extracting recurring routines from logged telemetry or documents.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Hierarchical Skill Discovery (HiSD), a two-stage unsupervised pipeline: ASOT-based temporal action segmentation converts raw observation features into discrete skill labels, then a modified Sequitur grammar induction (with a boundary-token fix) compresses concatenated skill sequences into a hierarchical CFG. The method is evaluated on Craftax and Minecraft (VPT) datasets, comparing segmentation quality (Table 1) and tree-structure metrics (Table 2) against CompILE and OMPN, and showing a downstream RL proof-of-concept in which HiSD-derived hierarchies improve PPO sample efficiency. The manuscript claims that HiSD 'consistently produces more structured and semantically meaningful hierarchies' than baselines and that the discovered hierarchies accelerate and stabilise downstream RL.

Significance. If validated, HiSD would provide one of the first fully observation-only pipelines for multi-level skill discovery in high-dimensional pixel domains, with the practical advantage of not needing actions, rewards, or interaction during discovery. The paper deserves credit for decoupling structure discovery from policy execution, evaluating on the full MineRL action space, and including a downstream RL demonstration; the segmentation results are anchored to externally derived inventory-delta labels, which is meaningful external grounding. However, the central hierarchy-quality claims currently lack external semantic validation, and the quantitative results in the noisiest environment contradict the 'consistently' qualifier. The core method is plausible but the current claims outrun the evidence.

major comments (3)
  1. [§5.4, Table 2, App. D.3] The 'ground truth structure' for hierarchy metrics is defined as 'generated by running our modified Sequitur on clean ground-truth labels' (§5.4). Consequently, every hierarchy-quality claim—unique tree counts, depth, size, branching, and the 'perfectly recovers' statement in App. D.3 (Fig. 12)—is a comparison of the pipeline's output against the same deterministic grammar applied to the same (clean) labels. This is a self-consistency check of the ASOT stage under label noise, not evidence that the induced non-terminals correspond to true subgoals. The abstract's central claim of 'semantically meaningful hierarchies' therefore rests on an unvalidated premise: that Sequitur's compression of the clean skill sequence is the correct semantic task decomposition. No independent semantic target (human-annotated subgoal structure, planner-generated HTN, or external benchmark) is used. Please add
  2. [Table 2, §6.2.2] Table 2's Minecraft-All row shows HiSD producing 500 unique trees (vs 293 for the ground-truth), mean tree size 283.25 (vs 22.69) and max branching 28.73 (vs 2.25). This directly contradicts the abstract's claim that HiSD 'consistently produces more structured and semantically meaningful hierarchies' and the §6.1.2 claim that HiSD 'matches or closely approximates the ground truth.' The text in §6.2.2 concedes that noise 'prevent[s] our deterministic grammar from discovering a consistent underlying structure.' The Mapped row is also far from the ground truth (500 vs 151 unique trees, size 83.85 vs 21.36). Thus, the hierarchy-quality advantage over baselines is not consistent across the evaluated settings. A revision should either restrict the claim to the simpler/low-noise tasks or provide evidence that the inflated trees are still semantically meaningful despite their size.
  3. [§7, Appendix C] The downstream RL comparison is not fully controlled. In App. C, OMPN hierarchies are not given learned low-level skill policies: 'we do not learn a BC model or a PU model for the OMPN hierarchies, only for the segmented skills OMPN produces.' OMPN leaf nodes are sequences of primitive actions, whereas HiSD's options are instantiated with BC-trained intra-option policies and PU-learned initiation/termination gates. Figure 5 therefore compares HiSD options (representation + learned controllers) with OMPN hierarchies (representation only). The observed RL improvement may be due to the added low-level controllers rather than the hierarchy structure. Please add an ablation where all methods share the same option-grounding mechanism (e.g., BC on the relevant segment labels) or explicitly state that the RL result is a combined proof of concept, not a hierarchy-quality comparison.
minor comments (5)
  1. [§5.3] F1 and mIoU under the 'Full' (global Hungarian) matching scheme need precise definitions: how are segment-level F1 and mIoU computed after global alignment? Please specify the matching and scoring procedure.
  2. [§6, first paragraph] Qualitative hierarchy analysis uses the single best-performing run per framework (selected via mIoU), which is a form of cherry-picking; the visual 'perfect recovery' examples in App. D.3 may not be representative. Please report how often the best-run hierarchy is typical, or provide aggregate visualizations.
  3. [§6.2.2] Typo: 'while HiSD still finds no unique trees' should read '500 unique trees' (Table 2 shows HiSD Mapped: 500 unique trees).
  4. [Table 19] The reported n-frames values for Craftax (60, 135, 110, 205) exceed the average episode lengths in Table 4 (14.9–33.0 steps). Please clarify what n-frames controls and how these values are applied to trajectories shorter than n-frames.
  5. [Appendix C, §7] In the RL evaluation, why is CompILE omitted from Figures 5 and 6 despite being a baseline in Table 1? Please state the reason (e.g., memory or training issues) for completeness.

Circularity Check

1 steps flagged

Hierarchy-quality claim rests on a self-consistency check: the 'ground truth' hierarchy is defined as the paper's own modified Sequitur run on clean labels, so 'perfect recovery' shows only label robustness under the authors' grammar, never external semantic validity.

specific steps
  1. self definitional [§5.4 (Hierarchy Metrics), Table 2, Appendix D.3 (Fig. 12)]
    "To assess structural quality and reuse, we compute tree-level metrics averaged over the dataset and compare them relative to the ground truth structure (generated by running our modified Sequitur on clean ground-truth labels). [...] Here, HiSD perfectly recovers the ground truth hierarchy (Figure 12C vs. Figure 12B)."

    The evaluation target ('ground truth structure') is defined as the output of the same modified Sequitur that HiSD uses for prediction, applied to clean labels. Hence 'HiSD perfectly recovers the ground truth hierarchy' is by construction the statement modified-Sequitur(ASOT-labels) ≡ modified-Sequitur(clean-labels): a self-consistency check of the segmentation labels under the paper's own grammar. It does not establish that the grammar's non-terminals correspond to true subgoals, because no external semantic target (human-annotated decomposition, HTN, planner grammar) ever validates Sequitur. The abstract's headline claim — 'semantically meaningful hierarchies' — is therefore anchored only to the algorithm's agreement with itself, so the hierarchy metrics measure robustness to label noise

full rationale

The segmentation evaluation (Table 1) is genuinely externally anchored: ground-truth labels are 'algorithmic annotations derived from domain knowledge' (§5) via 'inventory deltas and interaction logs', and HiSD competes against CompILE/OMPN on those labels — no circularity there. ASOT and Sequitur are cited external works (Xu & Gould 2024; Nevill-Manning & Witten 1997), not self-citations, so no self-citation chain is load-bearing. The downstream RL experiments (Figs 5–6) against sparse-reward tasks are empirical and not circular. The circularity is concentrated in the hierarchy-quality evaluation (§5.4, Table 2, App. D.3): the 'ground truth structure' is generated by the paper's own modified Sequitur on clean labels, so the headline claims of 'semantically meaningful hierarchies' and 'perfectly recovers the ground truth hierarchy' reduce to a self-consistency check — the grammar's output on noisy labels versus the same grammar's output on clean labels. Whether Sequitur's compression of the clean skill sequence is the correct semantic decomposition is assumed, never independently validated (no human-annotated subgoal decomposition, planner-generated HTN, or external benchmark). Additionally, the paper's own Minecraft-All row of Table 2 (HiSD: 500 unique trees, mean size 283.25; truth: 293/22.69) and §6.2.2's concession that the 'deterministic grammar' cannot find 'a consistent underlying structure' undercut the abstract's 'consistently produces... hierarchies' claim; likewise the Stone Pickaxe Static row (Truth 1 vs HiSD 36 unique trees) contradicts 'closely approximates the ground truth in simpler tasks'. Because the central semantic claim rests on a self-referential target, while the segmentation and RL components retain independent content, the paper is partially circular (6), not fully so.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The ledger shows the pipeline's real cost: a user-supplied skill count K, 17 per-task fitted hyperparameters, and a self-generated hierarchy ground truth. No genuinely new physical or mathematical entity is introduced; the 'entities' are learned representations whose external validation is limited to the reported metrics and the downstream RL check.

free parameters (5)
  • K (maximum number of skills) = Set to ground-truth skill count per task: 5 (Stone Pickaxe Static/Random), 2 (WSWS Random), 5 (Mixed), 44/14 (Minecraft
    User-supplied prior required by ASOT (§4.1); main results use the ground-truth count; Appendix D.1 shows WSWS Random avg mIoU drops from 0.72 (K=2) to 0.40 (K=3), so performance is sensitive to this input.
  • ASOT hyperparameters (17 per task) = Table 19: e.g., alpha-train 0.59, lambda-frame-train 0.10, n-frames 110 (Stone Rand); n-frames 9100 (Minecraft Mapped)
    Each task's hyperparameters were chosen by an Optuna sweep (Appendix E) to maximize validation segmentation; reported numbers are best-of-sweep per task.
  • PU decision thresholds = Selected by grouped 5-fold cross-validation maximizing F1 (App. C.1.1, C.2.1)
    Initiation and termination classifiers' thresholds are fitted on the segmented observations; they gate downstream RL option masking.
  • Option execution horizon and frameskip (Minecraft) = 64-step horizon, frameskip 8 (App. C.2.3)
    Hand-chosen; replaces unreliable termination PU models (PU-end Micro F1 about 0.09 in Minecraft).
  • PCA components (Craftax) = 650 components (~99% variance)
    Hand-chosen dimensionality for the feature stage (§5.1); the pipeline operates on these features rather than raw pixels despite the 'pixel-based' framing.
axioms (6)
  • domain assumption ASOT (Xu & Gould 2024) reliably yields latent skills aligned with semantically meaningful behavior when applied to PCA/MineCLIP features.
    HiSD's Stage 1 is a direct application of ASOT; the mapping from visual-temporal coherence to semantic skills is inherited unvalidated (§4.1, §5.1-5.2).
  • ad hoc to paper Sequitur compression of a clean skill sequence equals the correct semantic task hierarchy.
    The 'ground truth hierarchy' in §5.4 is defined as the output of the paper's own modified Sequitur on clean labels; no external validation ties grammar structure to true sub-goals.
  • domain assumption Ground-truth skill labels derived from inventory deltas and interaction logs are the correct segmentation target.
    All segmentation metrics compare against these programmatic annotations (§5, §5.2).
  • domain assumption Pretrained VPT and MineCLIP models encode demonstrations adequately for skill discovery.
    The Minecraft results depend on VPT-generated episodes and MineCLIP feature quality (§5.2, App. A.2).
  • domain assumption A* planner trajectories in modified Craftax are representative expert demonstrations.
    Craftax results are built on 500 A*-generated trajectories in a heavily modified, fully observable, deterministic environment (§5.1, App. A.1).
  • domain assumption Elkan-Noto PU learning on the discovered segments yields valid option initiation sets.
    Downstream RL masks options by PU initiation models; termination PU models are acknowledged unreliable and replaced by fixed horizons (App. C.2.1).
invented entities (3)
  • Latent skill prototypes (K per dataset) no independent evidence
    purpose: Atomic symbols of the induced grammar; learned by ASOT optimal transport over observation features
    Internal representation; its semantic validity is evidenced only by the paper's own segmentation metrics and qualitative figures — no external certificate.
  • Grammar non-terminals (composite subroutines) no independent evidence
    purpose: Mid-level options in downstream RL; represent recurring skill sequences such as Collect Wood -> Make Workbench
    Discovered by the modified Sequitur; hierarchy metrics compare them against the paper's self-generated ground truth (§5.4), and the RL experiment is the only external validation of their utility.
  • Boundary token phi no independent evidence
    purpose: Separates episodes in the corpus so grammar rules never cross episode boundaries
    Implementation mechanism (§4.2), not a falsifiable entity; included for ledger completeness.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Hierarchical Skill Discovery." pith.science (2026). https://pith.science/paper/RGP45SHS

@misc{pith2026260123156,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Hierarchical Skill Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RGP45SHS}},
  note         = {Machine review of arXiv:2601.23156}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We consider the problem of unsupervised skill segmentation and hierarchical structure discovery in reinforcement learning. While recent approaches have sought to segment trajectories into reusable skills or options, most rely on action labels, rewards, or handcrafted annotations, limiting their applicability. We propose a method that segments unlabelled trajectories into skills and induces a hierarchical structure over them using a grammar-based approach. The resulting hierarchy captures both low-level behaviours and their composition into higher-level skills. We evaluate our approach in high-dimensional, pixel-based environments, including Craftax and the full, unmodified version of Minecraft. Using metrics for skill segmentation, reuse, and hierarchy quality, we find that our method consistently produces more structured and semantically meaningful hierarchies than existing baselines. Furthermore, as a proof of concept, we demonstrate that these discovered hierarchies accelerate and stabilise learning on downstream reinforcement learning tasks.

Figures

Figures reproduced from arXiv: 2601.23156 by Benjamin Rosman, Branden Ingram, Damion Harvey, Geraud Nangue Tasse, Steven James.

Figure 1
Figure 1. Figure 1: Example of HiSD applied to a Minecraft trajectory. Step 1: HiSD segments the observational trajectory into distinct skills, such as get wood, craft tools, and gather stone. Step 2: Using these segmented trajectories, HiSD applies a grammar-based com￾pression algorithm to induce a hierarchy over the discovered skills, revealing reusable subroutines and their temporal organisation. approaches explore this vi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the HiSD pipeline. Demonstration trajecto￾ries are first segmented into skills. These skill sequences are then compressed and structured using a modified Sequitur algorithm, which identifies recurring mid-level subroutines across the dataset. The resulting grammar defines a hierarchical task decomposition. X(i) , we compute the optimal transport plan Γ ∗ by solving Equation 1. We obtain frame-l… view at source ↗
Figure 3
Figure 3. Figure 3: Example of the skill segmentation performance in the Stone Pickaxe Static Task in Craftax for all three baselines. Colours indicate discovered skills: wood, table, wooden pickaxe, stone, and stone pickaxe. 6.2. Minecraft Results We now present the same analysis in the Minecraft domain. To enable OMPN and CompILE to operate in this setting, we modify the dataset due to their reliance on discrete ac￾tion rep… view at source ↗
Figure 5
Figure 5. Figure 5: Mean episode rewards (±1 SD) over environment time steps for 10 random seeds on the Craftax Wooden Pickaxe task. The HiSD hierarchy (orange) achieves higher and more stable performance than both the OMPN hierarchy (green) and the Skills￾Only (HiSD) variant (purple), while closely matching the ground truth Hierarchy (blue). Primitive action PPO (cyan) fails to solve the task. ∼30k steps. Crucially, it outpe… view at source ↗
Figure 7
Figure 7. Figure 7: provides examples of the observations from both the Craftax (Matthews et al., 2024), and Minecraft (Baker et al., 2022) environments. (A) Example of an RGB observation from the modified Craftax environment (B) Example of an RGB observation from the Minecraft environ￾ment [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ground truth skill distributions. The histograms depict the frequency of skill usage across different Craftax tasks. The ‘wood‘ skill serves as a primary interaction mechanic in most configurations. Task Name Min Avg Max Stone Pickaxe (Static) 17 32.78 102 Stone Pickaxe (Random) 19 33.00 74 Mixed Task (Static) 3 14.91 58 Wood-Stone Coll. (Rand) 10 16.67 42 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ground truth skill distributions for Minecraft Tasks. The histograms show the frequency of skill usage. Again we see that the Mine Wood skills serve as a primary interaction mechanic in both settings. Mapped Category Fine-Grained Skills (Original Labels) crafted plank crafted acacia planks, crafted birch planks, crafted dark oak planks, crafted jungle planks, crafted oak planks, crafted spruce planks mined… view at source ↗
Figure 10
Figure 10. Figure 10: provides a qualitative visualization of the skills discovered by HiSD. A key strength of the method is its ability to learn object-centric behaviours that are robust to positional variance. As illustrated, HiSD consistently classifies the interaction with trees as a unified “Wood” skill, regardless of the agent’s approach vector or specific location in the grid. Furthermore, the model successfully disenta… view at source ↗
Figure 11
Figure 11. Figure 11: Figures from the Mixed Static task where the goal is to build a wooden pickaxe. (A) shows the skill segmentation between baselines. (B) is the ground truth tree, (C) is the tree discovered by HiSD, and (D) is the tree discovered by OMPN. Stone Pickaxe : Random Truth HiSD OMPN CmpILE (A) Skill Segmentation Comparison HiSD D1 D3 D2 Wood Wood Table W D5 ooden Pickaxe Wood Stone Stone Pickaxe (C) HiSD Discove… view at source ↗
Figure 12
Figure 12. Figure 12: Figures from the Stone Pick Random task where the goal is to build a stone pickaxe. (A) shows the skill segmentation between baselines. (B) is the ground truth tree, (C) is the tree discovered by HiSD, and (D) is the tree discovered by OMPN. We see HiSD matches the ground truth exactly, however, OMPN decomposes everything into one subtree (left), with no meaningful connection between subtrees and skills. … view at source ↗
Figure 13
Figure 13. Figure 13: Figures from the Wood Stone Wood Stone Random task where the goal in this case is to collect 2 wood, then 2 stone. (A) shows the skill segmentation between baselines. (B) is the ground truth tree, (C) is the tree discovered by HiSD, and (D) is the tree discovered by OMPN. In this case we see all implementations find the same tree decomposition. Stone Pickaxe : Static Truth HiSD OMPN CmpILE (A) Skill Segme… view at source ↗
Figure 14
Figure 14. Figure 14: Figures from the Wood Stone Pick Static task where the goal in this case is to collect a stone pickaxe. (A) shows the skill segmentation between baselines. (B) is the ground truth tree, (C) is the tree discovered by HiSD, and (D) is the tree discovered by OMPN. In this case, we see that even though the ground truth is a flat hierarchy, due to HiSD finding alternate skill sequencing, it leads to an informa… view at source ↗
Figure 15
Figure 15. Figure 15: Figures from the Minecraft Mapped task where the goal is to collect 2 stone pieces. (A) is the tree discovered by HiSD, (B) is the ground truth tree, and (C) shows the segmentation info. We note that the tree discovered by OMPN is omitted due to its size. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗

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

8 extracted references · 3 linked inside Pith

  1. [7]

    doi: 10.1609/ aaai.v31i1.10744

    ISSN 2374-3468, 2159-5399. doi: 10.1609/ aaai.v31i1.10744. URL https://ojs.aaai.org/ index.php/AAAI/article/view/10744. Pub- lisher: Association for the Advancement of Artificial Intelligence (AAAI). Torabi, F., Warnell, G., and Stone, P. Behavioral Cloning from observation.CoRR, abs/1805.01954, 2018. URL http://arxiv.org/abs/1805.01954. Xu, M. and Gould,...

    Pith/arXiv arXiv 2018

  2. [1933]

    Kanervisto, A., Karttunen, J., and Hautam ¨aki, V

    URL https://books.google.co.za/ books?id=qJfXAAAAMAAJ. Kanervisto, A., Karttunen, J., and Hautam ¨aki, V . Play- ing Minecraft with Behavioural Cloning. In Es- calante, H. J. and Hadsell, R. (eds.),Proceedings of the NeurIPS 2019 Competition and Demonstra- tion Track, volume 123 ofProceedings of Machine Learning Research, pp. 56–66. PMLR, 08–14 Dec

    2019

  3. [2016]

    doi: 10.1109/ICRA.2016.7487760. Ho, M. K., Abel, D., Griffiths, T. L., and Littman, M. L. The value of abstraction.Current Opinion in Behav- ioral Sciences, 29:111–116, 2019. ISSN 2352-1546. doi: https://doi.org/10.1016/j.cobeha.2019.05.001. URL https://www.sciencedirect.com/ science/article/pii/S2352154619300026. Artificial Intelligence. Hogg, C., Mu˜noz...

    arXiv 2016

  4. [2017]

    Lu, Y ., Shen, Y ., Zhou, S., Courville, A., Tenenbaum, J

    doi: 10.1109/CVPR.2017.113. Lu, Y ., Shen, Y ., Zhou, S., Courville, A., Tenenbaum, J. B., and Gan, C. Learning task decomposition with Ordered Memory Policy Network. InInternational Conference on Learning Representations, 2021. URL https:// openreview.net/forum?id=vcopnwZ7bC. Matthews, M., Beukman, M., Ellis, B., Samvelyan, M., Jackson, M., Coward, S., a...

    Pith/arXiv arXiv 2017

  5. [2019]

    org/CorpusID:199000710

    URL https://api.semanticscholar. org/CorpusID:199000710. Hayes, B. and Scassellati, B. Autonomously construct- ing Hierarchical Task Networks for planning and human- robot collaboration. In2016 IEEE International Confer- ence on Robotics and Automation (ICRA), pp. 5469–5476,

  6. [2020]

    Kipf, T., Li, Y ., Dai, H., Zambaldi, V ., Sanchez-Gonzalez, A., Grefenstette, E., Kohli, P., and Battaglia, P

    URL https://proceedings.mlr.press/ v123/kanervisto20a.html. Kipf, T., Li, Y ., Dai, H., Zambaldi, V ., Sanchez-Gonzalez, A., Grefenstette, E., Kohli, P., and Battaglia, P. Com- pILE: Compositional Imitation Learning and Execution. InProceedings of the 36th International Conference on Machine Learning, pp. 3418–3428. PMLR, May 2019. URL https://proceedings...

    Pith/arXiv arXiv 2019

  7. [2021]

    Correa, C

    URL https://proceedings.mlr.press/ v155/chen21d.html. Correa, C. G., Sanborn, S., Ho, M. K., Callaway, F., Daw, N. D., and Griffiths, T. L. Exploring the hierarchical struc- ture of human plans via program generation.Cognition, 255:105990, February 2025. ISSN 0010-0277. doi: 10. 1016/j.cognition.2024.105990. URL http://dx.doi. org/10.1016/j.cognition.2024...

    arXiv 2025

  8. [2022]

    All” (fine-grained) and “Mapped

    environments. (A)Example of an RGB observation from the modified Craftax environment (B)Example of an RGB observation from the Minecraft environ- ment Figure 7.Environment observation examples from Craftax (Matthews et al., 2024) (A) and Minecraft (Baker et al., 2022) (B). A.1. Craftax We model the Craftax environment (Matthews et al., 2024) as a fully ob...

    2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.