Pith. sign in

REVIEW 2 major objections 6 minor 20 references

Robot Pouring: Identifying Causes of Spillage and Selecting Alternative Action Parameters Using Probabilistic Actual Causation

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A probability-raising condition for actual causation can identify which robot pouring parameter caused a spill and select a changed value that prevents the spill in most reruns.

desk verdict First real application of probabilistic actual causation to robot action selection, with a solid pipeline but an evaluation that doesn't yet prove the method beats simpler baselines. read the letter →

arxiv 2502.09395 v3 pith:QTYNJP4Y submitted 2025-02-13 cs.RO cs.LG

Jaime Maldonado , Jonas Krumme , Christoph Zetzsche , Vanessa Didelez , Kerstin Schill This is my paper · ORCID
classification cs.ROcs.LG
keywords probabilisticactualcausationcausaldiscoveryrobotpouringactionselectionspillagepreventionBayesiannetworksneuralautoregressivedensityestimationcontrastiveexplanations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a formal criterion for singling out the cause of a particular observed event—probabilistic actual causation, stated as a probability-raising inequality—can tell a robot which task parameter caused a spill and which new value of that parameter should prevent the spill. It works through a complete causal model of a simulated pouring task: variables, a learned causal graph, and neural estimates of the conditional distributions. The quantitative support is a test set of 3000 pouring trials: from 1216 spillage trials, an alternative relative-diameter value was identified for 1191 trials and succeeded in 88.7% of reruns, and an alternative fullness value was identified for 718 trials and succeeded in 86.9% of reruns. If the approach holds up, corrective action in robotic manipulation can be selected from causal structure in a context-sensitive way, rather than by retrying blindly or using outcome probabilities that ignore the observed context.

What carries the argument

The engine is a probabilistic actual-causation inequality, in which $X = x$ rather than $X = x'$ is the actual cause of $Y = y$ when $P(Y \mid do(W = w^*, X = x, Z' = z^*)) > P(Y \mid do(W = w^*, X = x'))$ holds for every subset $Z'$ of the mediators along the chosen path. $W$ is the set of off-path variables, held at their observed values, so the test is context-specific rather than averaged. Feeding continuous contrastive values through this inequality yields a curve of spill probabilities for next actions; the automatic search then identifies the range where probability raising holds, restricts to values with predicted spill probability below a threshold, and picks the value closest to the current parameter. The DAG structure and the conditional distributions needed for the do-probabilities are learned from simulation data, with Monte Carlo integration over the estimated conditional distributions answering each query.

What would settle it

Run a controlled experiment in the same simulator where a hidden variable, for example marble diameter, varies between two distributions while all measured parameters are held at identical values; if the spill rate shifts systematically with the hidden variable, the DAG is not causally sufficient and the alternative parameters chosen by the actual-causation test should fail to transfer. A second check would be to compare the method's selected alternative value against a brute-force grid search over all parameter combinations in the simulator on the same spill trials.

Watch

Extended reading notes

Core claim

The paper's central claim is that the probabilistic actual-causation inequality can serve both as a diagnostic and as a prescription in a concrete manipulation task. In the learned causal graph, spillage has direct parents relative volume, fullness, and relative diameter, with relative capacity acting indirectly through relative volume; for a given spill trial the analysis fixes off-path variables and mediators at their observed values, then compares the reference probability $P(S \mid do(W = w^*, X = x, Z' = z^*))$ with $P(S \mid do(W = w^*, X = x'))$ over a grid of contrastive values $x'$. When probability raising holds, $x$ is an actual cause of that spill relative to $x'$, and the range of $x'$ where it holds contains candidate corrective values. The paper selects the value closest to the current one with predicted spill probability below 0.1, and shows that rerunning the simulated trials with those selected values succeeds most of the time, while the same causal probabilities averaged over contexts give much weaker and less actionable guidance. In short: the inequality identifies what to change and how much to change, in the context of the particular observed trial.

Load-bearing premise

The load-bearing premise is that the learned causal graph of the pouring task is the true causal structure, with no unobserved common causes; if a hidden factor jointly affects a task parameter and spillage, the interventional probabilities and probability-raising tests are biased and the suggested alternative parameter could be wrong.

Editorial extensions

If this is right

  • When the graph and probability estimates are reliable, the actual-causation search can automate corrective action selection after each failed trial, without random retries.
  • Direct causes are the most actionable: an alternative relative-diameter value was found in 97.9% of spill trials, an alternative fullness value in 59%, and an alternative relative-capacity value in only 2.7%.
  • The chosen alternatives produced empirical success rates of 88.7% for relative diameter and 86.9% for fullness, all above chance and in line with the probability threshold used.
  • Because the variable effects are non-linear, a small change in a parameter can separate a spilling from a successful pour, and the actual-causation scan finds such changes where human intuition might miss them.
  • Average causal probabilities are not enough: the causal probability of spillage given relative capacity stays near chance level, while the context-conditional actual-cause comparison reveals when and how much to change a parameter.

Reading between the lines

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

  • If this transfers to physical robots with accurate perception of container dimensions and fullness, the same inequality could choose corrective actions after failed manipulation attempts without random retries; that transfer is a testable extension the paper has not run.
  • The binary spillage variable discards severity, so a natural extension is a graded outcome such as number or fraction of particles spilled; the actual-causation test could then rank alternatives by expected severity reduction, and the authors themselves note that severity matters in some application contexts.
  • Because the graph is assumed causally sufficient, deployment would benefit from a robustness layer that compares the selected alternative against any other parameter change that also reduces predicted spill probability, so a misspecified graph cannot silently steer the correction.
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

2 major / 6 minor

Summary. The paper presents a complete pipeline for using Fenton-Glynn's probabilistic actual causation to diagnose spillage in a simulated robot pouring task and to select alternative values of task parameters. A causal DAG (RC, FU, RV, RD, S) is learned from simulated trials with the PC algorithm, conditional distributions are estimated with neural autoregressive density estimators, and the actual-causation inequality (1) is applied to test direct (RD) and mediated (FU) causes. On a held-out test set of 3000 trials (1216 spillage trials), an alternative RD value was identified in 1191 trials with 88.7% success when rerun, and an alternative FU value in 718 trials with 86.9% success. Four worked examples illustrate the context-sensitive selection behavior. The authors claim this goes beyond type-level causal probabilities because the analysis conditions on the actual values of other variables.

Significance. If correct, the paper provides one of the first practical implementations of token-level probabilistic actual causation for robotic action correction, and it demonstrates a complete causal-modeling pipeline (causal discovery, NADE-based do-probability estimation, and held-out simulation evaluation). Strengths include the use of a realistic physics simulation covering a large parameter space, bootstrapped structure learning, correct application of the Fenton-Glynn inequalities for both direct and mediated paths, and a held-out evaluation with a confusion matrix. The main unresolved issue is that the evaluation does not compare the proposed criterion against simpler baselines, so the paper's central claim that actual causation adds value beyond type-level causal probabilities is not yet established.

major comments (2)
  1. [Section 5.3; Figures 5e and 10] The evaluation reports only absolute success rates for the selected alternatives. Because the simulation is stochastic, a spillage trial may succeed on a simple retry; the retry success rate for the 1216 spillage trials is not reported. More importantly, the paper does not compare the proposed token-level selection with (a) changing the parameter to any low-probability value chosen by a type-level criterion such as P(S|do(RD)) below a threshold (which Section 5.1, Figure 5e, shows has a large low-probability region for RD greater than about 1.1), (b) a context-free threshold on the NADE prediction P(S|FU,RD,RV), or (c) changing a variable not identified as an actual cause. Without these baselines, the high success rates in Section 5.3 do not demonstrate that the actual-causation probability-raising test, rather than the low-probability threshold or the NADE predictions, is responsible for the performance. The authors should add these comparisons and report how often the probability-raising condition changes the selected value relative to a threshold-only rule.
  2. [Sections 4.4-4.5 and 5.1; inequalities (10)-(11)] The NADE models P(RV|RC,FU) as a Gaussian distribution, but in the data-generating process RV is a nearly deterministic function of FU and RC (Table 2), and the paper itself notes in Section 5.1 that the NADE for RV is smoothing the estimated probabilities. Because the FU actual-causation inequalities (10) and (11) integrate over P(RV|RC,FU), this smoothing can bias the probability-raising comparisons and therefore the conclusion that FU is an actual cause. The paper should quantify the approximation error (e.g., compare the NADE-implied conditional distribution with empirical conditional frequencies or use a conditional model reflecting the known near-determinism) and report sensitivity of the FU results to this choice.
minor comments (6)
  1. [Section 5.3] The phrase 'alternative DD values' is a typo and should read 'alternative RD values'.
  2. [Section 5.3] The sentence 'Running these trials with the alternative F Uvalues success rate of 86.9%' is missing a verb and should be 'produced a success rate of 86.9%'.
  3. [Section 5.3] Report 95% confidence intervals for the success rates (88.7% and 86.9%) and for the alternative-identification rates (97.9% for RD and 59% for FU); the histograms in Figure 10 help but only cover 100 trials.
  4. [Section 5.3] State explicitly that for FU the probability-raising test requires both inequalities (10) and (11), and report how often each is satisfied; the current text says only that 'probability raising holds' without distinguishing the two conditions.
  5. [Sections 4.1 and 5.3] Clarify the relationship between the 6000 simulated trials mentioned in Section 4.1 and the 3000-trial test set in Section 5.3, i.e., whether the training and test sets are disjoint and how many trials were used for training.
  6. [Section 5.1] The term 'chance level' is used without a definition; specify that it refers to a spillage probability of 0.5.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the actual-causation analysis is applied as stated, and the alternative-action success rates are evaluated on a separate physics-simulation test set rather than re-stating the model's own inputs.

full rationale

The derivation chain runs from (i) a DAG discovered with the PC algorithm on simulated trials (Fig. 3), (ii) NADE-estimated conditional distributions (Eq. 3), (iii) Fenton-Glynn probability-raising inequalities (Eqs. 9-11), to (iv) alternative-parameter selection using a fixed 0.1 probability threshold (Sections 2.1 and 5.3), and (v) empirical evaluation on a separate 3000-trial simulation test set. No load-bearing input is defined in terms of the target claim: the graph is learned from data rather than imposed to match the spillage outcome; the conditional distributions are estimated on training data and then used to select alternative RD/FU values; the actual-causation inequalities are applied exactly as stated by Fenton-Glynn and do not contain a fitted parameter later renamed as a prediction; and the reported 88.7% and 86.9% success rates are obtained by re-running the physics simulation with the suggested alternative values, which is an independent benchmark rather than a re-statement of the model's own probability estimates. The acknowledged limitations (no latent variables assumed, no ground-truth causal graph, binary spillage representation, and the absence of baselines such as random restarts or type-level causal-probability criteria) are correctness or external-validity concerns, not circularity. There are no load-bearing self-citations: the cited methodological works are external, and no uniqueness theorem from the present authors is invoked to force the modeling choice. The paper is therefore self-contained against an external simulation benchmark, and the correct circularity finding is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

All ingredients of the central claim come from three sources: the simulation (axioms on simulation fidelity and Gaussian representation), the causal-modeling pipeline (axioms on graph correctness and NADE accuracy), and the paper's own action-guidance heuristic (the probability-raising-to-alternative rule and the user-set probability threshold). There are no newly postulated physical entities.

free parameters (3)
  • probability threshold for alternative selection = 0.1 (evaluation), 0.2 and 0.5 (examples)
    In Sections 2.1 and 5.3, the authors require that an alternative value has spillage probability below a pre-defined threshold. The value is a user choice, not derived or fitted, but it determines which alternatives qualify and thus the reported success rates.
  • causal discovery edge-frequency threshold = 0.5 for stable edges, 0.75 to construct the reported DAG
    Supplementary Section 7.1: the DAG is constructed from edges present in more than 75% of 1000 bootstrap PC runs. A different threshold could yield a different graph and different actual-causation conclusions.
  • NADE hyperparameters = 2 hidden layers of 16 units, RMSProp, learning rate 0.01
    Supplementary Section 7.3: taken from Garrido et al. without systematic tuning. They affect the estimated conditional distributions that all causal probabilities are computed from.
assumptions (5)
  • domain assumption The PC-discovered DAG with edge frequencies above the bootstrap threshold is the true causal structure (RC->RV, FU->RV, RV->S, FU->S, RD->S), with no unobserved confounders.
    Supplementary Section 7.2 explicitly assumes no latent variables; Section 4.3 states that the validity of the actual causation analysis relies on the correctness of the DAG structure. If the graph is wrong, all do-operations and probability-raising tests are biased.
  • domain assumption The CoppeliaSim/ODE simulation is a faithful proxy for real robot pouring dynamics.
    All training and test data come from this simulation (Section 4.1); the paper claims practical use but never validates on a physical robot. If particle behavior in simulation diverges from reality, the quantitative success rates would not transfer.
  • ad hoc to paper Choosing a contrastive value where probability raising holds and predicted spillage probability is below threshold is a valid corrective action.
    Section 2.1 proposes this as the action-guidance rule. It is the central heuristic: the paper assumes that lower predicted probability in the observed context will change the outcome in the intended direction, and evaluates this empirically rather than proving it.
  • domain assumption The NADE conditional densities accurately represent the true conditional distributions over the support of the data.
    Section 4.4 relies on NADE as a universal approximator and on sufficient training support; the authors note in Section 5.1 that P(S|do(RV)) is smoothed compared with the expected step function, showing the approximation is imperfect.
  • domain assumption Continuous variables RC, FU, RV, and RD are well approximated by Gaussian conditional distributions.
    Section 4.5 says each continuous variable is modeled by a Gaussian output from its NADE; if the true marginals are non-Gaussian (e.g., RV is a ratio with a heavy tail), the causal probabilities will be biased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robot Pouring: Identifying Causes of Spillage and Selecting Alternative Action Parameters Using Probabilistic Actual Causation." pith.science (2026). https://pith.science/paper/QTYNJP4Y

@misc{pith2026250209395,
  author       = {Pith},
  title        = {Pith review of: Robot Pouring: Identifying Causes of Spillage and Selecting Alternative Action Parameters Using Probabilistic Actual Causation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QTYNJP4Y}},
  note         = {Machine review of arXiv:2502.09395}
}
read the original abstract

In everyday life, we perform tasks (e.g., cooking or cleaning) that involve a large variety of objects and goals. When confronted with an unexpected or unwanted outcome, we take corrective actions and try again until achieving the desired result. The reasoning performed to identify a cause of the observed outcome and to select an appropriate corrective action is a crucial aspect of human reasoning for successful task execution. Central to this reasoning is the assumption that a factor is responsible for producing the observed outcome. In this paper, we investigate the use of probabilistic actual causation to determine whether a factor is the cause of an observed undesired outcome. Furthermore, we show how the actual causation probabilities can be used to find alternative actions to change the outcome. We apply the probabilistic actual causation analysis to a robot pouring task. When spillage occurs, the analysis indicates whether a task parameter is the cause and how it should be changed to avoid spillage. The analysis requires a causal graph of the task and the corresponding conditional probability distributions. To fulfill these requirements, we perform a complete causal modeling procedure (i.e., task analysis, definition of variables, determination of the causal graph structure, and estimation of conditional probability distributions) using data from a realistic simulation of the robot pouring task, covering a large combinatorial space of task parameters. Based on the results, we discuss the implications of the variables' representation and how the alternative actions suggested by the actual causation analysis would compare to the alternative solutions proposed by a human observer. The practical use of the analysis of probabilistic actual causation to select alternative action parameters is demonstrated.

Figures

Figures reproduced from arXiv: 2502.09395 by the authors.

Figure 1
Figure 1. Example of actual causation test. The curve corresponds to the right side of inequality (1) as a function of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Course of pouring trial. (a) Source container is filled with simulated marbles and a target container of random dimensions is generated. (b) The source container is transported to the pouring position. (c) The source container is rotated to pour the marbles into the target container. (d) Trial end. In each trial, the source container (capacity = 514.72 cm3 ) was filled with simulated marbles (Figure 2a). The charact… view at source ↗
Figure 3
Figure 3. Discovered DAG A DAG model represents the conditional dependencies between the set of J random variables X1, · · · , XJ . The DAG induces a joint probability distribution P(X) of the variables, which can be factorized into the conditional distributions of each Xj conditioned on a function fj of its parents P A(Xj ) (the causal Markov condition): P(X) = Y J j P(Xj |fj (P A(Xj ))) (2) A NADE is a generative model that… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Implemented NADEs for the DAG nodes and auxiliary NADE used for causal effect estimation [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Causal probabilities of DAG variables. The probabilities P(S|do(F U)) and P(S|do(F U, RV )) were computed using equations (5) and (6), respectively, over a range of F U values covering low (F U < 0.4), medium (0.4 ≤ F U ≤ 0.6), and high (F U > 0.6) fullness levels of t…
Figure 6
Figure 6. Figure 6: Actual causation analysis of example 1. (a) Actual trial parameters of trial with spillage and (b) outcome frequencies over 100 replications. (c) Probabilities of actual causation inequality for RD. (d) Probabilities of actual causation inequality for F U. (e) Alternat…
Figure 7
Figure 7. Figure 7: Actual causation analysis of example 2. (a) Actual trial parameters of trial with spillage and (b) outcome frequencies over 100 replications. (c) Probabilities of actual causation inequality for RD. (d) Probabilities of actual causation inequality for F U. 14 [PITH_FU…
Figure 8
Figure 8. Figure 8: Actual causation analysis of example 3. (a) Actual trial parameters of trial with spillage and (b) outcome frequencies obtained over 100 replications. (c) Probabilities of actual causation inequality for RD. (d) Probabilities of actual causation inequality for F U. (e)…
Figure 9
Figure 9. Figure 9: Actual causation analysis of example 4. (a) Actual trial parameters of trial with spillage and (b) outcome frequencies over 100 replications. (c) Probabilities of actual causation inequality for RD. (d) Probabilities of actual causation inequality for F U. (e) Alternat…
Figure 10
Figure 10. Figure 10: Empirical success rates obtained over 10 replications. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Actual causation probabilities, actual trial parameters and target containers with different [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: PC and DG parameters specified in Tetrad. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Background knowledge specified in Tetrad. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Discovered edge type frequencies. the stochastic effect of 1) the containers’ characteristics (capacities and diameters, expressed as the variables RC and RD) and 2) the poured amount (fullness and volume, expressed as the variables F U and RV ) on the probability of …
Figure 15
Figure 15. Figure 15: Confusion matrix. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 10 canonical work pages

  1. [1]

    The Actual Cause, chapter 10, pages 309–330

    Judea Pearl. The Actual Cause, chapter 10, pages 309–330. Cambridge University Press, Cambridge, September 2009a. ISBN 9780521749190. doi: 10.1017/cbo9780511803161.012. Sander Beckers. Causal explanations and XAI. In Bernhard Schölkopf, Caroline Uhler, and Kun Zhang, editors, Proceedings of the First Conference on Causal Learning and Reasoning, volume 177...

    work page doi:10.1017/cbo9780511803161.012

  2. [2]

    4.2 DAG Variables

    the poured amount (fullness and volume, expressed as the variables F Uand RV ) on the probability of spilling (variable S). The randomness of the outcome S results from the interplay between RC, RD, F U, and RV and the behavior of the particles during the pouring movement. The parameters the physics engine uses (particle size and density) to simulate the ...

    work page 2021

  3. [7]

    Ramaravind Kommiya Mothilal, Divyat Mahajan, Chenhao Tan, and Amit Sharma

    doi: 10.48550/ARXIV .2401.13752. Ramaravind Kommiya Mothilal, Divyat Mahajan, Chenhao Tan, and Amit Sharma. Towards unifying feature attribution and counterfactual explanations: Different means to the same end. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, AIES ’21, pages 652–663. ACM, July

    arXiv doi:10.48550/arxiv.2401.13752

  4. [10]

    doi: 10.1109/lra.2022.3188889

    ISSN 2377-3774. doi: 10.1109/lra.2022.3188889. Eric Rohmer, Surya P. N. Singh, and Marc Freese. V-rep: A versatile and scalable robot simulation framework. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems . IEEE, November

    arXiv 2022

  5. [13]

    4.2 DAG Variables

    Note that within-tier edges are allowed in the background knowledge. To validate the causal relationships inferred from the data, we performed a bootstrapping analysis. Conducting a bootstrapping analysis is recommended for causal discovery[Malinsky and Danks, 2017, Glymour et al., 2019]. We ran the PC algorithm on 1000 bootstraps, producing 1000 differen...

    work page 2017

  6. [14]

    doi: 10.1002/widm.1449

    ISSN 1942-4795. doi: 10.1002/widm.1449. Daniel Malinsky and David Danks. Causal discovery algorithms: A practical guide. Philosophy Compass, 13(1), November

    doi:10.1002/widm.1449 1942

  7. [15]

    doi: 10.1111/phc3.12470

    ISSN 1747-9991. doi: 10.1111/phc3.12470. Joseph D Ramsey, Kun Zhang, Madelyn Glymour, Ruben Sanchez Romero, Biwei Huang, Imme Ebert-Uphoff, Savini Samarasinghe, Elizabeth A Barnes, and Clark Glymour. Tetrad—a toolbox for causal discovery. In 8th international workshop on climate informatics,

    work page doi:10.1111/phc3.12470

  8. [16]

    doi: 10.1515/jci-2020-0007

    ISSN 2193-3685. doi: 10.1515/jci-2020-0007. Santtu Tikka and Juha Karvanen. Identifying causal effects with the R package causaleffect. Journal of Statistical Software, 76(12):1–30,

    work page doi:10.1515/jci-2020-0007 2020

Show all 20 references
  1. [17]

    doi: 10.18637/jss.v076.i12. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamku...

    doi

  2. [18]

    URL https://proceedings.neurips.cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740-Paper.pdf. Supplementary Material 7.1 Causal discovery using the PC algorithm - Tetrad setup and bootstrapping results Figure 12 shows the exact parameters used in Tetrad to run th...

    2019

  3. [1991]

    doi: 10.1177/089443939100900106

    ISSN 1552-8286. doi: 10.1177/089443939100900106. Clark Glymour, Kun Zhang, and Peter Spirtes. Review of causal discovery methods based on graphical models. Frontiers in Genetics, 10, June

    doi

  4. [2005]

    doi: 10.1093/bjps/axi147

    ISSN 1464-3537. doi: 10.1093/bjps/axi147. Jan Borner. Causal explanations - how to generate, identify, and evaluate them. PhD thesis,

    doi

  5. [2013]

    www.coppeliarobotics.com

    doi: 10.1109/iros.2013.6696520. www.coppeliarobotics.com. Peter Spirtes and Clark Glymour. An algorithm for fast recovery of sparse causal graphs. Social Science Computer Review, 9(1):62–72, April

    2013

  6. [2017]

    doi: 10.1093/ bjps/axv056

    ISSN 1464-3537. doi: 10.1093/ bjps/axv056. Judea Pearl. Causal Diagrams and the Identification of Causal Effects, chapter 3, pages 65–106. Cambridge University Press, September 2009b. ISBN 9780521749190. doi: 10.1017/cbo9780511803161.005. Maximilian Diehl and Karinne Ramirez-A...

    doi

  7. [2019]

    doi: 10.3389/fgene.2019.00524

    ISSN 1664-8021. doi: 10.3389/fgene.2019.00524. 21 Robot Pouring - Probabilistic Actual Causation Ana Rita Nogueira, Andrea Pugnana, Salvatore Ruggieri, Dino Pedreschi, and João Gama. Methods and tools for causal discovery and causal inference. WIREs Data Mining and Knowledge D...

    2019

  8. [2020]

    A causal approach to tool affordance learning

    Jake Brawer, Meiying Qin, and Brian Scassellati. A causal approach to tool affordance learning. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 8394–8399. IEEE,

    2020

  9. [2021]

    Ossama Ahmed, Frederik Träuble, Anirudh Goyal, Alexander Neitz, Yoshua Bengio, Bernhard Schölkopf, Manuel Wüthrich, and Stefan Bauer

    doi: 10.1145/3461702.3462597. Ossama Ahmed, Frederik Träuble, Anirudh Goyal, Alexander Neitz, Yoshua Bengio, Bernhard Schölkopf, Manuel Wüthrich, and Stefan Bauer. Causalworld: A robotic manipulation benchmark for causal structure and transfer learning. arXiv preprint arXiv:20...

  10. [2022]

    doi: 10.1007/978-3-031-24670-8_9

    ISBN 9783031246708. doi: 10.1007/978-3-031-24670-8_9. Ehsan Zibaei and Robin Borth. Building causal models for finding actual causes of unmanned aerial vehicle failures. Frontiers in Robotics and AI, 11, February

    doi

  11. [2023]

    doi: 10.1016/j.robot

    ISSN 0921-8890. doi: 10.1016/j.robot. 2023.104376. Hugo Araujo, Patrick Holthaus, Marina Sarda Gou, Gabriella Lakatos, Giulia Galizia, Luke Wood, Ben Robins, Mohammad Reza Mousavi, and Farshid Amirabdollahian. Kaspar Causally Explains, pages 85–99. Springer Nature Switzerland,

    2023

  12. [2024]

    doi: 10.3389/frobt.2024.1123762

    ISSN 2296-9144. doi: 10.3389/frobt.2024.1123762. Hana Chockler, Daniel Kroening, and Youcheng Sun. Explanations for occluded images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1234–1243, October

    2024

Pith tools

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