Pith. sign in

REVIEW 2 major objections 36 references

A shared flow-matching policy can turn passive source noise into a real choice among robot futures by selecting only where the flow starts, not a mode-specific field.

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 ↗

T0 review · grok-4.5

2026-07-14 13:29 UTC pith:WMTM5XXZ

load-bearing objection Solid robotics methods paper: source-only handles plus Orthogonal Source Lifting make multimodal flow policies intervenable without mode-conditioned fields, with the right causal controls. the 2 major comments →

arxiv 2607.10206 v1 pith:WMTM5XXZ submitted 2026-07-11 cs.RO cs.AI

Source-Lifted Flow Matching for Intervenable Multimodal Imitation

He Zhang , Ying Sun , Pengteng Li , Ziyang Chen , Yiren Zhao , Ziyang Rao , Weiyu Guo , Yandong Guo
show 1 more author
Hui Xiong
This is my paper
classification cs.RO cs.AI
keywords flow matchingimitation learningmultimodal policiessource interventionorthogonal source liftingrobot controlgenerative policies
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.

Flow-matching policies can represent several valid robot actions for the same observation, yet that diversity is usually only random sampling: a user cannot deliberately pick which continuation happens from a fixed state. This paper argues that the choice can be exposed without splitting the model into mode-conditioned dynamics. The discrete handle selects only the source endpoint of a single shared, latent-free velocity field. Orthogonal Source Lifting raises handle-specific sources into auxiliary orthogonal coordinates while keeping every demonstrated target in the original action subspace, so branch identity survives path crossings and the demonstrated action distribution is left intact. A state-dependent source mixture plus a responsibility floor keeps every handle trainable. On crossing diagnostics and robot benchmarks, changing the handle redirects future routes in 91.1% of matched-prefix interventions while free-deployment performance remains competitive.

Core claim

Source-Lifted Flow Matching shows that a discrete handle which chooses only the source endpoint of a conditional flow—never conditioning the velocity field—is enough to make multimodal imitation intervenable, when handle-specific sources are lifted into orthogonal auxiliary coordinates and all targets remain on the original action subspace. One shared latent-free field then carries distinct branches through crossings without averaging them into composite trajectories, converting passive source randomness into an actionable local control variable.

What carries the argument

Orthogonal Source Lifting: each handle’s source is embedded with an auxiliary orthogonal coordinate while every target action stays at zero lift in the original action space; this keeps source identity through crossings so one shared velocity field can transport different branches without merging.

Load-bearing premise

That a modest number of state-conditioned Gaussian source components plus orthogonal lift coordinates is enough to keep chosen branches distinct through real closed-loop crossings without ever feeding the discrete choice into the velocity field.

What would settle it

On a same-prefix intervention protocol like D3IL Avoiding, if changing only the local source handle fails to redirect future routes in a large fraction of successful pairs—or produces composite averaged trajectories at known crossings—while free deployment still succeeds, the claim that source geometry alone yields an actionable handle would be falsified.

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

If this is right

  • A planner or operator can set a local source handle at a decision state to choose among valid futures that share the same prefix.
  • Multimodal robot policies need not condition the velocity field on a discrete mode to expose controllable branches.
  • Free sampling from the learned state-dependent prior recovers ordinary stochastic imitation while the same interface supports deliberate intervention.
  • A high-level selector can treat the source handle as a compact discrete action space over routes or subtasks with the low-level field frozen.
  • Crossing-induced composite trajectories that arise under a shared field can be removed by the lift construction without partitioning the target action distribution.

Where Pith is reading between the lines

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

  • Language, goal images, or other semantic signals could bias the state-dependent source prior to ground high-level intent without ever conditioning the shared velocity field on a mode label.
  • If lift dimension and mixture capacity scale with denser crossings in high-dimensional action chunks, the same source-only interface may extend to long-horizon generalist policies that currently rely on latent or mode-conditioned heads.
  • Same-prefix causal tests of the kind used here could become a standard check for whether any discrete “mode” variable in a generative policy is actually intervenable rather than merely correlated with outcomes.
  • The responsibility floor pattern may generalize to any overcomplete discrete interface attached only to generative sources, as a simple way to keep unused handles trainable.

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

2 major / 0 minor

Summary. The paper proposes Source-Lifted Flow Matching (SL-FM), a conditional flow-matching policy for multimodal imitation that exposes a discrete source handle z without conditioning the velocity field on z. A state-dependent Gaussian mixture defines selectable source endpoints; Orthogonal Source Lifting embeds handle-specific sources into auxiliary orthogonal coordinates while keeping all targets on the zero-lift plane (Eqs. 10–13), so one shared latent-free field can transport different branches without identity collapse at crossings. A responsibility floor (Eq. 14) keeps redundant handles trainable. Free deployment samples z from the learned prior; intervention sets z at a decision state and integrates the same field. On D3IL Avoiding, same-prefix interventions change future routes in 91.1% of pairs (same-handle control 0%), free-deployment success is competitive or best among same-harness baselines (Tables 1–2), and a frozen-policy high-level selector can use z for route control.

Significance. If the result holds, SL-FM supplies a clean, source-only intervention interface for generative imitation policies that is stricter than mode-conditioned fields or latent steering: the handle never enters the velocity field, the target action marginal is preserved, and same-prefix counterfactuals with a same-handle null control provide causal evidence of local controllability. The constructive method (Algorithm 1), w/o-lift ablation, MI diagnostics, and selector study make the contribution concrete and usable for planning or human-in-the-loop control on multimodal robot tasks. Strengths include the matched-prefix protocol, explicit preservation of pdata(a|s), and competitive free-deployment averages on Avoiding/Aligning/PushT.

major comments (2)
  1. The central claim that source geometry alone keeps branch identity through crossings rests on modest K and orthogonal lifts (Methodology, Eqs. 10–13; responsibility floor Eq. 14). The manuscript already shows this works on the reported regime (Avoiding 91.1% route change, w/o-lift drop 0.825→0.733, same-handle control 0%). However, the weakest assumption is left largely untested: denser crossings, higher-dimensional action chunks, or poorly aligned mixtures could still cause the shared field to average branches. A short stress test (e.g., synthetic denser-crossing diagnostic or higher-da chunk ablation) would substantially strengthen the load-bearing claim that the handle remains reliable without ever feeding z into vθ.
  2. Table 3 and the same-prefix protocol are the primary causal evidence, yet the intervention window is fixed to five consecutive policy calls at the first high-entropy pre-obstacle band. It is not shown how sensitive the 91.1% / 57.8% both-success rates are to window length, decision-point selection, or later obstacles (where free-rollout MI for digit2 is weaker). Clarifying robustness of the intervention protocol would make the causal claim more transferable beyond the Avoiding geometry.

Circularity Check

0 steps flagged

No significant circularity: SL-FM is a constructive method with independent empirical intervention metrics, not a result forced by definition or self-citation.

full rationale

The paper's load-bearing chain is constructive and empirically tested, not tautological. The source mixture (Eqs. 5–8) is fitted to demonstrated actions via Lsrc, Orthogonal Source Lifting (Eqs. 10–13) is an explicit geometric design that keeps targets on the zero-lift plane by construction (Eq. 11), and the shared field is trained with floor-weighted regression (Eqs. 14–16). None of these steps redefine the evaluation claims. Intervenability is measured on held-out closed-loop rollouts with a same-prefix counterfactual protocol, a same-handle null control (0% route change), a w/o-lift ablation, free-deployment benchmarks against external baselines, and a frozen-policy selector study. Route labels ρ are explicitly evaluation-time labels, not training targets. Self-citations (e.g., GoldenStart) appear only in related work and are not used as uniqueness theorems or load-bearing premises. Target-marginal preservation is a stated design property, not a circular prediction. No fitted parameter is renamed as a prediction of a closely related quantity, and no derivation reduces Eq. X to Eq. Y by construction in a way that forces the 91.1% intervention result.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claim rests on standard conditional flow matching plus several design choices that are free or paper-specific: a finite state-dependent Gaussian mixture as the selectable source, orthogonal auxiliary lift coordinates with anchors, a responsibility floor, and the modeling premise that source geometry alone can carry branch identity through crossings without conditioning vθ on z. No new physical entity is postulated; the invented pieces are algorithmic constructs whose only evidence is the reported robotics diagnostics.

free parameters (5)
  • K (number of source handles)
    Chosen by hand (defaults K=4 on Avoiding; K=8 Kitchen variant; sweeps K=4–10). Intervention capacity and dead-mode risk depend on this count.
  • responsibility floor α
    Hand-set mixing weight between soft responsibilities and uniform 1/K supervision (default α=0.25). Directly affects whether redundant handles remain trainable.
  • lift scale λ / tag scale
    Controls separation of lifted sources; sensitivity plots show unstable scales hurt robustness. Chosen, not derived.
  • source-loss weight w_src and anchor weight β
    Scalar trade-offs in the joint objective L = L_W-FM + w_src L_src + β L_anc; set by authors and swept in ablations.
  • Euler integration steps and source/component scales σ
    Deployment and source geometry knobs (e.g., 10 Euler steps, σ_iso=1.0 defaults) that affect measured success and intervention reliability.
axioms (5)
  • domain assumption Conditional flow matching with linear paths learns a velocity field transporting x0~p0(·|s) to demonstrated actions a~pdata(·|s).
    Preliminaries Eqs. 1–3; standard CFM setup from Lipman et al. / Tong et al., taken as given.
  • ad hoc to paper A shared velocity field that never receives discrete handle z can still realize selectable multimodal continuations if source endpoints are geometrically separated.
    Core interface requirement stated in Introduction and Methodology; not guaranteed by standard CFM and is what Orthogonal Source Lifting is designed to enforce.
  • domain assumption Demonstrated multimodal strategies can be adequately covered by a small state-conditioned isotropic Gaussian mixture over action-space sources.
    Modeling the Source Prior, Eq. 5; mixture-density style prior used as the handle set.
  • standard math Keeping all targets on the zero-lift plane preserves the imitation target distribution while lift coordinates may be discarded at execution.
    Eqs. 10–11: projection of p1 back to action space recovers pdata by construction of the embedding.
  • domain assumption Same-prefix route labels and success metrics are valid external tests of causal local intervention (not training labels).
    Experiments protocol for D3IL Avoiding matched-prefix interventions; evaluation-time ρ(·) defined as route/continuation label.
invented entities (3)
  • Orthogonal Source Lifting (auxiliary orthogonal lift coordinates for handle-specific sources) no independent evidence
    purpose: Prevent path-crossing identity collapse so one shared field can carry distinct branches without merging velocities at action-space crossings.
    Introduced in Methodology as the core mechanism; evidence is internal diagnostics (composite-trajectory removal, intervention rates), not an independently measured physical object.
  • Source handle z as a source-only intervention variable no independent evidence
    purpose: Expose an externally settable discrete choice of flow source endpoint under free deployment or intervention.
    Defined as selecting mixture component / lifted source only; independent evidence is behavioral (route change under matched prefix), not a prior established control primitive outside this construction.
  • Responsibility floor γk for floor-weighted flow training no independent evidence
    purpose: Give every handle weak transport supervision to mitigate dead modes when K is redundant.
    Eq. 14; training device validated by sensitivity sweeps, not an external scientific entity.

pith-pipeline@v1.1.0-grok45 · 19073 in / 3753 out tokens · 37873 ms · 2026-07-14T13:29:34.061720+00:00 · methodology

0 comments
read the original abstract

Flow-matching policies are promising for imitation learning because they model complex multimodal action distributions. However, their stochasticity is largely passive: repeated sampling may yield diverse behaviors, but users cannot directly choose among valid continuations from the same state. We propose Source-Lifted Flow Matching (SL-FM), a source-intervenable flow-matching policy that exposes such a handle while keeping the velocity field shared and latent-free. The handle selects only the source endpoint of the conditional flow, not a mode-specific field, preserving the standard formulation while avoiding decomposition into separate mode-conditioned dynamics. The core mechanism is \textbf{Orthogonal Source Lifting}, designed to prevent path-crossing ambiguity. Instead of partitioning target actions by mode, SL-FM lifts handle-specific sources into auxiliary orthogonal coordinates and keeps targets in the original action subspace. This preserves the demonstrated action distribution while allowing one shared field to carry different branches without merging at crossings. To keep handles usable across states, we learn a state-dependent source mixture end to end and use a responsibility floor, giving each handle weak supervision and mitigating dead modes. Experiments on crossing-flow diagnostics and robot-control benchmarks show that SL-FM converts passive source randomness into an actionable intervention variable. It removes crossing-induced composite trajectories, changes future routes in 91.1\% of matched-prefix interventions, and achieves strong free-deployment performance, with improvements in several benchmark settings. Overall, source geometry provides actionable multimodal control without conditioning the velocity field on the selected mode.

Figures

Figures reproduced from arXiv: 2607.10206 by He Zhang, Hui Xiong, Pengteng Li, Weiyu Guo, Yandong Guo, Ying Sun, Yiren Zhao, Ziyang Chen, Ziyang Rao.

Figure 1
Figure 1. Figure 1: Main idea of Source-Lifted Flow Matching. Stan [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Same-prefix source-intervention problem. Standard [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Method overview of Source-Lifted Flow Matching. SL-FM models a state-conditioned source prior, lifts handle [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: D4RL Kitchen diagnostic. A: 7-task completion [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hyperparameter and deployment-cost sensitivity on D3IL Avoiding, varying one setting per panel and keeping the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-step mutual information between sampled [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Fixed-state downstream selector diagnostic on [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional PushT branching examples. Columns show selected PushT states. Forced source choices in the source [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative successful downstream route-control trajectories on Avoiding. The low-level SL-FM policy is frozen, [PITH_FULL_IMAGE:figures/full_fig_p010_9.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

36 extracted references · 9 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , title =. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  2. [2]

    and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , title =

    Song, Yang and Sohl-Dickstein, Jascha and Kingma, Diederik P. and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , title =. International Conference on Learning Representations , year =

  3. [3]

    Lipman, Yaron and Chen, Ricky T. Q. and Ben-Hamu, Heli and Nickel, Maximilian and Le, Matt , title =. International Conference on Learning Representations , year =

  4. [4]

    Transactions on Machine Learning Research , year =

    Tong, Alexander and Fatras, Kilian and Malkin, Nikolay and Huguet, Guillaume and Zhang, Yanlei and Rector-Brooks, Jarrid and Wolf, Guy and Bengio, Yoshua , title =. Transactions on Machine Learning Research , year =. 2302.00482 , archivePrefix =

    Pith/arXiv arXiv

  5. [5]

    Pooladian, Aram-Alexandre and Ben-Hamu, Heli and Domingo-Enrich, Carles and Amos, Brandon and Lipman, Yaron and Chen, Ricky T. Q. , title =. Proceedings of the 40th International Conference on Machine Learning , series =. 2023 , url =

    2023

  6. [6]

    International Conference on Learning Representations , year =

    Liu, Xingchao and Gong, Chengyue and Liu, Qiang , title =. International Conference on Learning Representations , year =

  7. [7]

    and Vanden-Eijnden, Eric , title =

    Albergo, Michael S. and Vanden-Eijnden, Eric , title =. International Conference on Learning Representations , year =

  8. [8]

    , title =

    Guo, Pengsheng and Schwing, Alexander G. , title =. arXiv preprint arXiv:2502.09616 , year =. 2502.09616 , archivePrefix =

    Pith/arXiv arXiv

  9. [9]

    Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin C. M. and Song, Shuran , title =. Proceedings of Robotics: Science and Systems , address =. 2023 , doi =

    2023

  10. [10]

    International Conference on Learning Representations , year =

    Pearce, Tim and Rashid, Tabish and Kanervisto, Anssi and Bignell, Dave and Sun, Mingfei and Georgescu, Raluca and Macua, Sergio Valcarcel and Tan, Shan Zheng and Momennejad, Ida and Hofmann, Katja and Devlin, Sam , title =. International Conference on Learning Representations , year =

  11. [11]

    Proceedings of Robotics: Science and Systems , address =

    Reuss, Moritz and Li, Maximilian and Jia, Xiaogang and Lioutikov, Rudolf , title =. Proceedings of Robotics: Science and Systems , address =. 2023 , doi =

    2023

  12. [12]

    and Wahid, Ayzaan and Downs, Laura and Wong, Adrian and Lee, Johnny and Mordatch, Igor and Tompson, Jonathan , title =

    Florence, Pete and Lynch, Corey and Zeng, Andy and Ramirez, Oscar A. and Wahid, Ayzaan and Downs, Laura and Wong, Adrian and Lee, Johnny and Mordatch, Igor and Tompson, Jonathan , title =. Proceedings of the 5th Conference on Robot Learning , series =. 2022 , url =

    2022

  13. [13]

    Advances in Neural Information Processing Systems , volume =

    Shafiullah, Nur Muhammad Mahi and Cui, Zichen Jeff and Altanzaya, Ariuntuya and Pinto, Lerrel , title =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  14. [14]

    International Conference on Learning Representations , year =

    Jia, Xiaogang and Blessing, Denis and Jiang, Xinkai and Reuss, Moritz and Donat, Atalay and Lioutikov, Rudolf and Neumann, Gerhard , title =. International Conference on Learning Representations , year =

  15. [15]

    Streaming Flow Policy: Simplifying Diffusion/Flow-Matching Policies by Treating Action Trajectories as Flow Trajectories , booktitle =

    Jiang, Sunshine and Fang, Xiaolin and Roy, Nicholas and Lozano-P. Streaming Flow Policy: Simplifying Diffusion/Flow-Matching Policies by Treating Action Trajectories as Flow Trajectories , booktitle =. 2025 , url =

    2025

  16. [16]

    Proceedings of Robotics: Science and Systems , address =

    Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and Jakubczak, Szymon and Jones, Tim and Ke, Liyiming and Levine, Sergey and Li-Bell, Adrian and Mothukuri, Mohith and Nair, Suraj and Pertsch, Karl and Shi, Lucy Xiaoyang and ...

    2025

  17. [17]

    2504.16054 , archivePrefix =

    arXiv preprint arXiv:2504.16054 , year =. 2504.16054 , archivePrefix =

    Pith/arXiv arXiv

  18. [18]

    2503.14734 , archivePrefix =

    arXiv preprint arXiv:2503.14734 , year =. 2503.14734 , archivePrefix =

    Pith/arXiv arXiv

  19. [19]

    Advances in Neural Information Processing Systems , year =

    Du, Maximilian and Song, Shuran , title =. Advances in Neural Information Processing Systems , year =. 2506.13922 , url =

    arXiv

  20. [20]

    Proceedings of the 9th Conference on Robot Learning , series =

    Wagenmaker, Andrew and Zhang, Yunchu and Nakamoto, Mitsuhiko and Park, Seohong and Yagoub, Waleed and Nagabandi, Anusha and Gupta, Abhishek and Levine, Sergey , title =. Proceedings of the 9th Conference on Robot Learning , series =. 2025 , url =

    2025

  21. [21]

    Advances in Neural Information Processing Systems , year =

    Zhang, Tonghe and Yu, Chao and Su, Sichang and Wang, Yu , title =. Advances in Neural Information Processing Systems , year =. 2505.22094 , url =

    arXiv

  22. [22]

    2026 , url =

    He Zhang and Ying Sun and Hui Xiong , booktitle =. 2026 , url =

    2026

  23. [23]

    arXiv preprint arXiv:2602.01789 , year =

    Su, Entong and Westenbroek, Tyler and Nagabandi, Anusha and Gupta, Abhishek , title =. arXiv preprint arXiv:2602.01789 , year =. 2602.01789 , archivePrefix =

    arXiv

  24. [24]

    arXiv preprint arXiv:2508.01622 , year =

    Zhai, Xuanran and Zhao, Qianyou and Yu, Qiaojun and Hao, Ce , title =. arXiv preprint arXiv:2508.01622 , year =. 2508.01622 , archivePrefix =

    arXiv

  25. [25]

    and Lidard, Justin and Ankile, Lars L

    Ren, Allen Z. and Lidard, Justin and Ankile, Lars L. and Simeonov, Anthony and Agrawal, Pulkit and Majumdar, Anirudha and Burchfiel, Benjamin and Dai, Hongkai and Simchowitz, Max , title =. arXiv preprint arXiv:2409.00588 , year =. 2409.00588 , archivePrefix =

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2502.02538 , year =

    Park, Seohong and Li, Qiyang and Levine, Sergey , title =. arXiv preprint arXiv:2502.02538 , year =. 2502.02538 , archivePrefix =

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2502.09611 , year =

    Issachar, Noam and Salama, Mohammad and Fattal, Raanan and Benaim, Sagie , title =. arXiv preprint arXiv:2502.09611 , year =. 2502.09611 , archivePrefix =

    Pith/arXiv arXiv

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Luo, Gaoxiang and Cole, Frank and Zhang, Sihang and Wan, Yuxiang and Lu, Yulong and Sun, Ju , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    2026

  29. [29]

    arXiv preprint arXiv:2503.02353 , year =

    Wang, Luobin and Yu, Hongzhan and Yu, Chenning and Gao, Sicun and Christensen, Henrik , title =. arXiv preprint arXiv:2503.02353 , year =. 2503.02353 , archivePrefix =

    arXiv

  30. [30]

    arXiv preprint arXiv:2004.07219 , year =

    Fu, Justin and Kumar, Aviral and Nachum, Ofir and Tucker, George and Levine, Sergey , title =. arXiv preprint arXiv:2004.07219 , year =. 2004.07219 , archivePrefix =

    Pith/arXiv arXiv 2004

  31. [31]

    arXiv preprint arXiv:1707.06347 , year =

    Schulman, John and Wolski, Filip and Dhariwal, Prafulla and Radford, Alec and Klimov, Oleg , title =. arXiv preprint arXiv:1707.06347 , year =. 1707.06347 , archivePrefix =

    Pith/arXiv arXiv

  32. [32]

    Forty-second International Conference on Machine Learning , year =

    Efficient Skill Discovery via Regret-Aware Optimization , author =. Forty-second International Conference on Machine Learning , year =

  33. [33]

    , title =

    Rubinstein, Reuven Y. , title =. Methodology and Computing in Applied Probability , volume =. 1999 , doi =

    1999

  34. [34]

    , title =

    Bishop, Christopher M. , title =

  35. [35]

    and Jordan, Michael I

    Jacobs, Robert A. and Jordan, Michael I. and Nowlan, Steven J. and Hinton, Geoffrey E. , title =. Neural Computation , volume =. 1991 , doi =

    1991

  36. [36]

    and Welling, Max , title =

    Kingma, Diederik P. and Welling, Max , title =. International Conference on Learning Representations , year =