Pith. sign in

REVIEW 3 major objections 5 minor 47 references

EASE: Embodied Active Event Perception via Self-Supervised Energy Minimization

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read EASE claims that a single intrinsic signal—prediction error—can unify event segmentation, tracking, and summarization for an embodied agent.

desk verdict Useful integration of prediction-error segmentation and DQN active tracking, but the unvalidated salience-by-surprise premise keeps it from being fully convincing. read the letter →

arxiv 2506.17516 v1 pith:ZMO7ELG5 submitted 2025-06-20 cs.RO cs.CV

Zhou Chen , Sanjoy Kundu , Harsimran S. Baweja , Sathyanarayanan N. Aakur This is my paper · ORCID
classification cs.ROcs.CV
keywords eventperceptionself-supervisedlearningfreeenergyminimizationactiveinferencetrackingsegmentationembodiedAIpredictionerror
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

This paper proposes EASE, a framework that lets an embodied agent learn to detect, track, and summarize events in real time using only its own prediction errors, with no annotated data or external rewards. The central claim is that free energy minimization, instantiated as minimizing the gap between predicted and observed feature maps, is enough to drive both perception and control: the agent steers toward regions where its predictions fail, and treats peaks in the entropy of prediction errors as event boundaries. If this works, robots could acquire event perception in novel, unscripted environments without task-specific supervision, while retaining only sparse event summaries instead of raw video, which would reduce privacy risk. The paper reports simulation and real-world experiments showing that EASE tracks actors, segments their actions, and produces summaries, with behaviors such as backing up when a person gets too close emerging without being programmed.

What carries the argument

The load-bearing object is the spatial uncertainty distribution $\alpha_{ij}=\operatorname{Softmax}(\|f_{t,ij}-\hat{f}_{t,ij}\|^2/\tau)$, a softmax over per-location prediction errors on the CNN feature map. It carries the argument in three ways: it defines the reward for the DQN policy (centering the frame on its argmax $u_t$), it provides the input state $s_t=\{f_t,\alpha_{ij}\}$ for the policy, and—through its entropy over normalized errors ($H_t=-\sum_i p_{t,i}\log p_{t,i}$)—it supplies the event-boundary signal. The paper treats this 'surprise map' as a proxy for free energy, so that perception and control are both minimizing the same quantity.

What would settle it

Show EASE a scene containing a slowly moving person and a static, high-texture object such as a patterned curtain or bookshelf. If the agent consistently tracks or fixates on the static object—whose texture produces persistently high prediction error—while losing the person, then the equation 'maximum prediction error = salient actor' fails. A quantitative version: compute the fraction of frames in which the argmax of $\alpha_{ij}$ lies on a distractor rather than the human target, and compare against a random baseline; chance-level performance would falsify the salience-by-surprise premise.

Watch

Extended reading notes

Core claim

EASE's central claim is that a single intrinsic signal—spatiotemporal prediction error—can unify event segmentation, summarization, and active tracking in an embodied agent. The perception module, a recurrent generative model over visual features, predicts the next feature map and computes a spatial uncertainty distribution $\alpha_{ij}=\operatorname{Softmax}(\|f_{t,ij}-\hat{f}_{t,ij}\|^2/\tau)$. The motor module, a DQN with reward $r_t=-\|c_t-u_t\|$ where $u_t$ is the argmax of $\alpha_{ij}$, learns to center the frame on the most surprising region. Event boundaries are detected as peaks in the entropy of normalized prediction errors over a sliding window, and summary frames are the lowest-prediction-error frames within each segment. On the paper's account, these mechanisms operationalize free energy minimization without explicit latent-state inference, and the experiments show the resulting agent tracks salient actors, maintains target continuity, and summarizes events, with emergent behaviors like moving backward when the target is too close.

Load-bearing premise

The framework assumes that the location of maximum prediction error is the salient target worth tracking, and that event boundaries coincide with peaks in the entropy of normalized prediction errors; the authors note that when this assumption fails, the robot fixates on stairs or furniture instead of the person.

Editorial extensions

If this is right

  • Embodied agents can bootstrap event perception from scratch in novel environments, since prediction error and entropy are available without any annotation or reward engineering.
  • Streaming summarization that keeps only low-error keyframes and discards raw video becomes a viable privacy-preserving design for assistive and surveillance robots.
  • The same intrinsic signals can drive multiple tasks—tracking, segmentation, summarization—so a single self-supervised objective can replace separate supervised pipelines.
  • Emergent behaviors such as backward stepping and target continuity suggest that action policies for tracking can arise from perception-action coupling rather than explicit task specification.
  • Training in simulation transfers to a real robot, indicating the approach may generalize across domains without fine-tuning on real data.

Reading between the lines

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

  • The salience-by-surprise premise could be stress-tested against human-annotated salience benchmarks or standard visual attention models; if prediction-error maxima diverge from human gaze, the framework would need a learned salience prior to remain viable for human-centered tasks.
  • The entropy-peak boundary detector is a form of unsupervised change-point detection; comparing it against standard statistical change-point methods on the same feature stream would isolate how much of the segmentation performance comes from the predictive-coding representation versus the peak-picking rule.
  • The privacy guarantee is only as strong as what the retained snapshots and Q-value activations leak; a formal information-flow analysis of the summary frames would be needed to back the privacy claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes EASE, a self-supervised framework for embodied active event perception that couples a predictive-coding-based perception module with a DQN-based motor control module. The perception module computes prediction errors and an uncertainty distribution α_ij over spatial locations, and the control module selects actions to center the camera on the maximum-uncertainty location. Prediction-error entropy is used to detect event boundaries, and frames with minimal prediction error are selected as summaries. The authors evaluate EASE in UnrealCV simulation and on a real LoCoBot platform, reporting quantitative results for tracking, segmentation, and summarization, and claiming privacy-preserving, annotation-free operation with emergent behaviors such as implicit memory and target continuity.

Significance. If the central claims are sustained, the paper would make a useful contribution by demonstrating that a single self-supervised predictive-coding objective can drive active tracking, event segmentation, and streaming summarization without external rewards or annotations. The real-world deployment and the qualitative emergent behaviors (backing away from a close target, retaining the original target when a new actor enters) are interesting and potentially valuable for human-robot interaction and privacy-sensitive settings. However, the paper's significance depends on closing the gap between the free-energy formulation and the implemented learning objective, and on validating the salience-by-surprise premise, which the current evidence only partially supports.

major comments (3)
  1. [§III-A, Eq. (1)–(3)] The paper defines free energy minimization in Eq. (1) as an argmin over actions of a global prediction error plus a λ-weighted uncertainty term, but the implemented motor control is a DQN trained with the temporal difference loss in Eq. (3) on the reward r_t = -||c_t - u_t||, where u_t is the location of maximum α_ij. No argument or proof is given that maximizing this reward is equivalent to, or optimizes, the second term of Eq. (1), nor that the TD-optimal policy minimizes Eq. (1). Because the free-energy framing is the paper's central theoretical contribution, this gap must be closed with a derivation or an explicit relaxation statement that says what exactly the DQN optimizes and how it relates to Eq. (1).
  2. [§III-A.2 and §V-C] The entire task signal is the salience-by-surprise premise: the uncertainty distribution α_ij in Eq. (2) is a softmax over per-location prediction errors, and the DQN reward centers the camera on its argmax. The paper's own qualitative analysis in Section V-C concedes that in scenarios with subtle or limited target motion, EASE 'occasionally prioritizes novel elements in the environment, such as stairs or furniture, over the primary target.' There is no quantitative measurement of how often the argmax of α_ij coincides with the ground-truth target, no comparison against an independent salience benchmark, and no analysis of failure conditions. Since tracking, segmentation, and summarization all inherit this premise, the claim that EASE actively tracks salient actors is not yet supported without such validation.
  3. [§V-B, Table III] Real-world event segmentation is evaluated only against the three EASE variants; no comparison is made to prior self-supervised or weakly-supervised event segmentation methods, and no variance or inter-annotator agreement is reported for the human annotations. Given that the strict F1 score for EASE is 22.47 and for EASE-Hybrid is 32.06, it is unclear whether the method is competitive with existing segmentation approaches or merely internally consistent. Reporting statistical significance and external baselines is necessary to support the generality claim.
minor comments (5)
  1. [Table II] The abbreviations AR and AL are used in Table II but are not defined in the text or caption; please define them as average reward and average episode length, respectively.
  2. [§III-B, Eq. (4)] The sliding window size N appears in Eq. (4) but is only justified later in a sentence after Table I; please define N explicitly in the equation and clarify whether N=30 is a tuned hyperparameter or a fixed design choice.
  3. [§III-A] The statement that λ is 'a design choice rather than a tunable hyperparameter' is confusing because λ is then fixed to 1 with no sensitivity analysis; please clarify the role of λ and whether the reported results depend on this choice.
  4. [§V-C, Fig. 2] The text refers to 'Figure V-B' but the displayed figure is labeled 'Fig. 2'; please correct the cross-reference.
  5. [References] Reference [42] cites 'J. Chung' for the DQN, but the original DQN paper is by Mnih et al.; please correct the citation and verify the reference list for other formatting inaccuracies.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EASE's core equations are defined in-paper and benchmarked against external annotations and baselines; the salience-by-surprise premise is an assumption, not a circular derivation.

full rationale

EASE's derivation chain is not circular. The core objective in Eq. 1 and the uncertainty distribution in Eq. 2 are defined in the paper, and the DQN reward r_t = -||c_t - u_t||, where u_t is the argmax of alpha_ij, is a design choice that couples perception to action rather than a fit to the evaluation target. Event boundaries are computed from the entropy of prediction errors in Eqs. 4-5 and then evaluated against human-annotated boundaries and external baselines in Tables I-III; summarization keyframes in Eq. 6 are assessed by human annotators. Thus the central claims are tested against external evidence. The salience-by-surprise premise in Eq. 2 is load-bearing, because tracking success depends on whether the argmax of prediction error coincides with the intended actor, and the paper itself concedes in Section V-C that 'in scenarios with subtle or limited target motion, EASE occasionally prioritizes novel elements in the environment, such as stairs or furniture, over the primary target.' This is an unvalidated assumption and a robustness limitation, but it is not circularity by construction: no equation reduces to its own input, and no fitted parameter is renamed as a prediction. The authors cite their own prior work ([4], [5], [6], [9]) as related work on self-supervised event segmentation and active vision, but those citations are not used to justify EASE's equations or to forbid alternative formulations, so they are not load-bearing. Overall, the paper is self-contained against external benchmarks, and the admitted failure mode weakens the empirical claim without making the derivation circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the modeling assumption that prediction-error entropy marks event boundaries and that prediction-error magnitude marks salience. These assumptions are untested against independent salience or boundary benchmarks. The framework also introduces several free choices (tau, lambda, N) that are not rigorously ablated.

free parameters (3)
  • Temperature tau in uncertainty softmax (Eq. 2) = not specified
    Controls sensitivity of the uncertainty distribution alpha_ij to prediction errors; no value or search procedure is given.
  • Balance weight lambda in free-energy objective (Eq. 1) = 1
    Fixed to 1 as a 'design choice'; the paper argues intermediate values lack interpretation, but lambda is still a free choice affecting the objective.
  • Sliding window size N for entropy-based boundary detection (Eq. 4) = 30
    Set empirically and claimed to 'work best' with reference to Table I, which does not show a window-size ablation.
assumptions (4)
  • domain assumption Free energy can be approximated by squared L2 prediction error with unit precision (identity covariance)
    Invoked in Section III-A when defining LE = ||f_{t+1} - hat f_{t+1}||^2, ignoring precision weighting that classical free energy formulations include.
  • domain assumption Event boundaries coincide with peaks in the entropy of normalized prediction errors within a sliding window
    Adopted in Section III-B, Eqs. (4)-(5), without independent evidence that perceptual event boundaries align with prediction-error entropy peaks.
  • domain assumption The region of maximum spatial prediction error corresponds to the salient target to track
    Used in the DQN reward r_t = -||c_t - u_t||, where u_t is the argmax of alpha_ij; the paper acknowledges this can fail with salient distractors like stairs or furniture (Section V-C).
  • domain assumption A fixed discrete action space of seven motion primitives is sufficient for active tracking
    The DQN selects from a predefined set of actions as described in Section III-A; the paper does not justify that this action space suffices for general event perception.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EASE: Embodied Active Event Perception via Self-Supervised Energy Minimization." pith.science (2026). https://pith.science/paper/ZMO7ELG5

@misc{pith2026250617516,
  author       = {Pith},
  title        = {Pith review of: EASE: Embodied Active Event Perception via Self-Supervised Energy Minimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZMO7ELG5}},
  note         = {Machine review of arXiv:2506.17516}
}
read the original abstract

Active event perception, the ability to dynamically detect, track, and summarize events in real time, is essential for embodied intelligence in tasks such as human-AI collaboration, assistive robotics, and autonomous navigation. However, existing approaches often depend on predefined action spaces, annotated datasets, and extrinsic rewards, limiting their adaptability and scalability in dynamic, real-world scenarios. Inspired by cognitive theories of event perception and predictive coding, we propose EASE, a self-supervised framework that unifies spatiotemporal representation learning and embodied control through free energy minimization. EASE leverages prediction errors and entropy as intrinsic signals to segment events, summarize observations, and actively track salient actors, operating without explicit annotations or external rewards. By coupling a generative perception model with an action-driven control policy, EASE dynamically aligns predictions with observations, enabling emergent behaviors such as implicit memory, target continuity, and adaptability to novel environments. Extensive evaluations in simulation and real-world settings demonstrate EASE's ability to achieve privacy-preserving and scalable event perception, providing a robust foundation for embodied systems in unscripted, dynamic tasks.

Figures

Figures reproduced from arXiv: 2506.17516 by the authors.

Figure 1
Figure 1. Overview. The perception module predicts future observations (ot+1) and minimizes discrepancies (LE), while the motor control module selects actions (at), choosing from discrete motion primitives (forward, backward, stop, turn left, turn right, forward-left, and forward-right) to reduce control loss (LQ). Free Energy Minimization supervises both modules, enabling robust active event perception. storing personally id… view at source ↗
Figure 2
Figure 2. Qualitative visualization of emergent properties from free energy minimization. Top: EASE learns to move back when too close to the target. Middle: When a new actor is introduced, EASE keeps track of the old target until movement becomes predictable, reacts to the new target, and moves back to keep both actors in line of sight. Bottom: Entropy values and the summarization output (from robot POV) of the target perfor… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages

  1. [1]

    Learning hierar- chical world models with adaptive temporal abstractions from discrete latent dynamics,

    C. Gumbsch, N. Sajid, G. Martius, and M. V . Butz, “Learning hierar- chical world models with adaptive temporal abstractions from discrete latent dynamics,” inThe Twelfth International Conference on Learning Representations, 2023

    work page 2023

  2. [2]

    Event structure in perception and concep- tion

    J. M. Zacks and B. Tversky, “Event structure in perception and concep- tion.”Psychological bulletin, vol. 127, no. 1, p. 3, 2001. 8 IEEE ROBOTICS AND AUTOMATION LETTERS. PREPRINT VERSION. JUNE, 2025

    work page 2001

  3. [3]

    Emergence of sensory attenuation based upon the free-energy principle,

    H. Idei, W. Ohata, Y . Yamashita, T. Ogata, and J. Tani, “Emergence of sensory attenuation based upon the free-energy principle,”Scientific reports, vol. 12, no. 1, p. 14542, 2022

    work page 2022

  4. [4]

    A perceptual prediction framework for self supervised event segmentation,

    S. N. Aakur and S. Sarkar, “A perceptual prediction framework for self supervised event segmentation,”2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1197–1206, 2018

    work page 2019

  5. [5]

    Action localization through continual predictive learning,

    ——, “Action localization through continual predictive learning,” in European Conference on Computer Vision, 2020

    work page 2020

  6. [6]

    Self-supervised multi-actor social activity understanding in streaming videos,

    S. Trehan and S. N. Aakur, “Self-supervised multi-actor social activity understanding in streaming videos,” inInternational Conference on Pattern Recognition. Springer, 2025, pp. 293–309

    work page 2025

  7. [7]

    AD-V AT: An asymmetric dueling mechanism for learning visual active tracking,

    F. Zhong, P. Sun, W. Luo, T. Yan, and Y . Wang, “AD-V AT: An asymmetric dueling mechanism for learning visual active tracking,” in International Conference on Learning Representations, 2019. [Online]. Available: https://openreview.net/forum?id=HkgYmhR9KX

    work page 2019

  8. [8]

    Ad-vat+: An asymmetric dueling mechanism for learning and understanding visual active tracking,

    ——, “Ad-vat+: An asymmetric dueling mechanism for learning and understanding visual active tracking,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 5, pp. 1467–1482, 2021

    work page 2021

Show all 47 references
  1. [9]

    Towards active vision for action localization with reactive control and predictive learning,

    S. Trehan and S. N. Aakur, “Towards active vision for action localization with reactive control and predictive learning,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 783–792

    2022

  2. [10]

    An improved method based on deep reinforcement learning for target searching,

    X. Wei, X. Huang, T. Lu, and G. Song, “An improved method based on deep reinforcement learning for target searching,”2019 4th International Conference on Robotics and Automation Engineering (ICRAE), pp. 130– 134, 2019

    2019

  3. [11]

    Training an agent to find and reach an object in different environments using visual reinforcement learning and transfer learning,

    E. C. S. Batista, W. Caarls, L. Forero, and M. A. C. Pacheco, “Training an agent to find and reach an object in different environments using visual reinforcement learning and transfer learning,” inInternational Conference on Agents and Artificial Intelligence, 2021

    2021

  4. [12]

    End-to-end active object tracking via reinforcement learning,

    W. Luo, P. Sun, Y . Mu, and W. Liu, “End-to-end active object tracking via reinforcement learning,” inInternational Conference on Machine Learning, 2017. [Online]. Available: https://api.semanticscholar.org/ CorpusID:30180031

    2017

  5. [13]

    Towards distraction- robust active visual tracking,

    F. Zhong, P. Sun, W. Luo, T. Yan, and Y . Wang, “Towards distraction- robust active visual tracking,”ArXiv, vol. abs/2106.10110, 2021

    2021 arXiv

  6. [14]

    Pose-assisted multi-camera collaboration for active object tracking,

    J. Li, J. Xu, F. Zhong, X. Kong, Y . Qiao, and Y . Wang, “Pose-assisted multi-camera collaboration for active object tracking,”ArXiv, vol. abs/2001.05161, 2020. [Online]. Available: https://api.semanticscholar. org/CorpusID:210701739

    2001 arXiv

  7. [15]

    Coordinate- aligned multi-camera collaboration for active multi-object tracking,

    Z. Fang, J. Zhao, M. Yang, W. gang Zhou, Z. Lu, and H. Li, “Coordinate- aligned multi-camera collaboration for active multi-object tracking,” ArXiv, vol. abs/2202.10881, 2022

    2022 arXiv

  8. [16]

    Ad-vat: An asymmetric dueling mechanism for learning visual active tracking,

    F. Zhong, P. Sun, W. Luo, T. Yan, and Y . Wang, “Ad-vat: An asymmetric dueling mechanism for learning visual active tracking,” in International Conference on Learning Representations, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:53533716

    2018

  9. [17]

    Pointly-supervised action localization,

    P. Mettes and C. G. M. Snoek, “Pointly-supervised action localization,” International Journal of Computer Vision, vol. 127, pp. 263 – 281, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:44063701

    2018

  10. [18]

    Weakly supervised action localization by sparse temporal pooling network,

    P. Nguyen, T. Liu, G. Prasad, and B. Han, “Weakly supervised action localization by sparse temporal pooling network,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  11. [19]

    Weakly-supervised action localization with background modeling,

    P. X. Nguyen, D. Ramanan, and C. C. Fowlkes, “Weakly-supervised action localization with background modeling,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oc- tober 2019

    2019

  12. [20]

    Weakly-supervised action localiza- tion by generative attention modeling,

    B. Shi, Q. Dai, Y . Mu, and J. Wang, “Weakly-supervised action localiza- tion by generative attention modeling,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  13. [21]

    Weakly- supervised temporal action localization with multi-modal plateau trans- formers,

    X. Hu, K. Li, D. Patel, E. Kruus, M. R. Min, and Z. Ding, “Weakly- supervised temporal action localization with multi-modal plateau trans- formers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2024, pp. 2704–2713

    2024

  14. [22]

    Can mllms guide weakly-supervised temporal action localization tasks?

    Q. Zhang and Y . Qi, “Can mllms guide weakly-supervised temporal action localization tasks?”arXiv preprint arXiv:2411.08466, 2024

    2024 arXiv

  15. [23]

    Spatio-temporal event segmentation for wildlife extended videos,

    R. Mounir, R. Gula, J. Theuerkauf, and S. Sarkar, “Spatio-temporal event segmentation for wildlife extended videos,” inInternational Conference on Computer Vision and Image Processing, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:251286809

    2021

  16. [24]

    Unsupervised gaze prediction in egocentric videos by energy-based surprise modeling,

    S. N. Aakur and A. Bagavathi, “Unsupervised gaze prediction in egocentric videos by energy-based surprise modeling,”arXiv preprint arXiv:2001.11580, 2020

    2001 arXiv

  17. [25]

    Unifying short and long- term tracking with graph hierarchies,

    O. Cetintas, G. Bras ´o, and L. Leal-Taix ´e, “Unifying short and long- term tracking with graph hierarchies,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 877–22 887

    2023

  18. [26]

    Colorization as a proxy task for visual understanding,

    G. Larsson, M. Maire, and G. Shakhnarovich, “Colorization as a proxy task for visual understanding,”2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 840–849, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:12369376

    2017

  19. [27]

    Learning image represen- tations by completing damaged jigsaw puzzles,

    D. Kim, D. Cho, D. Yoo, and I. S. Kweon, “Learning image represen- tations by completing damaged jigsaw puzzles,” in2018 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE, 2018, pp. 793–802

    2018

  20. [28]

    Self-supervised video representation learning by pace prediction,

    J. Wang, J. Jiao, and Y .-H. Liu, “Self-supervised video representation learning by pace prediction,” inEuropean Conference on Computer Vision 2020. Springer, 2020, pp. 504–521

    2020

  21. [29]

    Video representation learning with visual tempo consistency,

    C. Yang, Y . Xu, B. Dai, and B. Zhou, “Video representation learning with visual tempo consistency,”arXiv preprint arXiv:2006.15489, 2020

    2006 arXiv

  22. [30]

    Tclr: Temporal con- trastive learning for video representation,

    I. Dave, R. Gupta, M. N. Rizve, and M. Shah, “Tclr: Temporal con- trastive learning for video representation,”Computer Vision and Image Understanding, vol. 219, p. 103406, 2022

    2022

  23. [31]

    Video contrastive learning with global context,

    H. Kuang, Y . Zhu, Z. Zhang, X. Li, J. Tighe, S. Schwertfeger, C. Stach- niss, and M. Li, “Video contrastive learning with global context,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3195–3204

    2021

  24. [32]

    Masked motion encoding for self-supervised video representation learning,

    X. Sun, P. Chen, L. Chen, C. Li, T. H. Li, M. Tan, and C. Gan, “Masked motion encoding for self-supervised video representation learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2235–2245

    2023

  25. [33]

    The free-energy principle: a unified brain theory?

    K. Friston, “The free-energy principle: a unified brain theory?”Nature reviews neuroscience, vol. 11, no. 2, pp. 127–138, 2010

    2010

  26. [34]

    Active inference and agency: optimal control without cost functions,

    K. Friston, S. Samothrakis, and R. Montague, “Active inference and agency: optimal control without cost functions,”Biological cybernetics, vol. 106, pp. 523–541, 2012

    2012

  27. [35]

    An empirical study of active inference on a humanoid robot,

    G. Oliver, P. Lanillos, and G. Cheng, “An empirical study of active inference on a humanoid robot,”IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 2, pp. 462–471, 2021

    2021

  28. [36]

    Active inference in robotics and artificial agents: Survey and challenges,

    P. Lanillos, C. Meo, C. Pezzato, A. A. Meera, M. Baioumy, W. Ohata, A. Tschantz, B. Millidge, M. Wisse, C. L. Buckleyet al., “Active inference in robotics and artificial agents: Survey and challenges,”arXiv preprint arXiv:2112.01871, 2021

    2021 arXiv

  29. [37]

    Hierarchical active inference: a theory of motivated control,

    G. Pezzulo, F. Rigoli, and K. J. Friston, “Hierarchical active inference: a theory of motivated control,”Trends in cognitive sciences, vol. 22, no. 4, pp. 294–306, 2018

    2018

  30. [38]

    T. Parr, G. Pezzulo, and K. J. Friston,Active inference: the free energy principle in mind, brain, and behavior. MIT Press, 2022

    2022

  31. [39]

    The free energy principle for perception and action: A deep learning perspective,

    P. Mazzaglia, T. Verbelen, O. C ¸ atal, and B. Dhoedt, “The free energy principle for perception and action: A deep learning perspective,” Entropy, vol. 24, no. 2, p. 301, 2022

    2022

  32. [40]

    Real-world robot control based on contrastive deep active inference with demonstrations,

    K. Fujii, T. Isomura, and S. Murata, “Real-world robot control based on contrastive deep active inference with demonstrations,”IEEE Access, 2024

    2024

  33. [41]

    Incremental learning of goal- directed actions in a dynamic environment by a robot using active inference,

    T. Matsumoto, W. Ohata, and J. Tani, “Incremental learning of goal- directed actions in a dynamic environment by a robot using active inference,”Entropy, vol. 25, no. 11, p. 1506, 2023

    2023

  34. [42]

    Playing atari with deep reinforcement learning,

    J. Chung, “Playing atari with deep reinforcement learning,”Comput. Ence, vol. 21, pp. 351–362, 2013

    2013

  35. [43]

    Stable-baselines3: Reliable reinforcement learning implementa- tions,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dor- mann, “Stable-baselines3: Reliable reinforcement learning implementa- tions,”Journal of Machine Learning Research, vol. 22, no. 268, pp. 1–8, 2021

    2021

  36. [44]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  37. [45]

    Unrealcv: Virtual worlds for computer vision,

    W. Qiu, F. Zhong, Y . Zhang, S. Qiao, Z. Xiao, T. S. Kim, and Y . Wang, “Unrealcv: Virtual worlds for computer vision,” inProceedings of the 25th ACM international conference on Multimedia, 2017, pp. 1221– 1224

    2017

  38. [46]

    Generic event boundary detection: A benchmark for event segmen- tation,

    M. Z. Shou, S. W. Lei, W. Wang, D. Ghadiyaram, and M. Feiszli, “Generic event boundary detection: A benchmark for event segmen- tation,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 8075–8084

    2021

  39. [47]

    End-to-end active object tracking and its real-world deployment via reinforcement learning,

    W. Luo, P. Sun, F. Zhong, W. Liu, T. Zhang, and Y . Wang, “End-to-end active object tracking and its real-world deployment via reinforcement learning,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 42, no. 6, pp. 1317–1332, 2019

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.