Towards Learning How to Properly Play UNO with the iCub Robot

Alessandra Sciutti; Pablo Barros; Stefan Wermter

REVIEW 2 major objections 5 minor 22 references

Towards Learning How to Properly Play UNO with the iCub Robot

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A proposed roadmap would give an iCub robot proper social responses in a multiplayer UNO game by learning from players' affective reactions.

desk verdict A clearly written roadmap for combining affective memory and actor-critic RL on an iCub UNO player, but with no implementation and a load-bearing assumption about legible in-game affect that the authors never test. read the letter →

arxiv 1908.00744 v1 pith:CZRLYKUW submitted 2019-08-02 cs.HC cs.AIcs.RO

Pablo Barros , Stefan Wermter , Alessandra Sciutti This is my paper · ORCID

classification cs.HCcs.AIcs.RO

keywords Human-RobotInteractionEmotionRecognitionAffectiveResponseMemoryReinforcementLearningiCubMultimodalPerceptionSocialRobotics

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 position paper proposes a research roadmap for giving the iCub humanoid robot the ability to respond “properly” in a social setting, using a competitive multi-person UNO game as the testbed. The paper argues that proper social responses can be learned by combining two mechanisms: an adaptive perception module that reads each player's facial, vocal, and bodily emotion expression and builds a personalized affective memory, and an actor-critic reinforcement learning module that chooses both game actions and affective expressions. The reward for each action comes from the observed affective reactions of the players and the change in game state, so the robot learns what counts as proper in context. If the roadmap is feasible, a companion robot could adapt to individual people, read the mood of a group, and choose behavior that is judged engaging, human-like, or winning depending on the goal.

What carries the argument

The load-bearing object is the affective memory: a Grow-When-Required (GWR) self-organizing network that learns prototype emotional concepts from each person in an online, unsupervised way, so the robot can adapt to how a specific individual expresses emotions without requiring a large labeled dataset of that person. Around it, the architecture adds gated recurrent units to compress asynchronous multimodal inputs, local and global attention to keep intense emotional changes from being averaged away, recurrent gamma connections to merge individual memories into a group affective state, and an actor-critic reinforcement learning network whose reward is derived from the observed affective reactions of the players. The paper also invokes predictive stimulation and a replay memory pool to let the actor-critic learn from the few data points a real interaction provides.

What would settle it

Record a set of real UNO games between humans under the same setup and have independent annotators mark the visible facial, vocal, and bodily affect per turn; if the majority of turns contain no recognizable affective expression, or if a real-time emotion recognizer on the iCub performs at chance on those recordings, the proposed loop lacks the input it needs and the roadmap's central premise fails.

Watch

Extended reading notes

Core claim

The central claim is that the action-perception cycle underlying natural human interaction—perceiving emotion, choosing a response, and reading the social impact of that response—can be instantiated in a real-time robot architecture and trained without thousands of human interactions. The paper's proposed hybrid framework uses deep convolutional channels to represent faces, bodies, and voice, and a self-organizing GWR network to learn, online, the idiosyncratic emotion expressions of each individual person; those individual affective memories are then merged into a group emotional read. An actor-critic reinforcement learner maps the perceived group state plus the relevant affective memory to an action, and the players' reactions serve as the reward signal that determines whether the action was proper. The authors describe this as a roadmap rather than a completed system, and they define an evaluation protocol with offline simulation plus repeated real-world sessions with shuffled participant groups, using objective measures and HRI questionnaires to assess the robot's acceptance.

Load-bearing premise

The whole perception-action loop depends on the players naturally and legibly showing their emotions during the game, and on the robot's sensors capturing those expressions in real time; if the emotional signal is weak or missed, the reward that teaches the robot what is proper never arrives.

Editorial extensions

If this is right

If the roadmap works, a robot can keep playing a competitive game against several humans at once while tracking each individual's emotional state and the group's overall mood in real time.
The robot's behavior could be tuned to three distinct goals—winning, keeping players positively engaged, or balancing both to appear human-like—simply by changing the reward function.
Because the affective memory adapts online, the same architecture could carry over to novel people without retraining the deep perception channels.
The evaluation protocol would provide quantitative evidence on whether affective-feedback-driven reinforcement learning yields socially accepted behavior in a multi-person scenario.

Reading between the lines

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

The UNO setting is a natural testbed for group affect because each turn creates a discrete, timed emotional event; the same reward design could transfer to turn-taking interactions like tutoring or therapy sessions, where a robot must read a learner's frustration or engagement.
A testable extension would be to vary the number of players and the game's randomness; if the robot's learned strategies depend strongly on game context rather than on stable player traits, the affective memory may be doing less work than the framework assumes.
The paper's reliance on natural affective expression suggests a clear baseline experiment: record human players in the same UNO setup and measure how often, and how legibly, they actually display facial and vocal affect; the result would set an upper bound on what the perception module can exploit.
If natural expressions prove too sparse, the roadmap could be salvaged by seeding the affective memory with acted or crowdsourced emotion data, but that would change the claim from learning proper responses to learning from scripted ones.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

This is a short position paper, not a research report. It lays out a plan to combine the authors' affective memory work with actor-critic reinforcement learning on an iCub playing UNO with three people. There is no system, no data, no results. Read it as a proposal, not a contribution.

The UNO scenario is actually a thoughtful choice: competitive, turn-based, rich in interpersonal reactions, and a natural source of repeated decision points. The paper is honest about being a roadmap. It also identifies a real limitation of deep emotion recognition — adapting to individual expressiveness without a large batch of personal data — and the proposed fix, GWR-based affective memories, is a legitimate extension of the authors' prior work. The evaluation plan is more concrete than most position papers: offline simulation plus longitudinal real-world sessions, with Asch and Godspeed questionnaires on top of objective measures.

The soft spot is exactly the one flagged in the stress-test note. The entire perception-action loop depends on reading spontaneous, temporally fine-grained affective reactions from face, body, and voice in real time, and using that read as a reward signal. The paper explicitly assumes natural affective behavior will emerge in the UNO setup (Section II) but offers no pilot data, no signal-to-noise estimate, and no discussion of how players may mask or feign emotions during a competitive card game. Positive group affect does not obviously equal a 'proper' response: a socially smooth losing move and an awkward winning move could both be mislabeled by that proxy. For a proposal this is the main empirical unknown, not an internal contradiction, but the authors do not acknowledge it as a central risk.

Minor quibble: the paper leans on 'we will investigate' throughout, which is fine for a roadmap, but it means there is little to verify on the page. The three target behaviors (win, be enjoyable, be human-like) are plausible but vaguely specified.

For a reader who wants a concrete example of how affective memory and RL might be posed for an HRI setting, this is a useful two-page sketch. It will not change anyone's research program. It is a workshop-level position paper, not a full archival contribution. I would not send it to a serious referee for a main track, but if it crosses your desk as a workshop submission, a quick read is enough; the authors have earned the benefit of the doubt from their earlier work.

Referee Report

2 major / 5 minor

Summary. This paper is a position/roadmap paper that proposes a research plan for enabling an iCub robot to play the card game UNO with three human players while providing "proper" social responses. The proposed framework has two main modules: an adaptive affective perception system based on growing self-organizing networks (GWRs), extended with temporal/attention mechanisms to integrate asynchronous multimodal stimuli and group-level affective states; and an actor-critic reinforcement learning module that selects game-related and affective actions, using the perceived affective reactions of the players as reward. The paper describes the UNO scenario, motivates it as a controllable and multi-turn context for HRI research, and outlines an evaluation protocol combining offline simulation/module-level tests with real-world sessions using questionnaires (Asch personality impression and Godspeed). No implemented system or experimental results are presented; the contribution is explicitly a roadmap rather than an empirical validation.

Significance. If realized, the proposed architecture would address a real gap in HRI: personalizing social perception to individual users and using spontaneous social feedback to shape robot behavior without large numbers of human-provided labels. The roadmap is coherently grounded in the authors' prior work on GWR-based affective memory, which is a genuine strength, and the chosen scenario (a competitive game with many turns) provides a rich, repeatable interaction context. The evaluation plan sensibly combines objective measures (emotion recognition accuracy, processing time) with subjective questionnaires. However, the central feasibility of the roadmap rests on an unvalidated assumption that players' affective reactions during a competitive card game are legible and usable as a reliable reward signal; the paper does not provide pilot data, signal-to-noise analysis, or a clear operational definition of "proper." These are not fatal flaws for a roadmap, but they are load-bearing points that need to be addressed before the proposed framework can be assessed as a sound research program.

major comments (2)

[II and III-B] The roadmap's central learning loop depends on the affective reactions of players being a reliable, distinguishable signal. Section II states the assumption that the UNO scenario will elicit natural affective behavior, and Section III-B makes the players' perceived reaction the reward that defines whether an action was proper. The paper provides no pilot data, annotator agreement, or signal-to-noise evidence for this assumption, and in a competitive card game players may mask or fake expressions as part of strategy. Since every learned behavior reduces to this reward channel, the roadmap should include a preliminary validation study of the affective reward signal (e.g., recordings with multiple raters or a baseline emotion-recognition evaluation) before the architecture can be judged feasible.
[III-B] The reward definition conflates "positive group affect" with "proper social response." A player may laugh politely at a robot's action while disliking it, and a socially smooth action may produce short-term positive affect but be strategically bad; conversely, a winning action may cause negative affect in opponents. Section III-B says a reward based on both game score and positive group affect will produce behavior "perceived as more human-like," but this is an unverified assumption and the paper does not specify how the multimodal perceived reaction is converted into a scalar reward, nor how the robot avoids exploiting the reward signal (e.g., by producing expressions that merely amuse the players). An operational definition of "proper," with concrete measurement, is needed.

minor comments (5)

[III-A] The text "A general attention mechanism will also important" is grammatically incomplete; rephrase to "will also be important."
[III-A] The phrase "the use of recurrent pooling layers will be used" is redundant and awkward; consider "we will integrate recurrent pooling layers that highlight...".
[IV] The sentence "objective measures such as ... the processing time will be maximized" is ambiguous and likely opposite to intent; processing time should be minimized or kept below a real-time bound, not maximized.
[II] The phrase "inverting the playing order" should be "reversing the turn order" or similar for clarity.
[IV] The evaluation plan does not state a target sample size, number of sessions, or a quantitative success criterion; adding these would make the roadmap more concrete and falsifiable.

Circularity Check

0 steps flagged · score 0.0 of 10

Roadmap paper with no derived predictions or fitted parameters; self-citations are motivators, not load-bearing reductions.

full rationale

This is an explicitly forward-looking position paper: it proposes a roadmap for a future hybrid affective-perception and reinforcement-learning architecture for the iCub playing UNO. There are no equations, no fitted parameters, no benchmark results that are 'predicted' from inputs, and no theorem or uniqueness claim. The prior works by the authors (Barros and Wermter on affective memory) are cited as existing components that the authors plan to build upon, but the central claim—that such a framework could empower the robot with proper social responses—is not derived from those papers. The affective memory results are described as evaluations on emotion recognition tasks, which are external to the proposed UNO scenario and are not used to define the target outcome by construction. The reward signal in Section III-B is proposed as a design choice, not as a quantity that the paper claims to derive from first principles. The feasibility assumption about natural affective expression is an empirical precondition, not a circular step. Therefore the paper exhibits no significant circularity.

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

The paper introduces no fitted parameters or invented entities because it only describes a plan. The central burden lies in four domain assumptions about the scenario and the proposed neural modules, none of which are backed by experiments in this paper.

assumptions (4)

domain assumption The UNO game scenario elicits natural affective expressions from participants.
Section II states the scenario 'will be an important characteristic to elicit natural affective behavior' without pilot data.
ad hoc to paper GWR-based affective memories can be extended to integrate asynchronous multimodal stimuli and group-level emotional states.
Section III-A proposes extending prior GWR affective memories with GRUs and attention, but this extension is not demonstrated.
domain assumption Reinforcement learning with a replay memory pool can learn socially appropriate actions from the limited interactions of a session.
Section III-B acknowledges sample complexity but proposes predictive stimulation as a remedy without evidence.
domain assumption Players' affective reactions provide a valid reward signal for what constitutes a 'proper' robot response.
Section III-B defines rewards based on maintaining positive group affective state, assuming this aligns with human-likeness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Learning How to Properly Play UNO with the iCub Robot." pith.science (2026). https://pith.science/paper/CZRLYKUW

@misc{pith2026190800744,
  author       = {Pith},
  title        = {Pith review of: Towards Learning How to Properly Play UNO with the iCub Robot},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZRLYKUW}},
  note         = {Machine review of arXiv:1908.00744}
}

read the original abstract

While interacting with another person, our reactions and behavior are much affected by the emotional changes within the temporal context of the interaction. Our intrinsic affective appraisal comprising perception, self-assessment, and the affective memories with similar social experiences will drive specific, and in most cases addressed as proper, reactions within the interaction. This paper proposes the roadmap for the development of multimodal research which aims to empower a robot with the capability to provide proper social responses in a Human-Robot Interaction (HRI) scenario.

Figures

Figures reproduced from arXiv: 1908.00744 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

[1]

Forming impressions of personality

Solomon E Asch. Forming impressions of personality. The Journal of Abnormal and Social Psychology , 41(3):258, 1946

work page 1946
[2]

Developing crossmodal expres- sion recognition based on a deep neural model

Pablo Barros and Stefan Wermter. Developing crossmodal expres- sion recognition based on a deep neural model. Adaptive behavior , 24(5):373–396, 2016

work page 2016
[3]

A self-organizing model for affec- tive memory

Pablo Barros and Stefan Wermter. A self-organizing model for affec- tive memory. In Neural Networks (IJCNN), 2017 International Joint Conference on, pages 31–38. IEEE, 2017

work page 2017
[4]

Measurement instruments for the anthropomorphism, animacy, likeabil- ity, perceived intelligence, and perceived safety of robots

Christoph Bartneck, Dana Kuli ´c, Elizabeth Croft, and Susana Zoghbi. Measurement instruments for the anthropomorphism, animacy, likeabil- ity, perceived intelligence, and perceived safety of robots. International journal of social robotics , 1(1):71–81, 2009

work page 2009
[5]

Human-robot collaboration: A survey

Balasubramaniyan Chandrasekaran and James M Conrad. Human-robot collaboration: A survey. In SoutheastCon 2015, pages 1–8. IEEE, 2015

work page 2015
[6]

Softmax regression based deep sparse autoencoder network for facial emotion recognition in human-robot interaction

Luefeng Chen, Mengtian Zhou, Wanjuan Su, Min Wu, Jinhua She, and Kaoru Hirota. Softmax regression based deep sparse autoencoder network for facial emotion recognition in human-robot interaction. Information Sciences , 428:49–61, 2018

work page 2018
[7]

Learning phrase representations using rnn encoder-decoder for statistical machine translation

Kyunghyun Cho, Bart Van Merri ¨enboer, Caglar Gulcehre, Dzmitry Bah- danau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 , 2014

arXiv 2014
[8]

Learning empathy-driven emotion expressions using affective modu- lations

Nikhil Churamani, Pablo Barros, Erik Strahl, and Stefan Wermter. Learning empathy-driven emotion expressions using affective modu- lations. In 2018 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2018

work page 2018

Show all 22 references

[9]

The inﬂuence of a socially assistive robot on mood, anxiety, and arousal in children

Molly K Crossman, Alan E Kazdin, and Elizabeth R Kitt. The inﬂuence of a socially assistive robot on mood, anxiety, and arousal in children. Professional Psychology: Research and Practice , 49(1):48, 2018

2018
[10]

Enhancing creativity: Proper body posture meets proper emotion

Ning Hao, Hua Xue, Huan Yuan, Qing Wang, and Mark A Runco. Enhancing creativity: Proper body posture meets proper emotion. Acta psychologica, 173:32–40, 2017

2017
[11]

Emotional contagion

Elaine Hatﬁeld, John T Cacioppo, and Richard L Rapson. Emotional contagion. Current directions in psychological science , 2(3):96–100, 1993

1993
[12]

Adaptive behavior acquisition of a robot based on affective feedback and improvised teleoperation

Masakazu Hirokawa, Atsushi Funahashi, Yasushi Itoh, and Kenji Suzuki. Adaptive behavior acquisition of a robot based on affective feedback and improvised teleoperation. IEEE Transactions on Cognitive and Developmental Systems , 2018

2018
[13]

Motivated attention: Affect, activation, and action

Peter J Lang, Margaret M Bradley, Bruce N Cuthbert, et al. Motivated attention: Affect, activation, and action. Attention and orienting: Sensory and motivational processes , 97:135, 1997

1997
[14]

Social robots for long- term interaction: a survey

Iolanda Leite, Carlos Martinho, and Ana Paiva. Social robots for long- term interaction: a survey. International Journal of Social Robotics , 5(2):291–308, 2013

2013
[15]

The icub humanoid robot: an open platform for research in embodied cognition

Giorgio Metta, Giulio Sandini, David Vernon, Lorenzo Natale, and Francesco Nori. The icub humanoid robot: an open platform for research in embodied cognition. In Proceedings of the 8th workshop on performance metrics for intelligent systems , pages 50–56. ACM, 2008

2008
[16]

Predictive learning: its key role in early cognitive development

Yukie Nagai. Predictive learning: its key role in early cognitive development. Philosophical Transactions of the Royal Society B , 374(1771):20180030, 2019

2019
[17]

Cognitive emotion regulation: Insights from social cognitive and affective neuroscience

Kevin N Ochsner and James J Gross. Cognitive emotion regulation: Insights from social cognitive and affective neuroscience. Current directions in psychological science , 17(2):153–158, 2008

2008
[18]

Parisi, J

G.I. Parisi, J. Tani, C. Weber, and S Wermter. Lifelong learning of spatiotemporal representations with dual-memory recurrent self- organization. In arXiv:1805.10966, 2018

2018 arXiv
[19]

Automatic analysis of facial affect: A survey of registration, representation, and recognition

Evangelos Sariyanidi, Hatice Gunes, and Andrea Cavallaro. Automatic analysis of facial affect: A survey of registration, representation, and recognition. IEEE transactions on pattern analysis and machine intel- ligence, 37(6):1113–1133, 2014

2014
[20]

Where to look: a study of human-robot engagement

Candace L Sidner, Cory D Kidd, Christopher Lee, and Neal Lesh. Where to look: a study of human-robot engagement. In Proceedings of the 9th international conference on Intelligent user interfaces , pages 78–
[21]

Emotional bodily expressions for culturally competent robots through long term human-robot interaction

Nguyen Tan Viet Tuyen, Sungmoon Jeong, and Nak Young Chong. Emotional bodily expressions for culturally competent robots through long term human-robot interaction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 2008–2013. IEEE, 2018

2018
[22]

Generic method for generating blended gestures and affective functional behaviors for social robots

Greet Van de Perre, Hoang-Long Cao, Albert De Beir, Pablo G ´omez Esteban, Dirk Lefeber, and Bram Vanderborght. Generic method for generating blended gestures and affective functional behaviors for social robots. Autonomous Robots , 42(3):569–580, 2018

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Forming impressions of personality

Solomon E Asch. Forming impressions of personality. The Journal of Abnormal and Social Psychology , 41(3):258, 1946

work page 1946

[2] [2]

Developing crossmodal expres- sion recognition based on a deep neural model

Pablo Barros and Stefan Wermter. Developing crossmodal expres- sion recognition based on a deep neural model. Adaptive behavior , 24(5):373–396, 2016

work page 2016

[3] [3]

A self-organizing model for affec- tive memory

Pablo Barros and Stefan Wermter. A self-organizing model for affec- tive memory. In Neural Networks (IJCNN), 2017 International Joint Conference on, pages 31–38. IEEE, 2017

work page 2017

[4] [4]

Measurement instruments for the anthropomorphism, animacy, likeabil- ity, perceived intelligence, and perceived safety of robots

Christoph Bartneck, Dana Kuli ´c, Elizabeth Croft, and Susana Zoghbi. Measurement instruments for the anthropomorphism, animacy, likeabil- ity, perceived intelligence, and perceived safety of robots. International journal of social robotics , 1(1):71–81, 2009

work page 2009

[5] [5]

Human-robot collaboration: A survey

Balasubramaniyan Chandrasekaran and James M Conrad. Human-robot collaboration: A survey. In SoutheastCon 2015, pages 1–8. IEEE, 2015

work page 2015

[6] [6]

Softmax regression based deep sparse autoencoder network for facial emotion recognition in human-robot interaction

Luefeng Chen, Mengtian Zhou, Wanjuan Su, Min Wu, Jinhua She, and Kaoru Hirota. Softmax regression based deep sparse autoencoder network for facial emotion recognition in human-robot interaction. Information Sciences , 428:49–61, 2018

work page 2018

[7] [7]

Learning phrase representations using rnn encoder-decoder for statistical machine translation

Kyunghyun Cho, Bart Van Merri ¨enboer, Caglar Gulcehre, Dzmitry Bah- danau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 , 2014

arXiv 2014

[8] [8]

Learning empathy-driven emotion expressions using affective modu- lations

Nikhil Churamani, Pablo Barros, Erik Strahl, and Stefan Wermter. Learning empathy-driven emotion expressions using affective modu- lations. In 2018 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2018

work page 2018