Abstracts Track 2021


Nr: 37
Title:

Learned Dynamics Models and Online Planning-based Animation Generation for Data-driven Virtual Character Agents

Authors:

Vihanga Gamage, Robert Ross and Cathy Ennis

Abstract: Data-driven agents for virtual character animation control offer great potential for both recreational and serious games and applications. For the characters to be most effective in these instances, the behaviour portrayed by output animation needs to be realistic, dynamic and responsive to live events and cues from the user. Current state-of-the-art work in the area has shown impressive results for using supervised learning to output select behaviours provided sufficient motion clips are available for training, but these methods only allow for limited dynamism. Also, deep reinforcement learning (RL) methods have been shown to be able to leverage physics engine-derived feedback and signals to generate animation portraying behaviours such as locomotion. RL-based approaches have the potential for greater generalization, leading to agents being able to learn a wide range of behaviours efficiently, and generate dynamic animation in real-time. However, current state-of-the-art RL agents depend heavily on feedback derived from physics engines. Animation portraying social, interactive behaviours do not elicit physics-driven signals that would allow for a policy to be shaped effectively, making these behaviours incompatible within a physics-driven paradigm for RL agents. In this work, we use gazing and pointing as exploratory tasks to explore the feasibility of a paradigm suitable for RL-based animation agents that learn latent dynamics that are applicable ubiquitously, from a modelling perspective, to a wider range of behaviours. We introduce a framework where agents learn generalizable animation dynamics required to portray different behaviours, and we propose a novel method for animation generation based on online planning in a beta distribution parameterised space. Agents learn a latent dynamics model enabling them to predict a character state, and generate animation via online planning, using a corresponding objective state to discern which candidate sequences represent the optimal animation trajectory. Purely through self-exploration and learned dynamics, agents created within our framework are able to output animations to successfully complete gaze and pointing tasks robustly while maintaining smoothness of motion, using a very low number of training epochs. In our experimental validation, we compared outputs generated from a trained agent to an Inverse Kinematics (IK)-based control implementation. We found that agents created using our method were significantly more computationally efficient, taking approximately 40 nanoseconds per frame to generate animation, compared to 2 milliseconds for the IK-based control. In our future work, we plan to develop methodologies to use motion capture data as an external source of information, in co-ordination with our model-based reinforcement learning training algorithm, to influence agents during training to account for realism. A video containing an overview of our work and examples of animation output can be found at https://virtualcharacters.github.io/links/ICAART2021

Nr: 26
Title:

Aristotelian Ethics in Social Robotics: Phronetic Robotics

Authors:

Roman Krzanowski and Pawel Polak

Abstract: This paper is the contribution to the discussion “How far can robots go--now and in the future--to fulfill the requirements of full-blown social agents?” focusing on “principles, and procedures” for social agency embodied in smart autonomous robots. We will be taking about specifically humanoid machines aka social robots that would enter the social fabric of the society in substitute roles for human agents. Such social robots may perform roles that have been the domain of humans such as teachers, personal assistants, helpers, companions, care givers, may be even offering some sort of psychological support or medical help.

Nr: 36
Title:

Visually-private Scene Classification with Agent-collected Weak-labels

Authors:

Shin Ando, Yusuke Hatae, Muhammad F. Fadjrimiratno, Qingpu Yang, Yuanyuan Li, Tetsu Matsukawa and Einoshin Suzuki

Abstract: In real-world applications, collecting and labeling on-site images raises concerns for privacy and security as well as cost. In this paper, we consider a scene classification application for private home and office environments facilitated by a mobile, monitoring agent, which replaces the visual input of the classifier with textual information to address these concerns. The monitoring agent is implemented with an on-board captioning model and the images it captures are discarded after generating caption texts, which serve as weak-labels, of the salient regions. Our intuition is that private aspects, such as personal profiles and proprietary equipments, will be protected by eliminating the visual input, but discriminative information can be retrieved from the patterns of the weak-label texts. The captioning model is pre-trained off-site to prevent identification of site-specific objects, and the exact coordinates of the extracted regions are also discarded to obscure their layouts as well. To maintain the classification performance with such limited input, we propose a recurrent neural network (RNN) framework, which preserves the spatial information with multiple sequences of texts. That is, we parametrize functions to project the coordinates of the salient regions, with which their captions are aligned to form a sequence, and those parameters are trained with feedbacks from the RNN classifying those sequences. We conduct an empirical study on two indoor datasets collected by an agent monitoring a laboratory environment. The results provides a proof-of-concept for the application and show that the performance of the proposed framework is competitive with the classifier exploiting visual input.

Nr: 18
Title:

Psychphysiological Modelling of Trust in Technology: Comparative Analysis of Algorithm Ensemble Methods

Authors:

Ighoyota B. Ajenaghughrure, Sonia Sousa and David Lamas

Abstract: Measuring user’s trust in technology in real-time using psychophysiological signals depends on the availability of stable, accurate, variance sensitive, and non-bias trust classifier model which can be achieved through ensembling several algorithms. Prior efforts resulted to fairly accurate but unstable models. This article investigates what ensemble method is most suitable for developing an ensemble trust classifier model for assessing users trust in technology with psychophysiological signals. Using a self-driving car game, a within subject four condition experiment was implemented. During which 31 participant were involved, and multimodal psychophysiological data (EEG, ECG, EDA, and Facial-EMG) were recorded. An exhaustive 172 features from time and frequency domain were extracted. Six carefully selected algorithms were combined for developing ensemble trust classifier models using each of the four ensemble methods (voting, bagging, stacking, boosting). The result indicated that the Stack ensemble method was more superior, despite voting method dominating prior studies.

Nr: 28
Title:

Bayesian Networks’ Fixed Structured to Manage with Missing Values in Environmental Data.

Authors:

Rosa Fernández-Ropero, M. Julia Flores and Rafel Rumi

Abstract: Environmental data often present missing values, mainly when data come from automatic data-collected systems. When they comprise a high percentage of the dataset, modelling tasks could be deeply affected. Under the framework of SAICMA project, data dealing with water management were collected from the Andalusia Hydrological Systems from October 2011 to September 2020. However, for one dam level variable, there is no information from May to September 2020. In order to avoid removing this variable from the entire model, or those 5 months, this abstract aims to input its values as accurately as possible. A complete dataset (October 2011 to September 2019) was used to develop Bayesian network (BN) regression models based on fixed structures (Naïve Bayes, NB, and Tree Augmented Naïve, TAN) using Elvira software. An scenario was carried out with new data (October 19 to March 2020) to impute those missing values. Results show both NB and TAN models are able to predict the behaviour of the objective variable, but slightly overestimating it, mainly because in this initial approximation, only dam’s inputs are considered, not the output (consume or evaporation). In order to check if this data imputation affects final environmental conclusions, a BN based on expert’s structural learning was learnt. Data with real and predicted values were used (October 2011 to April 2019 with real data, and from May to September 2019 was imputed). Results show the root mean square error of these models. Models learnt with data imputed improve the error of model from real data. Our work proves that BN based on fixed structures provide a robust and powerful tool able to deal with the lack of information. In particular, our application case shows how probabilistic models can be used to include new information or evidence that help to perform management or climate scenarios.