{"id":18338,"date":"2025-01-08T13:07:10","date_gmt":"2025-01-08T18:07:10","guid":{"rendered":"https:\/\/hexagram.ca\/morphoses-trois-creatures-artificielles-explorent-la-relation-entre-matiere-et-comportement\/"},"modified":"2025-01-29T11:49:34","modified_gmt":"2025-01-29T16:49:34","slug":"demo53-morphoses","status":"publish","type":"post","link":"https:\/\/hexagram.ca\/en\/demo53-morphoses\/","title":{"rendered":"DEMO53 Sofian Audry & Rosalie D. Gagn\u00e9 \u2014 Morphoses : Three Artificial Creatures Explore the Relationship Between Matter and Behaviour"},"content":{"rendered":"\n

Janurary 2025<\/p>\n\n\n\n

<\/div>\n\n\n\n

Morphoses<\/em> is a research-creation project involving artificial life forms. It explores the boundary between material and behaviour, through the adaptive movements of spheroid robotic agents covered in silicone. Artists Sofian Audry<\/a> and Rosalie D. Gagn\u00e9<\/a> describe this envelope as a \u201cskin,\u201d giving it irregular and varied shapes.<\/p>\n\n\n\n

Emerging from an artistic reflection on the phenomenology of living systems, Morphoses <\/em>offers a poetic and embodied experience of automated learning, fostering an affective relationship between human and non-human life forms.<\/p>\n\n\n\n

This DEMO traces the development of the project during a residency at the Milieux Institute for Arts, Culture and Technology<\/a> in the spring of 2024.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\n\n\n\n
<\/div>\n\n\n\n

Creation in the Age of Machine Learning<\/strong><\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

The rapid progress of artificial intelligence, driven by large deep learning models, has led to a \u201c4th Industrial Revolution\u201d (Schwab, 2016) where adaptive computational systems are quickly taking over intellectual tasks. At the heart of this revolution is machine learning, a branch of AI that allows machines to learn on their own from real-world data.<\/p>\n\n\n\n

Machine learning is increasingly being used for artistic creation, often by training strong generative models to produce various types of media content: text, images, music, etc. This hegemonic approach presents several problems: it is energy-intensive, does not respect copyright, and tends to homogenize culture.<\/p>\n\n\n\n

<\/div>\n\n\n\n

Learning Entities<\/strong><\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

Morphoses<\/em> deviates from this dominant approach by using the learning process itself to generate behaviours within an artificial life installation.<\/p>\n\n\n\n

Inspired by the artificial turtles of cybernetician Grey Walter (Walter, 1950) and the drawings of naturalist Ernst Haeckel (Haeckel, 1974), the project features autonomous spheroid agents whose behaviours emerge in real-time thanks to a reinforcement learning algorithm. These entities receive positive or negative rewards, visible by a green or red glow, based on their actions.<\/p>\n\n\n\n

<\/div>\n\n\n
\n
\"\"
Photo\u00a0credit: L\u00e9a Martin<\/sup><\/figcaption><\/figure>\n<\/div>\n\n\n
<\/div>\n\n\n\n

Aesthetics of Adaptive Behaviour<\/strong><\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

This project falls within the field of behaviour aesthetics (Penny, 1997). This emerging domain of digital arts uses the behaviours of artificial systems as artistic material. Rather than aiming for optimization or performance, Morphoses<\/em> leverages the learning process for an embodied aesthetic experience. By focusing on adaptive behaviours and their perceptual, sensitive, and emotional potential, the installation questions the boundaries between human and non-human. How do form and matter influence the behaviour of agents? What emotions are evoked by observing learning agents?<\/p>\n\n\n\n

By exploring the spatial and temporal aesthetic qualities of adaptive behaviours, the project invites us to rethink the nature of artificial intelligence, not as a utilitarian or threatening force, but as a fragile, evolving, and poetic process, opening new avenues for research-creation.<\/p>\n\n\n\n

<\/div>\n\n\n\n

Reinforcement Learning<\/strong><\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

The work relies on the real-time deployment of adaptive behaviours. The robotic agents were developed iteratively until functional prototypes were obtained, replicated in three copies.<\/p>\n\n\n\n

<\/div>\n\n\n
\n
\"\"<\/figure>\n<\/div>\n\n\n
<\/div>\n\n\n\n

These agents use reinforcement learning, a method that allows an artificial agent to learn on its own to achieve a goal (Sutton & Barto, 2018). For example, in preliminary experiments:<\/p>\n\n\n\n

    \n
  • A robot learned to stop.<\/li>\n\n\n\n
  • Initially, it knew nothing about its environment and activated its motors randomly.<\/li>\n\n\n\n
  • Gradually, it oscillated and eventually came to a complete stop.<\/li>\n<\/ul>\n\n\n\n
    <\/div>\n\n\n\n

    Although the goal seems simple, the resulting behaviour is often complex and unpredictable. Each behavioural mode was developed empirically, as the generated behaviours cannot be precisely anticipated, even when the defined goals are known.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    Configuring and Influencing Behaviour<\/strong><\/strong><\/h2>\n\n\n\n
    <\/div>\n\n\n\n

    Each behaviour results from the interaction between different components. Each behaviour is created from three main categories of constraints. First, the agent\u2019s observations<\/strong> (or states) determine the information it has access to for making decisions. The available actions<\/strong> impact its ability to act on its world. Finally, the reward<\/strong> function determines the success level of an action in response to an observation.<\/p>\n\n\n\n

    For example, to achieve the behaviour titled \u201cSwinging without Rolling,\u201d the observations are defined by internal movement; the possible actions are move forward, backward, left, right; and the agents are rewarded positively when they generate internal movement and negatively when they move horizontally on the ground.<\/p>\n\n\n\n

    In addition to these parameters directly controllable by the artists, at the physical level, the form and materiality of the agent (texture, weight, centre of mass, etc.) combined with its environment (movement surface, obstacles, space boundaries, etc.) are important elements influencing behaviour. For example, an agent will find it easier to perform the task \u201cSwinging without Rolling\u201d if it is stuck to an obstacle or if it is heavier.<\/p>\n\n\n\n

    <\/div>\n\n\n
    \n
    \"\"<\/figure>\n<\/div>\n\n\n
    <\/div>\n\n\n\n

    Visualization and Representation of Machine Learning<\/strong><\/strong><\/h2>\n\n\n\n
    <\/div>\n\n\n\n

    To present the different behaviours, the work unfolds as a performance of about 15 minutes, where the various behaviours follow one another sequentially.<\/p>\n\n\n\n

    For each behaviour, a short title is projected on the wall of the presentation space: \u201cBeing still,\u201d \u201cMoving toward the wall,\u201d \u201cBeing curious,\u201d to allow visitors to interpret the different behaviours or intentions of the agents.<\/p>\n\n\n\n

    Additionally, a visualization screen allows real-time tracking of the reward level (positive or negative) assigned to each agent, based on how their actions bring them closer to or further from the intended goal.<\/p>\n\n\n\n

    Finally, through the translucent silicone membrane that serves as their skin, the internal movements of the motors and counterweights can be discerned thanks to a system of light-emitting diodes that display different oscillating colours from red to green, depending on whether the actions taken bring them closer to or further from their objective.<\/p>\n\n\n\n

    <\/div>\n\n\n\n
    \n