DeepSelf: Emergence of Event-Predictive Agency in Robots

“控制自己的行为，并通过它们控制外部世界的事件的经验”（Haggard & Chambon 2012）是代理感的核心。虽然代理的形式可以在感觉运动经验的直接编码中找到，但我们认为，更明确的、可访问的形式需要从实际的感觉运动动态（即事件）中抽象出来。其结果可能被称为代理自我，它可以“意识到”自己的经历以及其在世界中的行为的后果。我们的目标是揭示关键的计算组件，包括学习和处理偏见，为机器人代理自我的发展。在三年的时间里，我们的目标是首先建模空间动作-效果绑定，实现一种简单的代理形式。然后，我们将增强该架构以模拟事件效应预期，重点关注预期跨模一致性范式，该范式显示了我们的大脑如何在开始执行所需的运动以达到位置之前将我们的身体部位投射到未来的位置。最后，我们将解决工具介导的事件效应预期，我们希望在人类参与者的实验中首先显示。我们的计算模型以意念运动理论、比较器模型和自由能原理（主动推理）为出发点。近年来，包括SPP第一个资助期的研究工作，我们已经在各种人工系统和机器人中实施了这些原则。我们的深度主动推理模型使机器人能够从连续的原始感官信息中学习生成模型，并以模型预测的方式进行规划。此外，受我们对事件预测认知理论贡献的启发，我们还实现了事件预测系统，该系统将相对距离和方向转换为事件编码，使代理能够在事件尺度上直接规划目标。通过将我们在自适应机器人和深度人工神经网络（Donders）方面的专业知识与我们在实验认知心理学和神经认知建模（t<s:1>宾根）方面的专业知识相结合，我们的目标是隔离“允许[人工]代理发展自我的机制和先决条件”，并仔细检查“代理角色”，促进更有效的事件控制的发展。此外，我们希望确定工具使用中自塑性的核心机制。同时，我们设想通过开发紧凑的事件预测编码来提高机器人的代理处理能力。在实际项目之外，我们期望对能够解释其对环境的影响并学会识别其因果关系的系统做出贡献。虽然我们将在这个项目中关注单个机器人，但我们希望代理自我的实现也将促进社会互动能力的发展——我们希望在第二个资助阶段与其他SPP项目讨论这些考虑因素。

The “experience of controlling one’s own actions, and, through them, events in the outside world” (Haggard & Chambon 2012) lies at the heart of the Sense of Agency. While forms of agency may be found in direct encodings of sensorimotor experiences, we propose that more explicit, accessible forms require abstractions away from the actual sensorimotor dynamics, i.e. events. The result may be called an agentive self, which can become ‘aware’ of its own experiences as well as the consequences of its actions in the world.We aim at revealing critical computational components, including learning and processing biases, for the development of an agentive self in robots. Over the three years, we aim at first modeling spatial action-effect binding, to implement a simple form of agency. We will then enhance the architecture to model event-effect anticipations, focusing on the anticipatory crossmodal congruency paradigm, which shows how our minds project our body parts onto future positions before even starting to execute the required motion to reach the position. Finally, we will tackle tool-mediated event-effect anticipations, which we expect to first show in experiments with human participants.Our computational model takes ideomotor theory, comparator models, and the free energy principle (active inference) as the point of departure. Over recent years, including research work within the SPP’s first funding period, we have implemented these principles in various artificial systems and robots. Our deep active inference model enables robots to learn generative models from continuous raw sensory information and to plan in a model-predictive manner. Furthermore, inspired by our contribution to theories of event-predictive cognition, we have also implemented event-predictive systems, which convert relative distances and orientations, into event encodings, enabling agents to plan goal-directly on an event scale.By combining our expertise in adaptive robotics and deep artificial neural networks (Donders) with our expertise in experimental cognitive psychology and neuro-cognitive modeling (Tübingen), we aim to isolate “the mechanisms and prerequisites that allow an [artificial] agent to develop a self” and the scrutinization of the “roles of agency”, fostering the development of more effective event control. Moreover, we expect to identify core mechanisms of self-plasticity in tool-use. Meanwhile, we envisage improving the robot’s agentive processing abilities via the development of compact event-predictive encodings. Beyond the actual project, we expect to contribute to systems that can explain their influence on the environment and that learn to identify its causality. While we will focus on individual robots in this project, we hope that the realization of an agentive self will also facility the development of social interaction competencies – considerations which we hope to discuss with other SPP projects during the second funding phase.