Periodic Reporting for period 1 - METATOOL (A metapredictive model of synthetic awareness for enabling tool invention)
Reporting period: 2022-10-01 to 2023-09-30
Around 3.3 million years ago our ancestors made the first tool. They imagined a new utensil and then knapped a stone until it became an efficient tool for cutting. Tool creation was an outstanding technological milestone for humanity providing us with unprecedented control over our environment. This ability required cognitive capabilities, such as prediction, metacognition, abstraction, and creativity—all of which are associated in humans with awareness. Current artificial intelligence systems and robots largely lack these capabilities and cannot even monitor and evaluate the consequence of their actions let alone develop new tools to address environmental challenges.
METATOOL aims to provide a computational model of synthetic awareness to enhance adaptation and achieve tool invention. This will enable a robot to monitor and self-evaluate its performance, ground and reuse this information for adapting to new circumstances, and finally unlock the possibility of creating new tools. Under the predictive account of awareness, and based on both neuroscientific and archaeological evidence, we will: 1) develop a novel computational model of metacognition based on predictive processing and 2) validate its utility in real robots in tool uses scenarios, such as tool selection, tool discovery, tool innovation and tool invention.
METATOOL will provide a blueprint for the next generation of artificial systems and robots that can perform adaptive, and anticipative, control with and without tools (improved technology), self-evaluation (novel explainable AI), and invent new tools (disruptive innovation). Tool-making and tool-invention are outstanding technological milestones in human history. A similar breakthrough can now be envisioned in engineering. We already have algorithms to enable machines to use tools and now it is time to develop robots that create tools.
METATOOL aims to provide a computational model of synthetic awareness to enhance adaptation and achieve tool invention. This will enable a robot to monitor and self-evaluate its performance, ground and reuse this information for adapting to new circumstances, and finally unlock the possibility of creating new tools. Under the predictive account of awareness, and based on both neuroscientific and archaeological evidence, we will: 1) develop a novel computational model of metacognition based on predictive processing and 2) validate its utility in real robots in tool uses scenarios, such as tool selection, tool discovery, tool innovation and tool invention.
METATOOL will provide a blueprint for the next generation of artificial systems and robots that can perform adaptive, and anticipative, control with and without tools (improved technology), self-evaluation (novel explainable AI), and invent new tools (disruptive innovation). Tool-making and tool-invention are outstanding technological milestones in human history. A similar breakthrough can now be envisioned in engineering. We already have algorithms to enable machines to use tools and now it is time to develop robots that create tools.
This first year has focused on understanding how ancient humans might have transited from tool use to tool creation and the development of the first computational model prototypes that derive from this knowledge.
For that purpose, the literature on early hominins and great apes' stone tool-making was revisited focusing on the connection to metacognition (i.e. thinking about thinking). This is, to study if ancient human tool creation was facilitated by their ability to think about their performance when using a tool. Furthermore, to reduce the interdisciplinary barriers of the project, a common ground terminology was created (a glossary and an ontology). For instance, to establish the definition of the tool-use concept from three axes: neuroscience, archaeology and robotics.
These first insights were transferred to two computational models inspired by how the brain may make causal inferences about the world. The first one describes how we can perform tool discovery (find tools in the environment) and generalize to tool innovation (e.g. use the acquired affordance knowledge to come up with a better combination of tools). The second one, inspired by human metacognitive performance, describes how robots can make use of control confidence to select the proper tool when the dynamics are known (e.g. selecting the tool that has less uncertainty involved in the outcome). However, the problems solved are still low-dimensional.
Finally, a bimanual humanoid was delivered and incorporated into a custom-built dual-arm robotic testbed for tool invention. This includes the physical scenario and its digital twin.
For that purpose, the literature on early hominins and great apes' stone tool-making was revisited focusing on the connection to metacognition (i.e. thinking about thinking). This is, to study if ancient human tool creation was facilitated by their ability to think about their performance when using a tool. Furthermore, to reduce the interdisciplinary barriers of the project, a common ground terminology was created (a glossary and an ontology). For instance, to establish the definition of the tool-use concept from three axes: neuroscience, archaeology and robotics.
These first insights were transferred to two computational models inspired by how the brain may make causal inferences about the world. The first one describes how we can perform tool discovery (find tools in the environment) and generalize to tool innovation (e.g. use the acquired affordance knowledge to come up with a better combination of tools). The second one, inspired by human metacognitive performance, describes how robots can make use of control confidence to select the proper tool when the dynamics are known (e.g. selecting the tool that has less uncertainty involved in the outcome). However, the problems solved are still low-dimensional.
Finally, a bimanual humanoid was delivered and incorporated into a custom-built dual-arm robotic testbed for tool invention. This includes the physical scenario and its digital twin.
Although project results are still preliminary, first, thanks to the literature review a new experiment that will aid in validating or falsifying our metacognitive hypothesis of tool use and tool making has been identified. Second, the two newly developed computational models (high-level planning and low-level control) utilise the monitored uncertainty to make informed decisions for tool selection and innovation. This goes beyond other models of awareness in robotics, providing a novel approach derived from neuroscience. Besides, the realization of common terminology and ontology in tool creation and invention provides a novel insight into material archaeology.