Developing a ROS-based Cognitive Architecture for AGI in Humanoid Robots. Using Deep Neural Networks for Representation and Reasoning.
Introduction: We do not need science-fiction to tell us that robots will one day walk among us. This means robots will need to interact seamlessly in a multi-agent dynamic environment, that is our world. AI has taken great strides in past few years, and is expected to mature into AGI eventually. While robots like Sophia by Hanson Robotics, uses OpenCog Prime to express behaviour, I doubt that it will serve as a platform for the open-source community to develop AGI. The goal of starting ROSCOG is to create an open-source cognitive architecture for humanoid robots. Perhaps, this new class of robots agents introduced into the society, will unite the human agents and end our social and political problems.
Timeline:
Mental Representations:
A mental representation, also called cognitive representation, in philosophy of mind, cognitive psychology, and neuroscience is a hypothetical internal cognitive symbol that represents external reality. [ya]
roslaunch cogarch diagnostics.launch #different modes: ON [awake (autonomous), diagnostics, sleep (minimal power)] | OFF [dead]
Inspired by human mental representations, the proposed framework also incorporates this for a humanoid robot's mind. Similar to the concept of life and death in humans, a robot has two basic modes: it can be OFF, where it's electromechanical parts are not being powered by electricity or it can be ON. In its ON state, it is switched on and the battery is being accessed. Furthermore, the ON state contains mental representations: Awake, Diagnostics, Robot and Asleep.
Awake Mode: Autonomous, Learning, Biological Fitness, Maslow Hierarchy of Needs, Personality Robot: Domain-Role: Process NLP Input, Load Memory/Experience, assume role, perform/act. Diagnostics: Control Mode, adjust memory, beliefs revision Asleep Mode: minimal power, repair process, charge battery, data defragment, LTM/Belief Revision
Use ROSCOG for input/output and processing of a robot.
- Metacognition to produce a Dec-POMDP for an agent for decision-making.
'Knowledge Representation and Reasoning' can be blockchain based, allowing "safe" AGI.
- Using Probablisitc Programming with Hypergraphs
- Use a DQN for decision-making and reinforcement learning.
Tools: Python + ProbLog + ROS + TensorFlow
please dont steal my ideas/work, i want to publish research on this, as this is might be my thesis (semi?/nano?), and eventually make a career out of this. contact me to collaborate instead ?
bibliography: [ya] Morgan, Alex (2014). "Representations Gone Mental" (PDF). Synthese 191.2: 213–44.