What is carl-simulation?

Carl Omega is a research project that simulates a virtual primate like robot with 28 moving joints inside MuJoCo, a physics engine used for robotics research. Instead of building a language model that only reads and writes text, this project tries to build something closer to a body and a brain working together. Carl has to deal with gravity, friction, and collisions, balance internal needs like hunger and fatigue, build a map of its surroundings, and control its arms well enough to reach for an object, grab it, and carry it somewhere else. The project describes its design as an eight layer stack, starting from raw physics at the bottom and building up toward more abstract thinking at the top. The lower layers handle the physical body, short term and long term memory, and a simulated set of brain chemicals such as dopamine, norepinephrine, and serotonin that influence how quickly Carl learns and how alert it becomes. Learning between simulated neurons happens through a technique called reward modulated spike timing dependent plasticity, where connections get stronger or weaker depending on the order in which neurons fire relative to a reward. Carl also has a memory system that stores what it experiences as very large vectors, split into separate registers for spatial information, emotional signals, and motor actions. A separate navigation system mimics grid cells and place cells, a type of brain cell involved in how animals map physical space, so Carl can track its own position and figure out safe paths. Arm movements are driven by a recurrent neural network that was first taught through imitation and then further trained using a reinforcement learning method called PPO. The repository is organized into folders for the brain related code, the core simulation environment, documentation, a dashboard for visualizing performance, and tests. To try it, you set up a Python virtual environment, install the listed dependencies, and run a single script that plays out a full search, approach, reach, grasp, lift, carry, and place sequence using the built in cognitive system. The full README is longer than what was shown.