Josias Moukpe

Research Interests

• Recursive Self-Improvement
  Architecting autonomous recursive evolution, infrastructure where agents design their own successor architectures. Focus on stabilizing post-training feedback loops where models fork, optimize, and merge weights, creating a flywheel of capability scaling that remains interpretable and steerable.
• Self-Play, Model-based RL & Generalist Agents
  Developing a generalist cognitive core as a foundation for generalist agents. The architecture integrates large multimodal VLMs with efficient world dynamics models and long-term memory to ground the agent in reality and past experiences. By leveraging Monte Carlo Tree Search guided by VLM-based value functions, the system performs deep lookahead search. An inverse dynamics action flow model produces the required action chunks to execute the strategies found. When embedded in an open-ended self-play RL environment, this creates an autocurriculum with unlimited experiential data, allowing the agent to iteratively discover novel strategies and scale its capabilities to superhuman levels unattainable by static datasets or traditional model-based RL alone.
• Embodied Foundation Models & Whole-Body Control
  Extending the generalist cognitive core into physical reality. Focusing on cross embodiment where generalist architectures are deployed across diverse morphologies, including humanoids, drones, and autonomous vehicles. The full architecture treats control as a joint generative process: the efficient world dynamics model generates latent video futures, while the morphology-grounded inverse dynamics action flow model produces the required sequence of actions. This paradigm enables zero-shot generalization in unstructured real-world environments without the need for explicit reward engineering or task-specific fine-tuning.
• Open-Ended Optimization
  Exploring algorithms that prioritize novelty and diversity to discover complex behaviors outside the training distribution. Moving beyond objective-based gradient descent to use evolutionary dynamics for automated architecture search and curriculum generation, drawing inspiration from natural evolution and its power to create complex well-adapted organisms from simple building blocks.