Mouhssine Rifaki

Current Research and Scholarly Interests

My current research utilizes RL in order to train embodied agents whose perception and foveation adapt while being deployed. I have made the claim that all of an agent's perception, decision-making, and physical actions should be designed simultaneously, as opposed to sequentially. The signal that connects each of the aforementioned components is prediction errors derived from a learned forward model.

In my work, I examine how adaptive sensor systems dynamically transition among different modalities based upon failures within a lightweight world model predicting what will be observed in the near-term future. I also explore foveated perception that devotes its high-resolution attention resources to those regions of space most likely to yield returns under conditions of distribution shift. Lastly, I investigate the design of real-time closed-loop control policies that utilize their dynamic sensing capabilities as input sources for their subsequent actions taken via the same senses.

Projects

• Adaptive Sensing for Physical AI, Stanford University (4/1/2026 - Present)

  Using fast–slow hierarchical models that use lightweight prediction errors to select modalities and resolutions used by a high-fidelity model. Developed proof-of-concept using multimodal gesture recognition; developed toward tactile sensing as well for robotic manipulation.

  Location

  Stanford, CA

  Collaborators

  • Amin Arbabian, Associate Professor, Stanford University
• Partner UED for Cooperative Multi-Agent Learning, New York University (3/1/2026 - 9/30/2026)

  PAIRED-style unsupervised environment design over partner policies in cooperative multi-agent learning, rather than over level layouts: the adversary samples co-players from a learned population, and the ego learns to coordinate with whatever partner it draws.

  Location

  New York, NY

  Collaborators

  • Eugene Vinitsky, Assistant Professor, New York University