A world model that moves with you
How predictive representations support navigation, planning, and embodied control.
Prediction begins with an action
A useful world model answers a practical question: what changes when I act? Reaching for a branch requires an estimate of the body, the branch, and the consequences of moving between them. Description becomes useful when it supports a decision.
Our approach at Hominid puts this loop at the center of intelligence. Perception supplies evidence; action tests an expectation; the next observation makes the result available for correction. This gives our research a concrete organizing principle: examine how a representation helps its owner do something, and how that representation changes when the world responds.
Predict. Act. Update.
The body predicts its next state
Wolpert, Ghahramani, and Jordan studied how people estimated the position of a hand after moving it in darkness, including movements under externally applied forces. The pattern of estimation errors supported an internal model that combines motor information with sensory evidence. Their experiment provides a precise starting point for studying a world model: the body's estimate of its own changing state. (Wolpert et al. 1995)
In this framework, a forward model predicts the next state from the current state and a motor command. That prediction is useful while sensory feedback is delayed. For Hominid, the design consequence is direct: action belongs inside the representation, because the same scene can lead to different outcomes under different movements. (Wolpert et al. 1995)
A map can represent what comes next
Stachenfeld, Botvinick, and Gershman developed a computational account in which hippocampal representations encode expected future occupancy of states. This predictive map connects a location to what tends to follow it under a pattern of behavior. It explains several properties of place-cell activity that a purely geometric map leaves unresolved. (Stachenfeld et al. 2017)
Pfeiffer and Foster supplied complementary evidence in rats navigating an open arena. Before travel, brief hippocampal sequences represented paths biased toward a remembered goal; those sequences predicted immediate future behavior, including unfamiliar combinations of starting point and destination. The study ties prospective neural activity to navigation without treating every imagined route as an executed plan. (Pfeiffer & Foster 2013)
Evaluate the consequences
These findings shape three questions for our world-model program. Can a representation anticipate the consequence of a movement? Can it connect present circumstances to a distant goal? Can it remain useful when the starting point, reward, or route changes? Each question asks for a distinct ability.
They also separate the evidence. Hand-position estimation tests sensorimotor integration. Neural sequences reveal prospective representations during navigation. A computational predictive map explains patterns across observations. Together, they establish a research direction grounded in control, memory, and planning. Our central design principle follows from that synthesis: judge a world model by the decisions it makes possible.
References
- Brad E. Pfeiffer and David J. Foster. Hippocampal place-cell sequences depict future paths to remembered goals. Nature 497, 74–79 (2013). doi:10.1038/nature12112
- Daniel M. Wolpert, Zoubin Ghahramani, and Michael I. Jordan. An Internal Model for Sensorimotor Integration. Science 269(5232), 1880–1882 (1995). doi:10.1126/science.7569931
- Kimberly L. Stachenfeld, Matthew M. Botvinick, and Samuel J. Gershman. The hippocampus as a predictive map. Nature Neuroscience 20, 1643–1653 (2017). doi:10.1038/nn.4650