Cognitive Models as Evaluation Primitives in Humanoid Factors

Foundation models are expanding what humanoid robots can do, but they also widen the gap between task completion and human-compatible behavior. In human environments, people coordinate using cognitive regularities---for instance, predictable timing and legible motion---that they can anticipate. Advocating for metrics beyond task completion, in this paper, we introduce several aspects of the Cognitive Pillar of our broader Humanoid Factors (HoF) framework, a cognitively grounded perspective in which explicit cognitive models serve as evaluation primitives for embodied reasoning and control. As a case study, we evaluate if a reaching task trained with a behavior cloning neural network on a Unitree G1 humanoid follows Fitts' Law, a classic human psychophysics model for motor control, relating movement time to task difficulty. In simulation, rollouts can appear acceptable under completion-centric metrics; in real-robot deployment, the expected Fitts relationship breaks, revealing hesitation and timing irregularity that standard robotics metrics often miss. Overall, these preliminary results motivate our broader goal of incorporating compact cognitive models as interpretable evaluation primitives for humanoids, with the longer-term aim of applying these principles to foundation model–driven control policies, where human-compatibility failures may be even harder to detect.Full paper draft: \url{https://anonymous.4open.science/r/humanoid-factors}