Control theoretical models for visuomotor control explains brain activity during naturalistic driving

Poster Presentation: Tuesday, May 21, 2024, 2:45 – 6:45 pm, Pavilion
Session: Action: Locomotor, flow, steering

Tianjiao Zhang1, Christopher A Strong1, Kaylene Stocking1, Jingqi Li1, Claire Tomlin1, Jack L Gallant1; 1UC Berkeley

Visuomotor control is crucial for successfully navigating through the world, as we must continuously adjust our actions to account for the behavior of other agents. Multiple brain regions, including the intraparietal sulcus (IPS), motor cortex, supplementary motor areas (SMA), and the prefrontal cortex (PFC), have been implicated in visuomotor control. However, the control algorithms implemented by these regions during interactions with other agents remain poorly understood. Here, we examined whether the brain may use algorithms similar to control theoretical models for car-following. We used fMRI to record brain activity from six participants performing a taxi-driver task in a large virtual world (110-180 minutes of data per participant). Virtual traffic required participants to constantly monitor other vehicles and adjust their own actions. We implemented three control theoretical car-following algorithms: the optimal velocity model (OVM) and intelligent driver model (IDM), two reactive dynamical systems models, and a model predictive control (MPC) model, a forward predictive model. In preliminary analysis in two participants, we tuned the parameters of the three control models to match the behavior of each participant, and used these tuned control models to create features for modeling brain activity. We used banded ridge regression (Nunez-Elizalde et al., 2019, Dupré la Tour et al., 2022) to estimate voxelwise encoding models for these control models along with 34 other feature spaces for the taxi-driver task. The MPC control model better matches the participants’ behavior than IDM or OVM, and the MPC encoding model better predicts brain activity than the IDM and OVM encoding models. Well-predicted regions include parts of the PFC, IPS, SMA, and motor cortex. Encoding model weights reveal multiple timescales of predictive control in the cortex. These results suggest that the human brain may implement a forward predictive algorithm similar to MPC for optimal visuomotor control during driving.

Acknowledgements: This work is supported by funding from the NEI, ONR, Ford URP, and NSF GRFP