Implications of recursive Bayesian Sensory Inference

Abstract

In this article, we examine how the balance between velocity-based and position-based proprioceptive feedback influences state estimation during motor control. We introduce a computational model of arm state inference grounded in Bayesian sensory integration and compare its behaviour to findings from three classical sensorimotor studies. The model allows us to contrast the predicted behaviour of two proprioceptive configurations: one relying primarily on position signals and another relying primarily on velocity signals, reflecting the distinct contributions of type II and type Ia muscle spindle afferents. Our simulations show that a system with strong reliance on velocity-based feedback tends to represent movement relative to previously estimated positions. In a simulated reaching task with briefly presented offset visual feedback, such an Agent produces systematic endpoint errors even after visual feedback is removed. In contrast, an Agent relying mainly on positional feedback is able to correctly update its inferred hand position after visual feedback is removed, thereby limiting this type of biased endpoint errors. When biased visual feedback remains continuously available, or when muscle vibration is simulated, the two configurations produce very similar behaviour. These results indicate that markedly different assumptions about the weighting of positional and velocity proprioceptive cues can yield similar observable behaviour in some of the often-used experimental setups designed to probe state inference processes. This underscores the importance of carefully considering the composition of proprioceptive signals when building computational models and interpreting human sensorimotor experiments. We highlight the task conditions under which our model predicts clear behavioural differences arising from the relative contribution of velocity versus positional feedback. 1 Author summaryDuring motor control, the central nervous system tracks both the position and movement of our limbs. Even with our eyes closed, we can bring the tips of our index fingers together, a feat that depends on sensory signals from specialised receptors in our muscles. One set of these receptors is most sensitive to the rate of muscle lengthening, providing information about movement, while another set signals the muscles current length, giving us a sense of posture. There has so far been limited discussion of the extent to which these two feedback channels are used to augment one another. For example, signals about the rate of muscle lengthening reflect how our pose is changing from moment to moment, but it remains unknown to what extent the brain uses this information to update its estimate of limb position. In this study, we simulate three classical sensorimotor experiments while varying the precision of these two types of sensory feedback. Two simulations show that different assumptions can yield similar movement patterns, highlighting the need for caution when interpreting such experiments. The third suggests that movement-related feedback may contribute more to our sense of limb position than previously recognised.