Coombs and Trestman present a convincing argument that complex cognition in vertebrates, cephalopod mollusks, and euarthropods emerges from pivotal sensory, motor, and neural traits, particularly emphasizing active visual sensing and visuomotor control. Their analysis highlights high-resolution vision, laminated visual brain structures, and advanced motor control as key drivers of cognitive evolution. However, their framework appears overly centered on vision and visual processing as drivers of cognitive evolution, which further leads to a view of cognitive processing as a sequential process. From our point of view, this framework would benefit from, first, considering sensory processing as multimodal in nature, which lays a foundation for grounded and functional internal models. Second, from an embodied cognition perspective (Barsalou, Reference Barsalou2008), cognition should not be reduced to a staged sensory-driven process. Instead, perception and action should be better viewed as inherently intertwined in control as well as in cognition.

Central to an embodied cognition perspective is the idea that control arises dynamically from integrated sensorimotor loops that are recruited in cognition for higher-level functions such as planning (Anderson, Reference Anderson2010). Sensorimotor loops are constituted by perceptual and motor processes that form hierarchical structures of functional representation (Merel & Botvinick, Reference Merel, Botvinick and Wayne2019). However, an embodied cognition perspective emphasizes that sensory and motor processing are tightly interconnected and interact dynamically across different levels of this hierarchy. Importantly, internal representations are grounded along the hierarchy as well, particularly serving immediate function in sensorimotor tasks on lower levels. As one example, the multimodal integration of sensory input constitutes an early form of such functional models (Ernst & Bülthoff, Reference Ernst and Bülthoff2004). Sensory information is often noisy and unreliable, which requires forms of integration that inherently involve internal modeling—exploiting relationships between modalities in order to derive more reliable signals. When focusing on the sensory part of the sensorimotor loops, a wide range of animal systems exhibit such forms of sensorimotor integration that is inherently multimodal. For example, studies of solitary foraging insects, such as ants and bees, illustrate sophisticated uses of olfactory, mechanosensory, and magnetic cues even in the absence of visual information (Buehlmann et al., Reference Buehlmann, Mangan and Graham2020). Experiments employing sensory deprivation further highlight remarkable cross-modal plasticity, demonstrating how animals adapt by enhancing or reconfiguring alternative sensory channels (Bavelier & Neville, Reference Bavelier and Neville2002). Such findings underscore that nonvisual modalities are equally fundamental, operating synergistically within sensorimotor loops to shape context-dependent dynamics. An overemphasis on a single modality as vision would miss out on the complexity arising from multimodal sensory integration as a form of representation building. Moreover, control mechanisms for adaptive locomotion in blind insects illustrate as well how internal models for movement and coordination evolved independently of vision (Dürr and Schilling, Reference Dürr and Schilling2018). Such forms of functional internal models already exist as traits in animal systems often assumed to be quite simple, nonetheless these are predictive in their nature (Webb, Reference Webb2004).

When turning the focus toward the control aspects of sensorimotor loops, we share the view of Coombs and Trestman that a hierarchical structure (Merel & Botvinick, Reference Merel, Botvinick and Wayne2019) can be found in motor control structures. These can serve as a scaffold for learning as well as cognition. However, we argue that such a hierarchy should be viewed within the broader context of concurrent sensorimotor loops operating at multiple levels of the hierarchy. Following a bottom-up perspective (Dickinson et al., Reference Dickinson, Farley, Full, Koehl, Kram and Lehman2000)—as often taken in embodied approaches—on a lower foundational level, embodied properties such as muscle elasticities facilitate immediate adaptive responses that, for example, directly provide stability during locomotion without the need for explicit control. Building upon this, reflex pathways contribute rapid, decentralized sensorimotor responses, while more complex coordinated movements arise through central pattern generators and higher-order control mechanisms. Importantly, these structures operate concurrently. The underlying distributed control structures—and the related internal functional models required for coordinated movements as well—are a trait found in many animals. For example, detailed behavioral and neuroscientific studies in stick insects (Schilling et al., Reference Schilling, Paskarbeit, Ritter, Schneider and Cruse2021) have shown that adaptive behavior can emerge from decentralized control systems.

Sensorimotor processing is inherently dynamic, hierarchical, and decentralized. Dickinson et al. (Reference Dickinson, Farley, Full, Koehl, Kram and Lehman2000) highlight that these control structures span multiple timescales and are shaped by recurrent feedback loops between control levels. These structures are not fixed output mechanisms executing predefined sequences. Instead, as discussed above for sensory processing, they actively construct and rely on functional internal models necessary for coordinated movement. Embodied cognition adopts a bottom-up perspective for representation as well, where higher levels of representation are grounded in lower levels of such functional internal models. This introduces forms of abstraction. A key example is the internal body model, required for coordinated movement and sensory integration (McNamee & Wolpert, Reference McNamee and Wolpert2019). These internal models serve an immediate functional role supporting targeted movements. Crucially, in an embodied view, cognition does not stem merely from the existence of such internal models or their predictive capabilities but from the flexible usage of these models outside of their original scope. While these control structures likely evolved to support adaptive behavior, they can be recruited for higher-order functions like planning and flexibly serve different tasks outside their original context. Internal models that support goal-directed behavior and sensory integration are common in many animals. However, the flexible application of such models is less widespread. The ability to recruit internal models for planning has been proposed as constituting cognition (McFarland and Bösser, Reference McFarland and Bösser1993). Neuroscientific evidence from higher animals adds to this view, revealing hierarchical internal models ranging from modality-specific to abstract, schematic representations (Desai et al., Reference Desai, Reilly and van Dam2018), which can be recruited in detailed internal simulations (Schilling et al., Reference Schilling, Paskarbeit, Ritter, Schneider and Cruse2021).

Overall, we propose that a comprehensive perspective on cognitive evolution should emphasize the integration of multimodal sensory inputs, hierarchical sensorimotor loops, and decentralized control structures. It further should focus on functional internal models as well as on how these can be flexibly recruited outside their original context. Such an approach enriches our understanding of cognitive mechanisms across diverse species, complementing and expanding upon Coombs and Trestman’s framework.