AnimatLab

A central goal of neuroscience is to understand how the nervous system is organized to control behavior. Behavior is controlled by neural circuits that link sensory inputs to decision networks, and link decision elements to motor networks and muscles. The dynamics of this interaction are central to the functional control of behavior. Each movement generates its own sensory input, both exteroceptive and proprioceptive, and changes the animal’s position and perspective in the world. To govern behavior correctly, the nervous system must both predict and respond to the consequences of the animal’s own movements and behavior, and do so on a millisecond to second time scale. Despite the importance of this dynamic relationship between nervous function and behavior, it is poorly understood because of technical limitations in our ability to record neural activity in freely behaving animals. The kinematics and dynamics of many behaviors are well understood, and the neural circuitry for behavior patterns in a variety of animals have been mapped and described in anesthetized, restrained animals, or in preparations where the nervous system has been isolated from the periphery. Investigators have then been left to imagine how the operation of neural circuits might produce the behavior patterns observed in the intact animal, but without any way to test those imaginings. 

The escape circuits and behavior of crayfish provide an example of this dilemma. The escape circuits are among the best understood neural circuits in any animal, and for 60 years have provided a model for sensorimotor integration (Fig. 1) (Edwards et al., 1999). Three distinct circuits govern escape of crayfish. The ‘lateral giant’ (LG) circuit responds to an attack on the rear of the animal with a short-latency abdominal flexion that throws the animal up and forward (Herberholz et al., 2004). The ‘medial giant’ (MG) circuit responds to an attack on the front of the animal with a short-latency abdominal flexion that throws the animal directly backward. The ‘non-giant’ (NG) circuits respond to less phasic stimuli with any one of a wide variety of longer latency escape responses that move the animal away from the direction of the attack. Each of these escapes is usually followed by a bout of repeated abdominal flexions and extensions, or ‘swimming’, that carries the animal farther away from the attack and towards shelter.

Despite the extensive knowledge base for these circuits, it remains an open question whether they can account for the actual escape movements of the animal. Although the spatial and temporal patterns of motor neuron activity appear to be appropriate for the initial fast abdominal flexion phase of the escape movements, the resulting patterns of muscle contraction and biomechanical force are unknown. Moreover, the subsequent abdominal re-extension and swimming movements are guided by as yet unknown patterns of exteroceptive and proprioceptive stimuli and central commands. 

AnimatLab was written to address this problem. It provides a software environment in which models of the body and nervous system interact dynamically in a virtual physical world where all the relevant neural and physical parameters can be observed and manipulated. The program contains a ‘body editor’ that is used to assemble a model of the body of an animal (or part thereof) in Lego^TM-like fashion, by attaching different sorts of parts to each other through a variety of joint mechanisms. Muscle attachments, muscles, muscle spindles, touch sensors, and chemical sensors can then be added to provide sensory and motor capabilities. A ‘neural editor’ is used to assemble virtual neural circuits using a variety of model neuron and synapse types. Model sensory neurons can then be linked to the body sensors, and motor neurons can be linked to the Hill model muscles to complete the loop. The body is situated in a virtual physical world governed by Vortex^TM, a physics simulator licensed from CM-Labs, Inc. Simulations can then be run in which the animat’s movements in the virtual environment are under neural control as it responds to simulated physical and experimental stimuli. The autonomous behavior of the animat is displayed graphically in 3-D alongside the time-series responses of any designated set of neural or physical parameters.