An Open-Source Joystick Platform for Investigating Forelimb Motor Control, Auditory-Motor Integration, and Value-Based Decision-Making in Head-Fixed Mice

Mice learn to discriminate between distinct sounds and corresponding actions while refining the kinematic parameters of joystick movements

Here, we introduce a two-alternative forced-choice auditory-motor discrimination task in which animals push or pull a joystick to indicate whether they have heard a high- or low-frequency tone. This setup allows for the analysis of exploratory trajectories, velocity, tortuosity, displacement patterns, and angular motion over the course of learning, providing rich insights into motor behavior and the cognitive processes underlying decision-making. Head-fixed water-restricted mice earn ∼10 μl water rewards by displacing a joystick in response to specific auditory cues. Joystick movements are categorized as anterior (push) or posterior (pull), corresponding to distinct high-frequency (12 kHz) or low-frequency (5 kHz) tones, respectively, each accompanied by five overtones. Reward delivery is controlled via a soundless solenoid valve equipped with a capacitive sensor at the lick spout. Auditory stimuli are presented through a front-mounted speaker controlled by a programmable sound card (Fig. 1A). The static joystick is positioned beneath the mouse's right paw, while the left paw rests on a body tube. This setup forces right-paw use, enabling neuronal contributions to be studied through recordings or manipulations on the contralateral or ipsilateral side relative to the movement.

The task begins with a variable pretrial interval of 2–5 s, followed by a 500 ms auditory cue. Mice are given a 5 s window to perform the correct joystick displacement. Correct responses trigger a 200 ms delay before reward delivery, while omissions result in a 5 s white noise signal and a reset intertrial interval (Fig. 1B, Extended Data Video 1, Movie 1). Mice are trained in daily sessions consisting of two single-association phases: low frequency–pull and high frequency–push. Both associations are trained each day, with the training order alternating daily. Sessions last for a maximum of 30 min or until 100 rewards are obtained, with expert animals completing 200 correct trials and consuming up to 2 ml of water per day.

With task learning, mice significantly reduce their reaction times, as evidenced by a decrease in the time taken to respond to stimuli (paired t test, p < 0.05; Fig. 1C, left). Additionally, during the initial training sessions—when mice only moved the joystick to obtain a reward—an analysis of the maximum number of joystick displacements revealed that most mice exhibited a preference for pulling rather than pushing (Fig. 1C, right).

Joystick movements are recorded in two dimensions (x- and y-axes), enabling the visualization of motor behavior trajectories. Push and pull actions are color-coded (red for pull, blue for push), and a temporal gradient highlights joystick movements 0.5 s before and after reaching the reward threshold. A black dot marks the point at which a choice was registered as either a push or a pull (Fig. 1D). Push actions involve downward–forward joystick displacement, while pull actions are characterized by upward–backward movement. This configuration provides a detailed two-dimensional representation of motor trajectories (Fig. 1D).

Exploratory behavior during learning was assessed by defining a workspace for all mice, based on the minimum and maximum x- and y-coordinates of joystick displacement across all mice. The joystick displacement workspace was binarized into smaller divisions, with each bin measuring 1 mm². The trajectory areas explored were then calculated. Mice (n = 6) showed a significant reduction in the area visited during push movements between the first and last day of training (paired t test, p < 0.0075). In contrast, no significant change was observed in the area visited during pull movements (paired t test, p = 0.15; Fig. 1E).

Performance was evaluated as the ratio of correct trials to the total number of correct trials and omissions. Mice exhibited significant performance improvement after session nine compared with the first day of training (two-way ANOVA, Dunnett's multiple comparison against first day, p < 0.05; Fig. 1F). Mice learn the auditory-motor association in 15 d, excluding 5 d of experimenter habituation during which the animals get used to handling, 2 d of head-fix habituation during which the mice freely drink water rewards while head-fixed, and 3–5 d of joystick association during which any displacement results in a reward, resulting in a month of training.

The differential reduction in the area visited—defined as how much mice explore the joystick movement—between pushing and pulling can be explained by two factors. First, the physical constraints of the pulling motion likely limit its range. Observations indicate that when pulling, mice tend to move the joystick primarily backward and slightly upward, whereas pushing involves more dynamic movements (down and forward), resulting in a larger initial exploration area.

Second, the initial preference for pulling observed in the mice may have reduced the potential for further refinement. Since mice were already more comfortable with pulling, there was less room for improvement in that motion compared with pushing, which had a greater scope for learning and optimization.

To quantify mouse behavior within the joystick workspace, we first identified trial-specific movement trajectories from the session data. Repeated coordinate pairs of the joystick's location were removed, and the coordinates were centered by subtracting the median x- and y-coordinates. Each trajectory was labeled as either a push or pull trial. These trajectories were then used to calculate the average tortuosity, where higher values indicate a longer, more circuitous route from the starting position to the point of maximum displacement and back. Tortuosity was computed as the ratio of the total path length to the Euclidean distance between the first and last points in the trajectory (Mosberger et al., 2024). Session averages of tortuosity revealed that mice initially exhibited high tortuosity, which decreased and stabilized as they became proficient in the task (two-way ANOVA, Dunnett's multiple comparison against first day, p < 0.05; Fig. 2A).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Joystick trajectory dynamics change during auditory-motor integration learning, indicating motor refinement. A, Push and pull tortuosity across training: mice show fewer circuitous routes for both displacements. B, Mean trajectory similarity increases across sessions, indicated by lower Fréchet distance over time. C, Velocity to reach the displacement threshold decreases during the first days of training. D, Joystick exploratory area visits decrease for push displacements as learning progresses. E, Mean angular deviation (MAD) remains stable throughout learning. n = 5. Vertical solid lines indicate p < 0.05.

To further analyze joystick displacement dynamics, we compared the pairwise similarity of joystick trajectories using the Fréchet distance (Ursell, 2025). For each session, we evaluated joystick displacement and measured the similarity across all possible combinations of trajectories and calculated a mean value per session. Results showed that mice increased movement consistency over time, reflected by a significant reduction in the Fréchet similarity index (two-way ANOVA, Dunnett's multiple comparison against first day, p < 0.05; Fig. 2B).

We also measured the velocity of joystick motions, defined as the distance between the first point and the point of maximum displacement divided by the corresponding time interval. Mice demonstrated a significant increase in movement velocity compared with their performance on the first day of training (two-way ANOVA, Dunnett's multiple comparison against first day, p < 0.05; Fig. 2C).

To define the joystick workspace, we calculated the minimum and maximum x- and y-coordinates across all animals. The explored area within this workspace was quantified by binning the joystick coordinates using MATLAB's “histcounts2” function, as described by Mosberger et al. (2024). Each bin measured 1 mm by 1 mm. The total explored area was calculated by summing the number of visited bins. Over successive training sessions, mice showed a significant reduction in the area explored, which eventually stabilized at a lower value, indicating reduced exploratory behavior only for the push displacement (two-way ANOVA, Dunnett's multiple comparison against first day, p < 0.05; Fig. 2D).

To assess directional consistency, we calculated the mean angular deviation for both push and pull motions using the CircStat MATLAB Toolbox for circular statistics (Berens, 2009). Mean angular deviation was calculated by taking the average of the angular deviation for the angles in each bin for each session. Angular deviation, ranging from 0 to √2, represents variability in directional movements, with higher values indicating greater variability. The angular deviation remained stable throughout training (Fig. 2E).

Together, these results demonstrate that mice effectively moved the joystick in two distinct directions, decreased trajectory tortuosity, increased movement similarity and velocity, and refined their displacement strategy by reducing the explored workspace. This evidence supports the task as a robust tool for analyzing motor output dynamics, offering high-quality, detailed behavioral data.

This auditory-motor discrimination task provides a robust framework for studying neural and behavioral mechanisms underlying auditory-motor associations, offering key insights into sound-driven action selection and motor learning.

Joystick kinematic parameters reflect total and relative value in value-based decision-making task

Here, we describe a two-armed bandit, joystick-based value-based decision-making task in mice that allows for the study of value-based modulation of motor execution, in addition to recapitulating known characteristics of conventional value-based 2-AFCs.

Head-fixed water-restricted mice obtain 10% sucrose solution rewards via anterior or posterior displacement of a retractable joystick. Reward is delivered via an optical lickometer setup that also tracks licking (Fig. 3A). At trial start, the joystick is made available to the mouse via anterior motion of the servo motor. Following a subsequent 100 ms wait period, a GO cue light on the lickometer signals the start of a 10 s window during which the mouse can register a choice via anterior (push) or posterior (pull) displacement of the joystick (Fig. 3B, Extended Data Video 2, Movie 2). There are four possible outcomes of a trial: (1) rewarded trial, followed by 2.5–8 s intertrial interval; (2) unrewarded trial, signaled by turning off house light, followed by 2.5–8 s intertrial interval; (3) omission, signaled by white noise, turning off house light, and 15 s time-out; and (4) premature trial in which mouse registers choice before Go cue, signaled by white noise, turning off house light, and 15 s time-out.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Mice integrate reward evidence across trials to guide next-trial choice and motor vigor. A, Photograph of our behavioral setup. B, Behavioral schematic highlighting two task variants. C, Males (n = 8) and females (n = 6) prior outcome to guide next trial strategy. D, Logistic regression weights of prior trial rewarded (undashed) and unrewarded (dashed) outcomes for predicting current trial repetition of prior choice. E, Raw joystick choice data. F, Demonstration that varying reward volume significantly shapes motor vigor, as measured by peak joystick displacement and, predominantly in females, choice latency.

Mice are trained in a sequential behavioral paradigm consisting of (1) probabilistic reversal and (2) changing volume phases (Fig. 3B). The probabilistic reversal phase consists of blocks in which one of two choices is more likely to be rewarded than the other (push blocks and pull blocks), with reward probabilities of 80%:20%. Each block has a minimum duration of 17 rewarded trials plus a geometrically distributed random variable (p = 0.4), after which the high reward probability side is reversed in an un-cued manner. As in the auditory-motor discrimination task, mice only register choices with their right forepaw. Mice in this phase integrate prior-trial evidence to guide next-trial decisions as evidenced by win-stay/lose-switch analysis as well as logistic regression of choice and reward history (Fig. 3C,D). The changing volume phase builds on this task structure by adding reward volume as an additional parameter that varies by block, using reward volumes of 2, 4, and 8 μl. Mice register choices with varying latencies and peak joystick displacement (Fig. 3E). We find that, in higher total value contexts, mice register choices with shorter choice latency and higher peak joystick displacement (Fig. 3F). Unlike choice latency, the peak displacement phenotype is robust across males and females, suggesting that a joystick-based design offers unique, key insights into animals’ regulation of motor vigor based on total value as compared with conventional, binarized lever press- or lick-based tasks (Wang et al., 2013; Alabi et al., 2020).

An advantage of our joystick-based value-based decision-making paradigm over conventional lever-press or lick-based paradigms in mice is the ability to read out kinematic parameters of operant choice. Vigor is known to reflect real-time internal value representations (Takikawa et al., 2002; Shadmehr et al., 2019).

In this task, we find that mice's joystick trajectories often exhibit extensive deliberation before crossing choice threshold (Fig. 4A). Anteroposterior joystick position relative to baseline can be plotted as a function of time and segmented into movement bouts to capture these deliberative movement bouts leading up to a threshold-crossing decisive bout. We defined bouts according to the following criteria: (1) movement is in one direction, (2) initiation speed is >7.5 mm/s, (3) speed is maintained at >2.5 mm/s for >50 ms, (4) bout ends with velocity sign change or joystick retracting. Of note, unlike the auditory-motor discrimination task, our value-based decision-making task constrains motion to the anteroposterior axis via addition of a metal bar underneath the joystick, limiting up-down joystick displacement.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Readouts of joystick trajectory in different relative value contexts reflect animal uncertainty. A, Segmentation of joystick position trace into movement “bouts” based on joystick velocity. B, Representative traces from high uncertainty (low |ΔQ|) and low uncertainty (high |ΔQ|) contexts. In higher-uncertainty contexts, mice register choices with greater mean velocity (D), greater number of movement bouts (E), lower consistency of direction of joystick motion across bouts (F), and greater path length (G). Peak joystick displacement is not significantly affected by degree of uncertainty (C). *p < 0.05; **p < 0.01; ***p < 0.001.

A range of kinematic parameters can be extracted from joystick movement traces. Peak displacement is defined as the maximum extent of displacement of the joystick away from baseline on a given trial. Number of bouts is defined as the number of movement bouts the animal initiates in a given trial. Directional consistency is the proportion of these bouts that occur in the higher frequency direction (Eq. 1), with a value of 1 implying that all bouts occur in one direction. Mean velocity is defined as the mean velocity of the joystick on the decisive bout. Path length is defined as the distance travelled by the joystick in any direction on a given trial.Directionalconsistency=n(boutsinmostfrequentdirection)/n(boutsinalldirections).

We found that our readouts of joystick trajectory reflect animals’ internal representations of certainty that one joystick direction is more likely to yield reward than another, i.e., relative value. Using a two-parameter Q-learning model with nondifferential forgetting (Ito and Doya, 2015; Choi et al., 2023), we generated trial-by-trial estimates of animals internal representation of the value of push and pull actions (Q_push and Q_pull). We computed the absolute value of the difference between Q_push and Q_pull (ΔQ = Q_push − Q_pull) to gauge the animals’ experienced uncertainty on a given trial, where low |ΔQ| implies more similar value representations of push and pull actions, and therefore greater uncertainty regarding which choice is more likely to be rewarded. Movement trajectories in high |ΔQ| (low uncertainty) and low |ΔQ| (high uncertainty) contexts are distinct, as is illustrated in example traces (Fig. 4B). In lower |ΔQ| trials, we found that animals trended toward lower peak displacement (Fig. 4C) and had significantly greater mean joystick velocity (Fig. 4D), a significantly greater number of joystick movement bouts (Fig. 4E), significantly lower directional consistency (Fig. 4F), and significantly greater path length (Fig. 4G), reflective of greater uncertainty. Our joystick kinematic parameters thus provided a key additional insight into animals’ dynamic representations of relative value.