Efficient visual learning by bumble bees in virtual‐reality conditions: Size does not matter

Recent developments allowed establishing virtual‐reality (VR) setups to study multiple aspects of visual learning in honey bees under controlled experimental conditions. Here, we adopted a VR environment to investigate the visual learning in the buff‐tailed bumble bee Bombus terrestris. Based on responses to appetitive and aversive reinforcements used for conditioning, we show that bumble bees had the proper appetitive motivation to engage in the VR experiments and that they learned efficiently elemental color discriminations. In doing so, they reduced the latency to make a choice, increased the proportion of direct paths toward the virtual stimuli and walked faster toward them. Performance in a short‐term retention test showed that bumble bees chose and fixated longer on the correct stimulus in the absence of reinforcement. Body size and weight, although variable across individuals, did not affect cognitive performances and had a mild impact on motor performances. Overall, we show that bumble bees are suitable experimental subjects for experiments on visual learning under VR conditions, which opens important perspectives for invasive studies on the neural and molecular bases of such learning given the robustness of these insects and the accessibility of their brain.


Introduction
Social insects have long attracted the interest of biologists due to their complex social organization and sophisticated communication codes, but also because some species have emerged as useful models for the study of cognitive capacities in small brained animals (Roussel et al., 2009;Giurfa, 2013;Giurfa, 2015).Such is the case of the honey bee Apis mellifera, an insect which exhibits remarkable learning and memory (Giurfa, 2007;Avarguès-Weber et al., 2011;Giurfa & Sandoz, 2012), which are the basis of its navigation skills and of flower constancy, that is, the fidelity to a flower species that is exploited through repeated foraging visits as long as it remains profitable in terms of nectar and/or pollen reward (Grant, 1950;Waser, 1986;Chittka et al., 1999).Decades of research on honey bee learning and memory have led to an important knowledge on the behavioral, cellular, and neural mechanisms of these capacities (Giurfa, 2007;Giurfa & Sandoz, 2012).The facility to raise honey bee colonies in apiaries and the establishment of multiple behavioral protocols for the controlled study of honey bee behavior, together with the development of invasive techniques to access and characterize the functional principles of the honey bee brain, justifies the historic success and the model status of the honey bee in the field of neurobiological studies on learning and memory (Giurfa, 2007).Yet, this intensive focus has led to a neglect of multiple social insect species, which differ in their social characteristics and life style, and which may thus enlighten comparative analyses of learning and memory and their underpinnings.Enlarging the spectrum of species for which a level of knowledge similar to that reached for the honey bee is therefore an important pending task.
The last decade has, nevertheless, witnessed an emergence of multiple studies focusing on bumble bees as an alternative model for the study of social insect behavior, including learning and memory (Goulson, 2003).Various species of bumble bees Bombus sp. have been used in studies focusing on cognitive problems because they ally the possibility of successful learning in behavioral protocols (Riveros & Gronenberg, 2009) with a relatively easy maintenance in the laboratory, where colonies can be purchased and kept without excessive investment (Goulson, 2003;Marder et al., 2010).
Visual learning and memory have been studied in bumble bees using protocols in which the bees freely fly in experimental arenas where they have to learn to discriminate rewarding from aversive options.This experimental approach has led to important discoveries about invertebrate cognition such as the presence of speed-accuracy trade-off in foraging decisions (Chittka et al., 2003;Dyer & Chittka, 2004a), cross-modal object recognition (Lawson et al., 2018;Harrap et al., 2019;Solvi et al., 2020) and social observation learning from conspecifics (Worden & Papaj, 2005;Leadbeater & Chittka, 2007;Leadbeater & Chittka, 2009;Avarguès-Weber et al., 2013), among others.Yet, as in the case of honey bees, in which the study of visual cognitive abilities was mainly restricted to free-flying animals, the study of visual learning in bumble bees has suffered from the fact that using free-flying insects precludes the access to the brain given the impossibility of coupling behavioral recording with invasive procedures.This limitation is regrettable as a detailed knowledge of the visual circuits in the bumble bee brain was achieved using electrophysiological recordings in immobilized animals (Paulk & Gronenberg, 2008;Paulk et al., 2008;Paulk et al., 2009a).
A recent development allowing to overcome this limitation in honey bees is the use of virtual-reality (VR) setups, in which these insects achieve visual discriminations while being tethered from the thorax and walking stationary on a treadmill (Schultheiss et al., 2017).The tethered honey bee is exposed to a controlled visual environment, which varies according to the walking movements of the bee on the treadmill, thus creating an immersive environment.The trajectory of the bee in this VR setup can be reconstructed by means of captors placed lateral or ventral to the treadmill.In this environment, honey bees have been trained successfully to discriminate visual targets differing in color and/or shape based on their different reinforcement outcomes (Buatois et al., 2017;Rusch et al., 2017;Schultheiss et al., 2017;Buatois et al., 2018;Zwaka et al., 2018;Rusch et al., 2021).Visual discriminations were either elemental (Buatois et al., 2017;Buatois et al., 2018) (i.e., based on simple associative learning) or nonelemental (Buatois et al., 2020) (i.e., based on configural learning) and their characterization could be combined with an identification of the different areas of the brain involved in these forms of problem solving using invasive approaches such as electrophysiology (Zwaka et al., 2018;Rusch et al., 2021), neuropharmacology (Rusch et al., 2021) and quantification of immediate early genes (Geng et al., 2022;Lafon et al., 2022), among others.The full control offered by VR allows referring neural and molecular events recorded under these circumstances to the specific problem being solved by the insects.Yet, among social insects, this research approach based on tethered individuals has been so far only pursued in the case of honey bees.Experiments on visual performances in a VR environment were performed in bumble bees (Frasnelli et al., 2018) but using free-flying individuals, which precludes the coupling with invasive methods to quantify neural activity.Thus, whether VR-based visual analyses, including on visual cognition, are extensible to other bee species walking stationary while being tethered remains so far unexplored.
To fill this void, we extended our studies of visual learning in bees under VR conditions to the buff-tailed bumble bee Bombus terrestris.We evaluated its capacity to learn visual discriminations in these restrictive conditions and thus its potential use for future experiments coupling behavioral analyses in VR conditions with invasive analyses of brain function.In addition, we investigated how body size variation, a characteristic trait linked to division of labor in bumble bee colonies (Jandt & Dornhaus, 2009;Jandt & Dornhaus, 2014) may affect visual learning in VR conditions.Previous studies found that body size is correlated with visual learning speed in freely flying bumble bees (Worden et al., 2005;Riveros & Gronenberg, 2012;Klein et al., 2017) and with the ability to negotiate obstacles using visual cues (Ravi et al., 2020;Giurfa & Luyat, 2021).This could be explained by differences in perceptual abilities (Riveros & Gronenberg, 2009) or brain size variation (Smith et al., 2016).Here we tested whether bumble bees learn visual discrimination in a VR environment to assess their suitability for further experiments employing this approach, and whether larger bumble bees would be more efficient at solving visual discriminations in this environment.

Study species and collection
Bumble bees B. terrestris from 4 commercial colonies (Koppert, Cavaillon, France) hosted in a laboratory room were collected each morning around 09 : 00 hours during November 2021.Each colony contained about 200 workers, brood, and 1 queen.Bees did not forage outside the box.Female worker bees were collected by placing a glass vial at the entrance of the nest box, which enclosed the workers leaving the nest.bumble bees were maintained and tested in the laboratory at 25 °C and 30%-40% relative humidity, under a 12 : 12 h light : dark photocycle.

Experiment 1: Appetitive and aversive reinforcement responses
In honey bees, sensitivity for aversive gustatory reinforcements is unclear (Ayestarán et al., 2010;Bestea et al., 2021).We thus aimed at verifying that the appetitive and aversive gustatory reinforcers used for visual conditioning-1.5 mol/L sucrose solution and 3.0 mmol/L quinine solution, respectively-were effective.In order to determine if bumble bees of our colonies showed appropriate responsiveness to these stimuli, we captured and enclosed them individually in 20 mL syringes with a perforated wall to facilitate respiration (Fig. 1A).Bees were weighed within the syringe and their individual weight (mg) was obtained by subtracting the weight of the empty syringe.This procedure allowed to refer food consumption to the initial weight of each individual.Enclosed bees were divided into 3 groups.One group was offered distilled water, another group received 3.0 mmol/L quinine solution and the remaining group received 1.5 mol/L sucrose solution.Solutions were delivered in an Eppendorf tip inserted in the open hub of the syringe so that the bee within the syringe could feed from it (Fig. 1B).bumble bees were kept for an hour under these conditions.For each individual, we measured its consumption (μL) of sucrose solution, quinine or distilled water at the end of this period (Desmedt et al., 2016;Bestea et al., 2022).

Experiments 2 and 3: General methods
Tethering procedure bumble bees captured for the VR experiments were anesthetized on ice for 5 min to facilitate tethering and subsequent positioning on a spherical treadmill (a Styrofoam ball).Bees were handled under red light, which ensured a dark environment to the insects.The absolute-irradiance spectrum (W/cm 2 /nm) of this light is shown in Suppl.Fig. S1, together with the spectral sensitivity curves of bumble bee photoreceptors (Skorupski et al., 2007).The red light was turned on when the animals were handled prior to the experiments.Otherwise, the room was dark and the only illumination available came from the video projector used in the VR setup (see below).
A 0.06 g steel needle, 0.5 mm in diameter and 40 mm in length, was fixed to the thorax of the bee by means of melted bee wax (Lafon et al., 2021).The needle was placed within a 3D-printed resin tube (Black tough resin Prusa, Prusa Research a. s., Czech Republic), 7 mm inner diameter, 1 cm outer diameter, and 55 mm in length, which was fixed on a holding frame placed above the treadmill (Fig. 1C).This system allowed the bee to adjust its position in the vertical axis once placed on the ball but did not allow rotational movements (Fig. 1D).The holding frame consisted of a vertical black, plastic half frame made of 2 vertical rectangular supports, 105 mm in length, connected to an upper, horizontal rectangular support, 120 mm in length.The latter held the black cylinder in the middle (Fig. 1D).After being attached to its tether, and before starting the experiments, each bee was placed on a 50 mm diameter Styrofoam ball for familiarization during 3 h and provided with 5 μL of 1.5 mol/L sucrose solution.This provisory setup was maintained in the dark to deprive the bees of visual stimulations.
Virtual-reality setup.After the 3-h period, the bees were moved to the VR setup (Lafon et al., 2021).The VR relied on a custom software developed using the Unity engine (version 2020.3.4f1), which updates the position of the bee within the VR every 0.017 s.The open-source code is available at https://github.com/G-Lafon/BeeVR.In addition, it includes information about troubleshooting and about the customization of the 3D models used in the VR setup.
The VR apparatus consisted of a spherical Styrofoam ball (diameter: 50 mm, weight 1.07 g), floating on an air cushion.The ball was positioned within a 3D-printed, hollow, cylindrical support (cylinder: 50 mm high, 59 mm diameter).The cylinder allowed distributing an upward airflow of 33 L/min produced by an AquaOxy 2000 aquarium pump and released through a small hole at the base of the cylindrical support at a pressure of 1.221 bar.The movements of the ball were recorded by 2 infrared optic-mouse sensors (Logitech M500, 1 000 dpi) placed at a distance of 7 mm from the sphere and forming an angle of 90°relative to each other (i.e., 45°from the bee body axis; see Fig. 1C).In this way, ball movements due  and 3).1: Plastic cylinder held by the holding frame; the cylinder contained a glass cannula into which a steel needle was inserted.2: The needle was attached to the thorax of the bee.3: Its curved end was fixed to the thorax by means of melted bee wax.(E) Absolute irradiance (W/cm 2 /nm) of the lights used to assess spontaneous color preference of bumble bees in the VR setup (Experiment 2).In order to find a pair of colors that elicited the same amount of attraction, we tested 2 pairs of green and blue virtual cylinders.In 1 pair, 1 cylinder was Green 1 (R: 0, G: 100, B: 1.0; dominant wavelength: 528 nm; irradiance: 300 μW/cm 2 ) and the other Blue 1 (R: 0, G: 80, B: 254; dominant wavelength: 449 nm; irradiance: 2500 μW/cm 2 ) while in the other pair, 1 cylinder was Green 2 (R: 0, G: 51, B: 1.0; dominant wavelength: 527 nm; irradiance: 130 μW/cm 2 ) and the other Blue 2 (R: 0, G: 0, B: 255; dominant wavelength: 446 nm; irradiance: 2700 μW/cm 2 ).The right, green axis refers to Green 1 and Green 2. The blue, left axis refers to Blue 1 and Blue 2. (F) Location of the color used in Experiment 2 in the color hexagon model using Bombus terrestris photoreceptor sensitivity.Photoreceptor excitations (S: short-wave; M: mid-wave; L: long-wave) are plotted at an angle of 120°.Within each color pair, colors were well discriminable as shown by their distant loci in the color space.
to the displacements of the bee could be recorded and walking trajectories reconstructed.
The ball was positioned in front of a half-cylindrical vertical screen, 268 mm in diameter and 200 mm height, which was placed at 9 cm from the bee.The screen was made of semitransparent tracing paper, which enabled projecting a 180°visual environment to the bee (Fig. 1C).The projection was achieved by means of an Acer K135 video projector (Lamp: LED, maximum vertical sync: 120 Hz, definition: 1 280 × 800, minimum vertical sync: 50 Hz, brightness: 600 lumens, maximum horizontal sync: 100.10 3 Hz, contrast ratio: 10 000 : 1, minimum horizontal sync: 30.10 3 Hz), which was connected to a laptop.The delay between the motion of the bee and the update of the visual surrounding was 18.00 ± 2.53 ms (mean ± SE; n = 10) (Lafon et al., 2021).
Virtual colored cylinders had a 5-cm-diameter base and occupied the entire vertical extent of the screen irrespective of the position of the bumble bee.At the beginning of each test, each cylinder subtended a horizontal visual angle of 6.5°and was positioned either to the left (−50°) or the right (+50°) of the tethered insect placed on the treadmill.When the bee walked toward the cylinder, its horizontal extent expanded at a rate of 1.7°/cm.A choice was recorded when the bee approached the cylinder within an area of 3 cm surrounding its virtual surface and faced directly at its center.
Each insect was subjected to 2 consecutive tests in nonreinforced conditions, each with a different pair of green and blue cylinders (see above), to evaluate its spontaneous preference.Each test lasted 180 s and the intertest interval was 10 s.The order of presentation of the color pairs and the side of each color were randomized between tests.In each test, we recorded the color first chosen by each bumble bee.

Experiment 3: Differential color conditioning
Having chosen a pair of colors that elicited the same amount of spontaneous attraction (see above), we trained bumble bees to discriminate 2 cylinders differing in color and in reinforcement type (see training procedure below).The cylinders had the same initial positioning and angular size as those displayed in the previous experiment; 1 was Green 2 (see above) and the other Blue 2 (see above).
Bumble bees were trained during 6 conditioning trials using a differential conditioning procedure in which 1 of the cylinders (either Green 2 or Blue 2) was paired with 1.5 mol/L sucrose solution (the appetitive conditioned stimulus or CS+) while the other cylinder was paired with 3.0 mmol/L quinine solution (the aversive conditioned stimulus or CS−).The concentration of quinine solution was chosen based on previous reports on visual learning by free-flying bumble bees where it was used as aversive reinforcement (e.g., Chittka et al., 2003;Dyer & Chittka, 2004b).Both CS were delivered to the insect by means of a toothpick soaked in either solution.Color contingencies (i.e., blue and green equally rewarded across bees) were based on a random assignment by the VR software and were balanced across bees.As in the previous experiment on spontaneous preferences, a choice was recorded when the bee reached the CS+ cylinder within a surrounding area of 3 cm (i.e., the cylinder subtended a horizontal visual angle of 53°) and centered it on the screen.In this case, the screen was locked for 14 s, which allowed the delivery of the sucrose solution.If the bee chose the CS− cylinder, the choice was recorded as incorrect and the screen was locked again to facilitate the delivery of the quinine solution.Solutions were delivered by the experimenter sitting behind the bee by means of a toothpick contacting the antennae and the proboscis during 3 s.If the bee did not extend its proboscis (e.g., in the case of quinine solution), the toothpick was set in contact with its mouthparts during 3 s.
At the beginning of the conditioning sequence, bees were presented with a dark screen during 60 s.During training trials, each bee was presented with the 2 cylinders and it had to learn to choose the CS+ cylinder by walking toward it and centering it on the screen.Each training trial lasted until the bee chose one of the stimuli or until a maximum of 180 s in which case a "no choice" (NC) was recorded.Trials were separated by an intertrial interval of 60 s during which the dark screen was presented.Bees that were unable to choose a stimulus in at least 3 trials were excluded from the analysis.From the 315 trained bumble bees, 235 were kept for the analyses (∼75%).
Sixty seconds after the 6th and last training trial, each bee was subjected to a nonreinforced retention test, which contrary to the training trials, had a fixed duration of 180 s.During this test, 2 variables were recorded: the first choice (as defined above) and the time spent fixating the rewarded and the nonrewarded cylinder.Both variables have been used to characterize honey bee performance in prior works performed in our VR setup as they reveal different aspects of learning and memory retention (Buatois et al., 2017;Buatois et al., 2018;Buatois et al., 2020;Lafon et al., 2021;Lafon et al., 2022).Fixation time (s) was defined as the time spent by each cylinder at the center of the screen (± 2.5 mm) where it was brought by the bee's actions (Lafon et al., 2021).A ray-casting approach was used to decide if the cylinder was indeed at the center of the screen, which corresponded to the intersection between a ray corresponding to the forward vector of the moving bee and the center of the cylinder (Lafon et al., 2021).At the end of the final test, each animal was frozen at −20 °C to allow weighing and size measurements.

Body size and weight measuring
Body size was assessed by measuring the distance between the 2 wing joints (Cane, 1987;Ravi et al., 2020) using the Toupview software (ToupTek Photonics, Zhejiang, China).After size measurements, insects were placed in an oven at 70 °C for 4 h in order to evaporate all water from their bodies.Dry weight (Sage, 1982) was measured with a precision scale (OHAUS, Nänikon, Switzerland).

Statistical analysis
Statistical analyses were performed using R software, version 4.0.2(R Development Core Team, 2016).
In Experiment 1 (food consumption), consumption of water, sucrose, and quinine solutions (μL/bee) were compared using a one-way analysis of variance (ANOVA) based on the null hypothesis of an equivalent consumption for all 3 solutions delivered.Tukey's tests were used for post hoc comparisons.The relation between solution consumption and body weight was analyzed by means of a Pearson's correlation (R) as data were continuous and normally distributed.In Experiment 2 (spontaneous color choice), we counted the number of green and blue choices for each color pair and tested the null hypothesis of a random distribution of choices (frequency of 0.5) between colors using a χ 2 test.In Experiment 3 (differential conditioning of colors), the first choice in each trial and in the retention test was categorized as choice of the CS+, of the CS− or no choice (NC).Thus, a bee choosing the CS+ was recorded as (1, 0, 0) for choice of the CS+, CS−, and NC, respectively.A bee choosing the CS− was recorded as (0, 1, 0) while a bee not making a choice was recorded as (0, 0, 1).Data were bootstrapped to plot the proportion of bees in each category with their corresponding 95% confidence intervals.Learning was analyzed using generalized mixed linear models (GLMM) with a binomial error structure-logit-link function (glmer function of R package lme4) (Bates et al., 2015).The in-dependent variables (fixed factors) were the trial number (Trial, 1 to 6), the choice category (Choice) and the color of the CS+ (Color, Blue or Green).Individual Identity was included as a random factor to account for the repeated-measure design.Post hoc ANOVAs were performed on those models to assess the impact of each factor.During the retention tests of Experiment 2, in addition to the color choice of each bee, we recorded the time spent fixating each color (CS+ vs. CS−).Fixation times were compared using a Wilcoxon signed-rank test.
During acquisition trials, we also recorded the total Distance Walked during a trial, and the Walking Speed of each bee.In addition, we recorded the Latency to Make a Choice, starting from the beginning of a trial to the moment in which a choice (either for the CS+ or the CS−) was made.NC data were excluded from the latency analysis.The analysis of these continuous variables was done using a linear mixed model (lmer function of R package lme4), in which Individual Identity was a random factor and Trial Number (Trial, 1 to 6) was a fixed factor.
To explore the effect of size and weight on learning performances, we defined a learning score as the number of CS+ choices minus the number of CS− choices from the second trial to the test.The first trial was excluded as choice in that trial could only be random in the absence of reinforcement experience.Then, we analyzed the potential effect of Size and Weight on learning scores using a Spearman's rank correlation.We used the same approach to assess the effect of Size and Weight on Latency to Make a Choice.To analyze the effect of Size and Weight on Speed and Distance, we used a linear model (lm function R package stats), in which Size and Weight were fixed factors.Post hoc ANOVAs were performed to assess the impact of each factor.Due to the multiple comparisons performed in the analysis of the effects of Weight and Size, Bonferroni correction was applied, and we thus use α = 0.00625 for these analyses.

Experiment 1: Appetitive and aversive reinforcement responsiveness
Prior to the VR experiments, we determined the individual consumption of 1.5 mol/L sucrose solution (n = 21), 3.0 mmol/L quinine solution (n = 21) and distilled water (n = 20) by enclosing bumble bees in individual syringes providing these solutions for 1 h (Fig. 1A, B).Fig. 2B shows that the consumption of all 3 groups differed significantly (one-way ANOVA: F 2,59 = 57.22,P < 0.0001).The consumption of sucrose solution was significantly higher than that of quinine solution (Tukey's test, P < 0.001) and distilled water (P < 0.001), while the consumption of quinine solution and water did not differ (P = 0.12, n.s.).Consumption of sucrose solution was variable as it correlated with the initial weight of bumble bees differing in size (Fig. 2C); large bumble bees consumed more sucrose solution than smaller bumble bees (Pearson's R = 0.5175, P = 0.016) while the consumption of water and quinine solution was unaffected by body weight (R = −0.3235,P = 0.164 and R = 0.0496, P = 0.831, respectively).Overall, these results indicate that the bumble bees used in our experiments had the appropriate motivation toward the appetitive and aversive reinforcements delivered in the VR setup.bumble bees preferred sucrose solution to water, thus showing that this preference was selective and was not driven by thirst.

Experiment 2: Choosing colors with spontaneous equal attraction in the VR setup
We then recorded the spontaneous choice of bumble bees presented with 2 shades of green and 2 shades of blue in the VR setup: Green 1 versus Blue 1 and Green 2 versus Blue 2. The irradiance difference between green and blue was 10 times higher for Blue in the first pair and 26 times higher in the second pair.Each bee was presented a single time with each color pair following a random sequence that varied from bee to bee.As this sequence had no significant effect on the choice of bees (z = 1.094,P = 0.27), we pooled the data according to the color pair.

Experiment 3: Differential color conditioning
Having calibrated the light stimuli to be used in the VR setup, we subjected our bumble bees to a differential color conditioning.Bumble bees had to learn to discriminate between a Green 2 and a Blue 2 cylinder presented against a dark uniform background during 6 conditioning trials, each lasting up to 180 s.One color (CS+) was rewarded with 1.5 mol/L sucrose solution while the other (CS−) was punished with 3 mmol/L quinine solution.A nonreinforced test followed the last conditioning trial.In all cases, we recorded the first choice of the bees, either for the CS+ or the CS−, or if the bees did not make any choice (NC).
Choices during conditioning trials were not affected by the association of a particular color with reward or punishment (χ 2 = 5.57, df = 2, P = 0.62).Therefore, we pooled performances according to choices of the CS+ and the CS−, irrespective of their color.Fig. 4 shows that CS+ choices evolved differently from both CS-choices (z = −1.46,P < 0.0001) and nonchoices (NC) (z = 5.22, P < 0.0001), consistently with the fact that bees learned to respond more to the CS+ than to the CS−.Choices of the CS− did not evolve differently from NC (z = 1.458,P = 0.14).The analysis of the trajectories followed by the bumble bees in the VR setup reported as heat maps in Fig. 4B showed that bees walked toward the cylinders and interacted with them during the conditioning trials (see V-shaped cumulative trajectories).A more detailed analysis of trajectory parameters during conditioning trials showed that the distanced walked decreased significantly along trials (Fig. 4C; χ 2 = 271.62,df = 1, P < 0.0001) as bumble bees took more direct paths toward the cylinders.Speed increased significantly with trials (Fig. 4D; χ 2 = 10.08,df = 1, P < 0.01) in parallel to the increase in the proportion of CS+ choices.Finally, the latency to make a choice decreased significantly across trials (Fig. 4E; χ 2 = 90.25,df = 1, P < 0.0001), consistently with an improvement in choice performance and the execution of faster and direct choices.
After the last training trial, each bee was subjected to a nonreinforced test, in which choice of the green and the blue cylinders was recorded during 180 s.During the test, we quantified the percentage of bees choosing first either the CS+ or the CS−, or not making any choice (NC), as well as the time spent fixating the CS+ and the CS−.Fig. 5A shows that the proportion of bees choosing first the CS+ was significantly higher than that of bees choosing first the CS− or not making any choice (C+ vs. CS−: z = 7.59, P < 0.0001; CS+ vs. NC: z = 10.96,P < 0.0001).The proportion of bumble bees not making a choice (NC) was also lower than that of bumble bees choosing the CS-(z = 5.62, P < 0.00001), suggesting that, in general, bees kept the motivation to interact with the virtual cylinders in the VR setup.Consistently with these data, bees fixated the CS+ for longer than the CS− (Fig. 5B; V = 5781, P < 0.00001).Nevertheless, trajectories in the corresponding cumulative heat map (Fig. 5C) show that bumble bees explored and interacted with both cylinders during the test.
Overall, these data indicate that bumble bees learned efficiently the color discrimination in the VR setup and that they retained the learned information at least in the short delay.

Effect of size polymorphism on visual performances
To test the influence of body size on visual performances, we first established a learning score for each bumble bee.The score was computed as the number of CS+ choices minus the number of CS-choices from the second trial to the test.The interwing distance of the bee did not significantly affect these learning scores (Fig. 6A; S = 2293709, ρ = −0.06,P = 0.36).Similarly, dry weight had no significant effect on learning scores (Fig. 6B; S = 2336901, ρ = −0.08,P = 0.22).

Discussion
The present results show that bumble bees had the proper appetitive motivation to engage in the VR experiments as they consumed the positive reinforcement of sucrose solution while they rejected the quinine solution.Bumble bees learned efficiently elemental visual discriminations in the VR context.In doing so, they reduced the latency to make a choice, increased the proportion of direct routes toward the virtual stimuli and walked faster toward them.Performance in a short-term retention test showed that bumble bees chose correctly the stimuli learned in the absence of reinforcement and that they fixated longer on the correct stimulus.Size and weight, although variable across individuals, did not affect cognitive performances (e.g., learning score) and had a mild impact on motor performances as the only significant effect was related to the walking speed, which was faster in larger individuals.

Bumble bees as a model for visual research under VR conditions
These results confirm that bumble bees are a suitable model for experiments on visual learning under tethering conditions in a VR environment.Visual learning in several species of bumble bees, including under VR conditions (Frasnelli et al., 2018), has been mostly restricted to free-flying individuals performing different kinds of visual discriminations in mazes or experimental arenas (e.g., Chittka, 1998;Spaethe et al., 2001;Chittka et al., 2003;Dyer & Chittka, 2004a;Solvi et al., 2016;Lawson et al., 2018;Harrap et al., 2019).In this sense, research on visual cognition has faced the same problem as that experienced for many decades by studies on this topic in honey bees, namely the impossibility of developing invasive analyses of neural activity in freely flying animals.Yet, bumble bees may offer advantages for research under VR conditions compared with honey bees (Buatois et al., 2017;Rusch et al., 2017;Schultheiss et al., 2017;Buatois et al., 2018;Zwaka et al., 2018;Lafon et al., 2021;Rusch et al., 2021;Geng et al., 2022;Lafon et al., 2022).For instance, bumble bees seem to be more resilient to the potential stress induced by the tethering condition, and thus more disposed to engage in the VR experiments than honey bees.In the present conditioning experiments, from 315 trained bumble bees, 235 were kept for the analyses (∼75%).In the case of honey bees subjected to the same kind of training in a similar setup but with motion cues provided by the background (Lafon et al., 2021), from 216 and 272 honey bees trained in 2 different experiments, 75 (∼35%) and 67 (∼25%), respectively, were kept for the analyses.These different percentages should be taken cautiously as they were not obtained in simultaneous experiments.Yet, they seem to indicate a potential advantage of bumble bees for VR experiments.
A fundamental argument for developing VR setups in the case of honey bees was the contrast between the richness of their visual cognitive performances in free-flight conditions and the limited visual performances exhibited under immobility conditions (Avarguès-Weber et al., 2012).Indeed, harnessed bees perform rather poorly in the protocol of visual conditioning of the proboscis extension reflex, in which visual stimuli are paired with sucrose solution delivered to the antennae and proboscis (e.g., Hori et al., 2006;Hori et al., 2007;Niggebrügge et al., 2009;Mota et al., 2011).This has probably reduced the amount of explorations of the neural substrates underlying visual cognition, which require coupling representative and reliable behavioral outcomes with invasive methods.While VR constitutes a promising alternative to overcome this limitation, bumble bees offer the additional advantage of being physically robust and thus more resistant for long-lasting brain invasive recordings than honey bees (Riveros & Gronenberg, 2009).Accordingly, electrophysiological long-lasting recordings in various visual processing stages have been performed in the bumble bee brain, yet in immobilized individuals (Paulk & Gronenberg, 2008;Paulk et al., 2008;Paulk et al., 2009a;Paulk et al., 2009b).In addition, the development of connectomic and projectomic approaches for visual centers in the bumble bee brain such as the central complex (Sayre et al., 2021) provides novel maps of visual-neuron connectivity and of neural projection patterns, which adds to the appeal of bumble bees as a model for visual research.Taking advantage of these features under VR conditions could open new perspectives for neural access in behaving animals.

Size polymorphism and visual performances
Contrary to honey bees, bumble bees exhibit important differences in body size and weight that relate to division of labor.While larger workers perform mainly foraging tasks (Spaethe & Weidenmüller, 2002), smaller workers mostly participate in in-hive tasks such as nursing (Jandt & Dornhaus, 2009).These differences have been related to learning abilities (Worden et al., 2005;Riveros & Gronenberg, 2009;Riveros & Gronenberg, 2012;Klein et al., 2017) as foraging brings more opportunities for associative learning to take place (e.g., associating floral sensory traits with food reward as well as environmental cues with navigation goals) and because body size was found to be positively correlated with brain size, which in turn has been considered as a proxy of learning abilities (Collado et al., 2021).In our study, bees exhibited a 2.5 fold difference in body size and a 3 fold difference in dry weight, values which are in accordance with natural size and weight variation found in bumble bee colonies (Jandt & Dornhaus, 2009).These differences did not have an impact on their visual learning performances, consistent with results reported for color discrimination by freeflying bumble bees foraging on artificial flowers (Raine & Chittka, 2008) and in contrast with the clear effects that these parameters have for visually guided maneuvering (Ravi et al., 2020) and flight speed (Crall et al., 2015).
The absence of a size-weight effect on learning goes against the prediction that larger bees would be better learners than smaller ones.Yet, individual experience may be a more decisive factor as larger bumble bees, which are more likely to become foragers, have the opportunity to gather more learning experiences in the course of their activities.Importantly, the bumble bees used in our experiments were all naïve as they did not have the opportunity to leave their colony before being tested so that the factor experience was homogenized among them.The fact that in these conditions no learning differences were observed between bees of different size and weight suggests that the primary role for visual cognitive differences does not refer necessarily to brain size but to individual experience.Alternatively, the tethering condition imposed by the VR setup-which restricted maneuverability, and thus the possibility of learning via active vision-may have overshadowed size-dependent differences in learning.Finally, the elemental learning discrimination task offered to the bees might have been too simple for detecting fine differences in cognitive performances related to size.Using higher-order discriminations (Buatois et al., 2020) may eventually allow discarding or not the role of body and brain sizes as determinant factors of cognitive success in bumble bees.

Conclusion
The present study shows that bumble bees are suitable experimental subjects for experiments on visual learning under VR conditions.This conclusion opens important perspectives for uncovering the neural and molecular bases of such learning via invasive studies in behaving animals, given the robustness of bumble bees and the accessibility of their brains.Honey Bee and bumble bee brains have a similar organization so that the knowledge acquired in both species may lead to a better understanding of insect brains and their cognitive feats.

Fig. 1
Fig.1(A) Feeding experiment (Experiment 1).bumble bees individually enclosed in perforated syringes for experiments on feeding responses to sucrose solution, quinine solution and water (Experiment 1).(B) Feeding method (Experiment 1).The solutions were delivered by means of an Eppendorf tip inserted into the syringe hub, allowing the bee access and to feed on them.Different groups of bees were offered 1.5 mol/L sucrose solution, 3 mmol/L quinine solution and distilled water to evaluate appetitive and aversive responses and responses based on thirst.(C) Global view of the virtual-reality (VR) system (Experiments 2 and 3).1: Semicircular projection screen made of tracing paper.2: Holding frame to place the tethered bee on the treadmill.3: The treadmill consists of a Styrofoam ball positioned within a cylindrical support (not visible) floating on an air cushion.4: Two infrared mouse optic sensors allow recording the displacement of the ball and reconstructing the bee's trajectory.5: Air arrival.The video projector displaying images on the screen from behind can be seen on top of the image.(D) The tethering system (Experiments 2 and 3).1: Plastic cylinder held by the holding frame; the cylinder contained a glass cannula into which a steel needle was inserted.2: The needle was attached to the thorax of the bee.3: Its curved end was fixed to the thorax by means of melted bee wax.(E) Absolute irradiance (W/cm 2 /nm) of the lights used to assess spontaneous color preference of bumble bees in the VR setup (Experiment 2).In order to find a pair of colors that elicited the same amount of attraction, we tested 2 pairs of green and blue virtual cylinders.In 1 pair, 1 cylinder was Green 1 (R: 0, G: 100, B: 1.0; dominant wavelength: 528 nm; irradiance: 300 μW/cm 2 ) and the other Blue 1 (R: 0, G: 80, B: 254; dominant wavelength: 449 nm; irradiance: 2500 μW/cm 2 ) while in the other pair, 1 cylinder was Green 2 (R: 0, G: 51, B: 1.0; dominant wavelength: 527 nm; irradiance: 130 μW/cm 2 ) and the other Blue 2 (R: 0, G: 0, B: 255; dominant wavelength: 446 nm; irradiance: 2700 μW/cm 2 ).The right, green axis refers to Green 1 and Green 2. The blue, left axis refers to Blue 1 and Blue 2. (F) Location of the color used in Experiment 2 in the color hexagon model using Bombus terrestris photoreceptor sensitivity.Photoreceptor excitations (S: short-wave; M: mid-wave; L: long-wave) are plotted at an angle of 120°.Within each color pair, colors were well discriminable as shown by their distant loci in the color space.

Fig. 2
Fig. 2 Feeding responses of bumble bees to distilled water, sucrose solution and quinine solution (Experiment 1).(A) Food consumption (μL/bee) after 1 h enclosure.Bumble bees were enclosed within a syringe in which solutions (distilled water: n = 20; 3.0 mmol/L quinine solution, n = 21; 1.5 mol/L sucrose solution: n = 21) were delivered by means of an Eppendorf tip inserted in the syringe hub.Bees consumed significantly more sucrose solution, while their consumption of distilled water and quinine solution was low.Box plots show the median (horizontal line) and the 10th, 25th, 75th, and 90th percentiles; 5th/95th percentiles are shown as outliers.Vertical scatter plots show the actual data distribution.Different letters on top of box plots indicate significant differences (Tukey's test; P < 0.05).(B) Food consumption (μL/bee) as a function of individual weight (mg).Bumble bees were individually weighed when placed in the syringe and their solution consumption (see above) was represented as a function of weight.Consumption of water and of quinine solution did not vary with individual weight.In contrast, consumption of sucrose solution depended on the weight of bumble bees, with larger bees consuming more than smaller bees.Lines represent linear fit based on a Pearson's correlation analysis.Red lines correspond to a significant correlation (Pearson's correlation; P < 0.05); blue lines correspond to nonsignificant correlations (P > 0. 05).

Fig. 4
Fig. 4 Bumble bees' behavior during the learning trials.(A) Evolution of first choices across trials.The graph shows the percentage of bees responding to the CS+ (red), to the CS− (black) or not making any choice (NC; gray) during the training phase.(n = 235).(B) Cumulative heat maps of bumble bees' trajectories for the 6 learning trials (n = 235).Warmer colors correspond to higher density of visits (see color scale).(C) Distance walked for each trial.Total distance walked on average by bumble bees during each trial (n = 235).(D) Mean speed for each trial.Evolution of the average walking speed across trials (n = 235).(E) Latency to make a choice.Average time elapsed between the start of the trial and the first choice (n = 235).For A, C, D and E the dashed lines above and below the curves represent the 95% confidence intervals.

Fig. 5
Fig. 5 Bumble bees' behavior during the test phase.(A) Choices during the test.The graph shows the percentage of bees responding to the CS+ (red), to the CS− (black) or not making any choice (NC; gray) during the test phase (n = 235).(B) Fixation time.Shows the time spent fixating each stimulus during the test.Error bars indicate 95% confidence intervals.*P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.(C) Cumulative heat map of bumble bees' trajectories during the memory test (n = 235).The color scale indicates the cumulative frequency of bee locations.

Fig. 6
Fig. 6 Impact of bumble bee's body size and weight on the motor and temporal performances during the test.After the last trial, we measured the interwing distance and dry weight for each bee, reported here as Size and Weight.This figure shows the various parameters measured during the test: distance walked, speed and choice latency as a function of insect's size (Left) or weight (Right).Each point is a bee and the line shows the linear regression.Data for 235 bumble bees are represented here.(A, B) Distance walked.(C, D) Walking speed.(E, F) Latency to make a choice.Red lines correspond to a significant correlation (P < 0.05); blue lines correspond to nonsignificant correlations (P > 0. 05).

Fig. 7
Fig. 7 Impact of bumble bee's body size and weight on the learning performances.Learning score is shown as a function of the insect's size (A) or weight (B).For each bee, we computed the learning score as the difference between the number of CS+ and CS− choices from the second trial to the test.Each point represents a bee (n = 235).Blue lines correspond to nonsignificant correlations (P > 0.05).