Choosing the Right Reality: A Comparative Analysis of Tangibility in Immersive Trauma Simulations

Technical skills are important to be trained for first responders, as muscle memory is an important aspect when performing technical skills [25]. Interestingly, according to Baetzner et al. [3] in disaster response training practical skill training was never the sole objective, but combined with other aspects like decision-making or communication.

Haskins et al. [18] provided a concept of use for VR training for first responders (police, firefighters, medical first responders) and noted that ‘VR can teach how to use specific equipment in a stressful but safe simulation’ [18, p. 60]. Physical mockups and passive haptics were especially mentioned as possible solutions. However, this - still - is an area where real-world training often has better results. Recently - due to technical improvements in the area of MR - more approaches towards training technical skills in MR have been published.

Escobar-Castillejos et al. [13] provides a review on haptic simulations in medical training. Examples of MR training in the medical domain are often in surgery - e.g. [17, 19, 22, 38, 40]. In the area of medical first responder training dedicated skills have been addressed. This includes first aid for patients with seizures [2], handling of equipment and tools[25] first aid and reanimation [6], COVID19 Nasopharyngeal swab [7, 62], and FR responses to mass casualty incidents [58]. An approach using a physical manikin in VR in combination with data gloves and a head-mounted display was described by [46]. Similarly, Uhl et al. [54] describe a manikin integrated inside VR using chroma-keying and Scherfgen et al. [45] proposed a system that ‘estimates the pose of a medical manikin’ and ‘haptically augments a 3D human model’ for VR based training, allowing users to physically touch a virtual patient. Apart from using immersive approaches [46, 54] Gasques et al. [14] proposed ‘HoloSim’ where a physical manikin is augmented with visual overlays.

Based on related work, we suggest that training technical skills, especially the combination of training technical skills with other skills [3], is an important aspect that can be achieved well through using MR training. Thus, for the Green Manikin solution presented in this work, we included the usage of real tools (e.g. stethoscope, tourniquet), in combination with the other training modalities.

Due to their exposure to dramatic, critical incidents and chronic stressors, first responders have a higher prevalence of negative health outcomes - due to stress or trauma - compared to other professions. Psychological resilience and training for psychological resilience are important factors in preventing negative occupation-related health effects [21]. Therefore training for psychological skills and especially resilience against stress and trauma is an important aspect in current research towards training for first responders.

A review by Wild et al. [57] showed only limited ‘evidence for interventions aimed to improve well-being and resilience to stress’. However, they also noted that ‘current trials [...] show promise in preventing stress-related psychopathology’, but further evaluations in high-quality trials are needed. Multiple VR training solutions to improve resilience against stress and trauma have been developed, e.g. in the military domain [5, 39], for police forces [37].

Based on the aforementioned work we can draw the conclusion, that immersive technologies are well suited for psychological training, due to the highly immersive environment. Especially stress inducting environments [5, 37] can support the resilience of responders for upcoming stressful situations. Therefore we integrated a realistic stressful environment in our Green Manikin solution, to support training for stressful real-world situations [21].

Haskins et al. [18] describe training challenges for VR training identified through user research with first responders. According to Haskins et al., ‘the most common response we got when we asked what the biggest pain point is for first responders is communication.’

Communication is a complex process that is influenced by many factors and variables [44, 50], misunderstandings [51] or the knowledge base [44, p. 31]. When looking into communication during first responder operations, we can distinguish two categories: (1) inward-facing communication (communication within participating first responders) and (2) outward-facing communication (e.g. people in civil society) [23].

There are examples of training for communication inside a hospital - e.g. diagnosing patients [29] or training for the patient consultation process before surgery [28]. However, they do not address first responders in their complex, stressful and sometimes dangerous context. In the FR domain, [34] evaluated an interactive course for the whole disaster management chain, including communication and coordination. Uhl et al. [54] integrated communication with patients in their MR simulation, where the trainer role-plays the patients using a microphone. Also (commercial) approaches for training communication in stressful scenarios exist - e.g. Enhanced Dynamic Geo-Social Environment EDGE ¹, ETC Simulation - Advanced Disaster Management Simulator ² or XVR Simulation ³.

Communication training is often done via real-world roleplay. However, Uhl et al. [52] could show that VR can lead to similar results in terms of social presence as real-world roleplay, thus communication training can be done in MR if the dialogue and the responses of NPCs are properly designed. In the Green Manikin solution, we integrated the possibility of talking to the victim (see Section 3.2.4) to address outward-facing communication. Inward-facing communication was not a focus of the presented study.

Decision-making skills, including the knowledge of processes, rules, and situational demands, are crucial in the education and training of medical first responders, as well as for nurses, physicians, and medical students. This is particularly true in mass casualty incidents and disaster situations, where decision-making and triage are essential skills.

Various forms for training triage exist including video [1], real-world high fidelity simulation [8, 9], educational review sessions [10], exercises with patient dummies [11], card-based training [24], serious games [24], educational courses [60] and Mixed Reality (MR) [54].

Special situations are also a focus of training. For example, Montan et al. [34] describe an interactive course that covers the entire ‘chain of response’ from the scene to the hospital, including communication, coordination, and command, using magnetized cards in a tabletop simulation. Motola et al. [35] evaluated training of medical first responders to CBRN responses using a training video that corresponds to the trained CBRN scenario. Jones et al. [20] assessed training for Emergency Medical Services’ response tactics to active shooter incidents, and Edinger et al. [12] evaluated online education for dealing with individuals with developmental disabilities.

Simulations, increasingly including computer/VR/MR-based methods, are common for decision-making training. We suggest that MR is a fruitful medium for training in medical decision-making, as the immersive environment enhances the perception of stressful situations and supports realistic decision-making processes. The Green Manikin solution we present supports decision-making training on multiple levels, from situation assessment to planning and executing the appropriate response, including the selection and application of the right tools for patient treatment.

The MR training prototype was developed using Unity 3D version 2022.3.7f1 set to High-Definition Rendering Pipeline (HDRP) for enhanced graphical fidelity. The Varjo XR-3 head-mounted display (HMD) was employed in conjunction with the application Varjo Base for tracking, depth-sensing and calibration, and Varjo Lab Tools to enable chroma-key masking surfaces to combine the real and virtual worlds. The system ran on a Dell Alienware Aurora R12 PC with an 11th generation Intel i9 CPU, 64GB of RAM and a Geforce RTX 3080 graphics card which resulted in approximately 40 frames per second rendering capabilities during play time.

For chroma-keying, a specialized green screen setup is used, comprising green fabric and floor tape. This enables the integration of physical objects, such as a stethoscope, into the virtual 3D environment. Additionally, it allows for a layered composition in which the physical Green Manikin is augmented with a virtual avatar. The depth-testing feature of the Varjo XR-3 supports the seamless integration of the user’s real hands and medical instruments into the virtual environment, utilizing depth information captured by the device’s integrated LiDAR sensors.

Positioning and orientation of tools like the stethoscope are tracked through the use of printable fiducial markers attached to the objects. These markers are identified by the pass-through cameras integrated into the Varjo HMD. To provide an immersive auditory experience, the trainee is equipped with an audio headset that delivers spatialized sounds, i.e., ambient sounds, breathing noises (internal and external) and voice from the virtual agent had distance attenuation and sound source localization properties.

The contextual 3d avatars (NPCs) were produced using Realillusion’s Character Creator 4⁵ software and custom-animated with Realillusion’s iClone 8⁶. The main virtual agent simulating the victim had no pre-programmed animations, but reacted to sound cues during run-time using SALSA LipSync Suite animation tool. For example, it compressed and expanded its chest according to a set intensity value from integrated breathing sound patterns, it also moved its lips and jaw according to the voice generated by the Text-to-Speech engine responsible for generating the voice.

Two Wallimex Pro Daylight 1260 photography studio-lights were used when lighting conditions affected the shades of green, and thus more control over the environmental brightness and shadows was required.

We used a simple full-body training manikin, made of hard plastic and moveable limbs. In the first step, the manikin was painted green, using plastic primer and green acrylic spray paint for the individual parts. After assembly, we installed the Vive trackers on the now Green Manikin using Vive’s body tracking straps, attached to hands, feet and hip. An additional tracker was mounted on top of the manikin’s head. To ensure tracking even when the users would cover the trackers from multiple sides, four lighthouses were placed around the manikin and green screen floor.

In Unity, the inverse kinematics package VR IK was used to map the Vive trackers on the manikin’s body (see Figure 2) to the respective parts of the virtual 3D model in Unity. The mapping process also required the scaling of the existing 3D model to match the proportions of the individual parts of the manikin. Ideally, the manikin should be scanned to produce a 3D model with exact proportions that can be further modified. Doing this should result in a more accurate and faster mapping process. Once finished, the manikin and the 3D model move in sync and further natural interactions are possible, e.g., lifting the hand to have a closer look at the virtual pulse oximeter on the finger, or moving the leg for better placement of the tourniquet. However, the trackers need to always be visible by the surrounding base stations. If their path gets blocked, the mapping between the manikin and the 3D avatar is broken which leads to awkward positions of the avatar. This is a limitation, given that some treatments performed by MFRs require that they adopt positions that block the trackers, e.g., turning the patient on the back for visual inspection of injuries, or placing their knees around the head of the patient to use a bag valve mask. Additionally, a calibration offset which can occur for several reasons, e.g., relocating the manikin, or inadvertently moving trackers or base stations, will cause a mismatch between the manikin and the 3D avatar. In these cases a re-calibration of the play area or the trackers will be needed.

A total of four (physical) medical tools were integrated into the training scenario, which would be stored in a first responder backpack. To enable our research on tangible tool interaction, we used a mix of wizard-of-oz simulation of the interaction and automated interactions.

A tourniquet could be applied to the manikins leg to stop a (virtual) bleeding (see 3.4) on the virtual patient, which would be stopped by the operator manually. On the stethoscope a fiducial marker was applied, allowing it to be tracked in space. A simple collision detection script, enables a tangible way of blending in the internal lung sounds, once the stethoscope is placed on the chest of the manikin. The bag valve mask could be placed on the patient’s mouth and nose and squeezed to ventilate the patient. This would be synced with the virtual avatar’s chest moving up and down, which was achieved manually by the operator moving a slider. Lastly, to communicate vital signs like heart rate and blood oxygen levels, a virtual oximeter was added to the scene, that would dynamically display the vitals of the patient, depending on their state.

In order for the users to talk to the virtual patient, we employed a mixture of speech recognition, a specifically trained Large-Language-Model (LLM) on the basis of ChatGPT 3.5⁷, and voice output. This system will be described in detail in future work, so we will only give a short summary here.

The system captures a user’s voice through a noise-cancelling microphone and sends the audio to OpenAI’s Whisper⁸ for initial processing. Whisper, an Automatic Speech Recognition module, converts the speech into text. This transcribed text is then merged with role characteristics set in a prompt for ChatGPT. The prompt serves as a constraining feature to guide the model’s responses, which can be tailored to evoke a specific emotional tone or context. ChatGPT then generates a text-based response that considers both the transcribed message and the prompt. Finally, ElevenLabs’ Text to Speech engine⁹ transforms the text response back into audio, offering various customization options like pitch and emphasis. Finally, the resulting audio file is played in the virtual scenario, automatically moving the patient’s lips. The voice used by ElevenLabs was custom recorded with a hurting, shocked timbre, to simulate a hurting patient.

The Virtual Reality prototype was developed using Unity 3D version 2022.3.7f1 set to Universal Rendering Pipeline (URP) to optimize graphical performance and resolution. The Meta Quest Pro HMD was employed as a standalone device with its integrated hand-tracking, depth sensing, play area calibration, and speakers. The Quest Pro´s internal Qualcomm Snapdragon XR2+ with 12 GB of RAM resulted in approximately 30 frames per second rendering capabilities during play time. No controllers were used in the VR modality since participants only needed to move freely within a 2m x 2m play area, i.e., no teleportation was necessary. All visual elements, e.g., NPCs and the main virtual agent simulating the victim with its corresponding sound, reactive chest movements and facial gestures, remained the same as in the MR modality, and only the materials, textures, skybox and lightning systems were adjusted to URP.

The same four medical tools as the other modalities, i.e., tourniquet, stethoscope, bag valve mask and oximeter, were available in VR as virtual models. Participants were able to manipulate these 3D objects with a natural hand grab gesture and could release them anywhere by spreading the fingers apart. No snapping to the virtual patient was applied to these objects in order to study their ease of use with free rotation and translation affordances. These interactions were designed to better map to the other modalities that use real-world physical objects.

The same mode of communication as in the MR modality was employed for the VR version (see Section 3.2.4).

In the real training simulation, a role-player was recruited to lie down in an office room. A bicycle was placed beside him, and printed images of a wound and a blood puddle were positioned on his thigh and adjacent to him on the floor, respectively (see Figure 4 for an overview of the steps in the real training scenario). The role-player was thoroughly briefed regarding the scenario, and was given the same information that made up the prompt of the speech interaction system in the virtual modalities. The scenario was performed as in the other modalities, with information about vitals (like the heart rate, respiratory rate or whether the bleeding had stopped) being given by the trainer when asked, as these can’t be ‘acted’. The participants would get the contextual information of the scenario (weather, time of day, information of the environment and how they arrived on the scene) from the trainer before entering the office room. Upon entering the room, they would approach the role-player (Figure 4 b) and talk to the role-player, who would make loud pain noises and answer within the constraints of the information given to him during the briefing. Participants would the apply pressure on the simulated wound (often first with their knees, see Figure 4 c) and then apply the tourniquet. They would check the airway and use the stethoscope, to be given the respiration rate by the trainer (Figure 4 d). The decay of the patient was actively reported by the trainer (e.g. ‘His skin is getting pale and sweaty and his oxygen level is dropping.’), which would also lead to the role-player not answering anymore, and falling unconscious. Some participants first used an oxygen mask (Figure 4 e, but ultimately switched to the ventilation bag, when the vitals (especially the oxygen level) would not improve.

After equipping the Varjo XR-3, participants would first see the passthrough video of the lab environment. The MR application was launched directly from the PC, which would blend the passthrough to the MR view of the accident environment (Figure 3 a). Because of the chroma-keying of the green floor and manikin, participants could see their own hands and body overlaid over the virtual scene.

They would further see the physical backpack with their tools next to the tangible patient. As the manikin is green, it would be replaced with the virtual patient, whereas the physical zip hoodie would be visible around the virtual, but tangible patient (see Figure 5 bottom row). To stop the bleeding they would take out the tourniquet and apply it on the manikin, with the tourniquet being overlaid on top of the virtual thigh (see 5 bottom left). They would talk to the patient via the system described in Section 3.2.4. When placing their hands on the chest where the bump mark was, the patient would make a sound of pain. Participants would uncover the chest by opening the (physical) zip hoodie and proceed to use the stethoscope. When the stethoscope was placed on the chest, the internal breathing sounds would become audible. Lastly, upon the decay of the patient, participants would apply the bag valve mask on the tangible manikin (see Figure 5 bottom right), and upon pressing the bag valve bag, the chest of the cyclist would be moved up and down accordingly by the operator (i.e. wizard-of-oz).

Upon equipping the Meta Quest Pro, participants would first see the passthrough video of the lab environment. The VR application would be launched directly from Unity, which would switch the headset to VR. Just as in the MR modality, participants would appear at the scene of the accident on a square (see Figure 3 a). Participants would first look around the virtual scene and then attend the patient. The medical tools were placed near to the patient as 3D assets, matching the look of the real tools (see Figure 3 b). Using the hand-tracking feature of the Meta Quest Pro, participants could grab the virtual tools and apply them to the patient by moving them to the appropriate location. The patient wore a t-shirt, which could be removed by the participants by means of a pinching gesture, that was explained to them during the briefing. When touching the bump mark on the patient’s chest with their virtual hands, the patient would grunt with pain. The tourniquet, when moved to the correct position on the leg would trigger the blood flow to stop after 2 seconds. For the stethoscope, the internal lung sounds would be audible if the virtual stethoscope collided with various areas on the chest of the patient and be inaudible once the stethoscope was removed. The ventilation bag could be moved to the correct position on the patient’s mouth, with the participants then performing a pumping gesture on the ventilation bag, which would animate the bag accordingly, as well as trigger chest-movements according to the movement. The patient’s vitals could be checked on a oximeter on the patient’s left hand. Once the initial treatment (i.e., once the tourniquet and stethoscope were used), the experimenter would trigger the patient’s decay, visualized through changing vitals on the oximeter, the patient turning blueish, sweaty and finally unconscious. This would trigger the participants to use the ventilation bag, which then marked the end of the scenario.

To evaluate the acceptance of first responders for the three training modalities (RQ1), we utilized the framework provided by the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2), as described by Venkatesh et al. [55]. We adapted items from the original UTAUT2 model based on their German translation by Harboth & Pape [16] to better align with the training contexts of the three modalities. Four sub-dimensions of the UTAUT2 model—Habit, Social Influence, Price Value, and Use Behavior—were deemed irrelevant to our research focus and were therefore excluded. The scales used in our study thus included Performance Expectancy (PE), Effort Expectancy (EE), Hedonic Motivation (HM), Facilitating Conditions (FC), and Behavioral Intention (BI). We recorded responses to the Technology Acceptance items on a 5-point Likert scale.

To measure the sense of presence experienced by participants in the three training modalities (RQ1), we utilized the German-translated version [56] of the Multimodal Presence Scale (MPS) originally developed for VR settings by Makransky et al. [30]. This scale is rooted in Lee’s unified theory of presence [27], and categorizes presence into three distinct dimensions: physical, social, and self-presence. Physical presence relates to how participants experience objects and environments in the virtual setting. Social presence covers the sense of interaction with other virtual characters or users. Self-presence measures how much participants feel they are part of the virtual environment. Responses to MPS items were collected using a 5-point Likert scale.

To gain qualitative feedback for our research questions–mainly RQ2–three open-ended questions were answered by the participants at the end of the questionnaire:

To assess the impact of the three different conditions (MR, VR, REAL) on the multiple dependent variables regarding presence and technology acceptance, we employed Friedman tests, which is a non-parametric statistical method used for analyzing ordinal data across multiple groups. This test is particularly useful when the data violate the assumptions of normality and homogeneity of variance, allowing for a robust evaluation of differences among the training modalities.

For the variable Physical Presence, the Friedman test showed no statistically significant differences, with χ²(2) = 5.92 and p = 0.052. However, for Self Presence, a statistically significant difference was observed, χ²(2) = 8.67, p = 0.013. The test for Social Presence also revealed a highly statistically significant difference, with χ²(2) = 26.07 and p < 0.0001. For Behavioral Intention, the test was not significant, showing χ²(2) = 5.15 and p = 0.076. A highly significant difference was detected for Effort Expectancy, with χ²(2) = 18.73 and p < 0.0001. The test for Facilitating Conditions yielded a significant result, χ²(2) = 7.68 and p = 0.021. Lastly, no significant differences were found for Hedonic Motivation and Performance Expectancy, with χ²(2) = 3.21, p = 0.201 and χ²(2) = 4.29, p = 0.117, respectively. Means and standard deviations are presented in Table 1, boxplots and distributions of the scales for presence and technology acceptance are displayed in Figures 6 and 7 respectively.

To further explore the observed differences, post hoc Wilcoxon signed-rank tests were performed, and resulting p-values were corrected using the False-Discovery-Rate (FDR) [4], which is a less conservative alternative to the Bonferroni correction for multiple comparisons. The results for the significant variables are detailed in Table 2.

The post hoc tests revealed significant differences in Self Presence between MR and VR as well as VR and REAL, with medium effect sizes (|r| > 0.24). Similarly, Social Presence demonstrated significant differences across all pairwise comparisons with medium effect size for MR and VR (r = 0.24) and high effect sizes for MR and REAL (r = −0.416) and VR and REAL (r = −0.493). For Effort Expectancy, significant differences were found between MR and REAL (r = −0.280) with a medium effect size, and VR and REAL (r = −0.449), with a large effect size. Facilitating Conditions also showed significant differences between MR and VR, and VR and REAL, with moderate effect sizes.

With the first open-ended question, the goal was to gather overall feedback for each of the three modalities.

Real Simulation was often referred to as the established standard in training, familiar to most participants (4 mentions). It was lauded for its naturalistic human interaction and tangible experience but noted for its limitations in simulating complex or high-risk scenarios (4 mentions).

VR generated mixed responses, leaning towards the view that the hand tracking implementation has not yet reached its full potential (8 mentions). Participants pointed out issues like a lack of haptic feedback, problems in communication, and limited equipment as areas that need development (4 mentions). Despite these limitations, several participants recognized its potential for organizational and leadership training (3 mentions).

MR received the most favorable reviews, often described as the best of both worlds, combining elements from real-life scenarios and virtual elements (8 mentions). Participants noted that MR could simulate scenarios more realistically and provide an effective training environment (3 mentions).

Participants’ responses provide insightful perspectives on the utility and limitations of Real Simulation, VR, and MR in training scenarios. Real Simulation, while reliable, is limited in its ability to handle complex training needs. VR shows promise, particularly for ‘simulation of organizational and leadership tasks’. MR stands out as the most balanced and promising modality, particularly for ‘training of patient treatment’, gaining the highest number of positive mentions. The choice of modality should therefore align with specific training objectives to maximize effectiveness.

The second open-ended question sought to understand the participants’ preferred training modality when preparing for real-life medical scenarios. Participants were asked: ‘If you had to choose a training method to prepare for real scenarios, which one would it be and why?’. The preferences were categorically divided into MR, VR, and Real Training for the purpose of analysis.

Mixed Reality. A majority of the participants, 11 out of 16, indicated a preference for MR as their training modality of choice.

Realistic Simulation: The ability of MR to seamlessly integrate real and virtual elements was cited as a significant advantage. Participants highlighted that MR provides realistic patient and environmental details (7 mentions).

Haptic Feedback: MR’s feature of allowing actual ‘hand movements’ and tactile feedback was noted as crucial for effective training (3 mentions).

Resource-Intensive but Worthwhile: Although one participant mentioned that MR is resource-intensive, they noted that the high-quality training experience justified the resource expenditure (1 mention).

Speech Interaction: One participant suggested that MR could be even more effective if improvements in voice interaction are made (1 mention).

Real Training. Four participants expressed that Real Training would be their preferred modality.

Immediate Feedback: Participants appreciated the instant communication and feedback possible with real patients (2 mentions).

Physical Interactions: The ability to perform physical tasks like bandaging and carrying patients was highlighted as an advantage (2 mentions).

Scenario-Specific: One participant additionally noted that Real Training is optimal for certain types of scenarios like confined spaces and neurological conditions (1 mention).

Virtual Reality. Notably, none of the participants selected VR as their preferred training modality. However, some did mention its utility for specific types of training scenarios, e.g. mass casualty events, as in this case the focus is on triage and not treatment, making the tangible aspect less important while still being able to simulate large environments with multiple injured persons.

Two participants indicated that a combination of MR, VR, and Real Training in a training curriculum could be beneficial, suggesting a multimodal approach depending on the specific training needs and scenarios.

Participants often mentioned MR and VR in the same context when it comes to simulating a virtual environment effectively (4 mentions). They indicated that both these modalities are superior for creating an immersive scenario, particularly when it comes to generating a ‘general impression of an emergency scene’ and mimicking real-world distractions. Moreover, these modalities were said to offer flexibility and realism in training scenarios that are otherwise difficult or costly to stage in real life, such as mass casualty incidents or dangerous situations (1 mention).

A synergy between MR and Real Training was noted in several responses (3 mentions). MR was praised for its realistic simulation of the environment and specific medical tasks, while Real Training was cited as essential for hands-on practice with actual medical equipment and patient care. The sentiment was that, while MR offers a ‘training advantage’ by portraying the environment more realistically, the care of the patient still feels more natural in a real-world setting (1 mention).

MR was specifically highlighted for its realistic depiction of the environment (3 mentions) and for skill-specific training such as stemming a bleeding wound. It was perceived as particularly useful for training where understanding the context or environment is crucial.

VR was appreciated for its organizational training potential (2 mentions). It was also favored for the assessment of emergency patients, particularly in continuing education where less equipment is needed (1 mention). It allows for the simulation of factors like skin color and respiratory rate more effectively compared to other methods.

Real Training was explicitly mentioned for its benefit in the initial stages of training (1 mention). It was also touted for its unique ability to facilitate genuine physical interactions with patients, an aspect that is crucial in medical practice but challenging to simulate effectively in MR or VR (1 mention).

In summary, the findings reinforce the idea that each training modality has unique strengths that make it particularly well-suited for specific aspects of emergency medical training. As one participant succinctly put it: ‘VR [is best] for organizational and leadership tasks, MR [is best] for immersive patient treatment and real training [is best] for entry-level training’. This statement encapsulates well the general sentiment expressed across the responses and affirms that the best training approach depends on what needs to be achieved, while still highlighting MR as a promising future direction for MFR training.