Abstract
How do humans integrate landmarks to update their spatial position during active navigation? Using immersive virtual reality and high-density mobile EEG, we investigated the neural underpinnings of landmark-based recalibration during path integration in freely moving male and female humans. Participants navigated predefined routes, indicated the start position to quantify accumulated errors, and once per route corrected their estimate using a visual landmark. Our findings reveal that homing error accumulated along the course of navigation, but a briefly presented intramaze landmark effectively corrected accumulated errors. However, this effect was transient and less pronounced when participants were highly confident in their self-motion-based spatial representation suggesting that internal priors can hinder the assimilation of novel spatial cues. Theta activity in the retrosplenial complex supported the realignment of internal spatial representations by anchoring self-motion-derived estimates to visual landmark cues. Increased theta power and phase resetting upon landmark presentation accompanied subtle corrections, supporting a smooth realignment of spatial representations, whereas diminished synchronization marked the need for more extensive spatial updating. In addition, we identified motor-related theta response that scaled with rotational acceleration. Taken together, these findings highlight the dual role of theta oscillations in flexibly integrating multimodal signals, supporting both the recalibration of spatial representations to external cues and the encoding of self-motion information during naturalistic human navigation.
Significance Statement
Spatial navigation depends on the continuous integration of internally generated self-motion cues with external environmental landmarks to construct and maintain spatial representations. Despite its central role in adaptive behavior, the neural dynamics underlying this intermodal integration remain unclear. Using noninvasive mobile EEG with immersive virtual reality, we identify a dynamic mechanism within the retrosplenial complex that realigns internal and external spatial representations based on visual landmarks during naturalistic navigation. This process is mediated by theta-band activity, which not only supports visuospatial realignment but also encodes vestibular signals. These dual theta signatures demonstrate how the human brain flexibly integrates multimodal information to update spatial representations, highlighting theta oscillations as a unifying neural substrate for both perceptual and motor components of navigation.
Introduction
Spatial navigation relies on the integration of multiple sources of idiothetic (i.e., body-based) and allothetic (i.e., environment-based) information. When allothetic information such as landmarks are absent, individuals rely exclusively on continuous integration of idiothetic information from self-motion including optic flow, vestibular and proprioceptive information to maintain orientation (Wolbers and Hegarty, 2010), a process known as path integration. However, the reliability of self-motion cues declines with distance as sensory noise accumulates, leading to greater navigational errors (Stangl et al., 2020). To recalibrate the path integrator and stop error accumulation, visual landmarks play a pivotal role in humans (Ekstrom, 2015; Wang, 2016). However, despite the pervasive nature of this intermodal recalibration, how the human brain integrates and combines external landmarks with self-motion cues remains elusive.
A substantial body of work identified the brain regions supporting the isolated processing of idiothetic or allothetic information. These studies highlighted hippocampal and extrahippocampal contributions when using self-motion cues (Baumann and Mattingley, 2021; Parra-Barrero et al., 2023). Conversely, when landmarks are the only source of information, high-level visual regions are recruited, including parahippocampal place area (Epstein and Kanwisher, 1998), retrosplenial complex (RSC; Vann et al., 2009), and occipital place area (Julian et al., 2018). Although distinct neural correlates of idiothetic and landmark integration suggest separate mechanisms for self-motion and visual-based navigation, these processes are instead considered complementary (McNaughton et al., 2006). In rodents, self-motion cues provide fundamental spatial information, whereas external visual cues recalibrate the path integration system (Campbell et al., 2018; Jayakumar et al., 2019). Recent animal studies highlighted the RSC’s role in combining landmark and self-motion cues for navigation (Alexander et al., 2023; Sit and Goard, 2023; Dubanet and Higley, 2024). The integration of these cues is supported by the RSC's connectivity with cortical and subcortical regions, including the hippocampus, parahippocampal structures, and the visual system (Campbell et al., 2021). This hub role supports the binding of allothetic and idiothetic cues into flexible spatial representations that accommodate shifts in sensory input and context-dependent navigation demands (Stacho and Manahan-Vaughan, 2022).
In humans, the RSC has been primarily linked to the visuospatial processing of landmarks (Auger et al., 2012) and recently extended to the positional coding of self-motion cues to track translation and rotation (Chrastil et al., 2016; Chen et al., 2024). However, most studies relied on stationary protocols, overlooking key self-motion cues (vestibular input, proprioception, and motor efference copies) essential for naturalistic navigation (Taube et al., 2013; Iggena et al., 2023). Advances in EEG and virtual reality now enable studying brain dynamics during unrestricted movement (Gramann et al., 2011; Stangl et al., 2023). Studies using these methods extended the role of the RSC to active navigation, highlighting its involvement in full-body rotations computation, reflected in slow-frequency oscillations like theta activity (Delaux et al., 2021; Gramann et al., 2021). Despite these initial findings suggesting the importance of RSC theta activity for human spatial navigation, it remains unclear how the human brain dynamically combines information conveyed by self-motion cues and landmarks to update the path integration system in naturalistic human navigation.
To address this issue, we combined immersive virtual reality with high-density mobile EEG to investigate the neural mechanisms of path integration. Participants navigated predefined routes and periodically pointed to the remembered starting position, allowing us to quantify accumulated positional errors. Once per route, immediately after a pointing response, participants were shown a previously encountered landmark and allowed to update their pointing. We hypothesized that landmarks would reduce accumulated errors by realigning estimates of orientation, thereby improving spatial accuracy. We further predicted that correction magnitude would vary with participants' subjective confidence reported before landmark presentation. Path geometry was manipulated to selectively increase travel distance or rotational displacement, isolating distance- versus rotation-specific error accumulation, which likely reflects distinct neural coding of self-motion cues. Finally, we predicted that angular correction following landmark presentation would be indexed by theta-band modulation within the RSC, underscoring its role in realigning internal spatial representations.
Material and Methods
Participants
We recorded the EEG activity of 29 healthy young participants with no history of neurological pathology and normal or corrected-to-normal vision. This number was determined based on a priori power analysis conducted with G*Power 3.1 which indicated that a sample size of n = 23 would provide 80% power to detect a within-subject effect size of d = 0.60 at an alpha level of 0.05. This estimate was based on prior findings demonstrating modulation of occipito-parietal theta activity by rewards measured with mobile EEG (Lin et al., 2022). Because we planned to perform source reconstruction using dipole clustering, we increased our target sample size. Indeed, Delaux et al. (2021) reported that, when using the same methodology, an average of only 13.3 out of 16 participants (∼83%) were retained in the final clustering solution. To ensure a final sample of 23 participants after clustering, we therefore aimed to recruit 28 participants (i.e., 23/0.83 ≈ 28). They all signed an informed consent form approved by the Ethics Committee of the Technical University of Berlin (BPN_GRA_231204, Institute of Psychology & Ergonomics, Technical University of Berlin, Germany). They were recruited locally via an online platform and received monetary compensation (€12/h) for their participation. We excluded one participant from our sample due to excessive artifacts in the EEG recording, resulting in a final sample of 28 participants (11 females; mean age, 25.92 ± 3.75). Participants completed a motion sickness questionnaire to ensure they reported no symptoms (Kennedy et al., 1993) and a computerized version of the Perspective Taking Task/Spatial Orientation Test to assess their visuospatial abilities in a 2D environment (Friedman et al., 2020).
Experimental design
The virtual environment was created using the Unity3D game engine (version 2021.3.8f1 for Windows) from Unity Technologies, San Francisco, California, USA. Rendering was performed using an HTC Vive Pro head-mounted display with a 90 Hz refresh rate, dual 3.5 inch AMOLED screens with 1,440 × 1,600 pixel resolution, 615 ppi, and a 110° nominal field of view. The HTC Vive Pro was connected to a VR-ready backpack computer (Zotac PC) equipped with an Intel 7th Gen Kaby Lake processor, GeForce GTX 1060 graphics, 32 GB DDR4-2400 memory support, and running Windows 10 OS (ZOTAC Technology). This setup was battery powered and remotely controlled. The participant's head movements were recorded using an integrated HTC Lighthouse motion tracking system, which included four cameras operating at a sampling rate of 90 Hz and covering an area of 8 × 12 m.
The virtual environment consisted of an infinite floor with a grass texture and some added grass tufts to allow the integration of optic flow by the participants (Fig. 1A). The task consisted of an extended triangle completion task with four stops, each marked by a yellow ring that appeared along the path. At each stop, participants were asked to turn in the direction indicated by an arrow (either in the shortest direction in the normal and long-leg condition or in the longest in the long-rotation condition) in order to face their (unmarked) starting point as accurately as possible. Participants then pressed a button on the controller to submit their response, from which we calculated the angular error as the deviation between the true angle and their reported angle. They were then presented with a progress bar and had to indicate their subjective confidence (i.e., the longer the bar, the more confident they were). They were carefully familiarized with the subjective confidence indicator beforehand to ensure that they understood it well. Participants had to walk from one ring to another without any visual cues. After reaching each ring, they were presented with an arrow for 1 s, which indicated the direction they had to rotate. At either the second or third stop, we presented them with a landmark (i.e., a stone) that they had seen at the beginning of the trial. Participants had to wait one second to look at the landmark before being asked to report again the position of their starting point and their subjective confidence, using the spatial information conveyed by the landmark to correct their previous pointing (Fig. 1B). After the four stops, participants then had to physically return to their starting point, validate their response, and report their confidence in their homing task. From this validation position, we were able to calculate the distance homing error (Bierbrauer et al., 2020), which is the error in total distance independent of any angular error:
Finally, they were presented with a red ring indicating the true position of the starting point and had to walk to this ring to start a new trial. The task included three different conditions, with trial order counterbalanced so that half of the trials began on the left and half on the right side of the physical space. We used three path configurations: a standard path with legs of three meters each, a long-leg path in which the length of each leg was increased to 5 m, and an increased-rotation condition in which participants had to perform rotations at a larger angle (Fig. 1C). In addition, we included two landmark configurations in which the landmark was presented at either the second or the third stop. This design resulted in a total of 12 unique paths derived from the combination of 2 start positions (left/right) × 2 landmark positions (second stop/third stop) × 3 path types (standard/long leg/increased rotation). Each of these 12 unique paths was presented in a randomized order within each run, resulting in a total of 60 paths per participant, divided into five runs presented in a pseudorandomized order, each consisting of 12 paths to complete. On average, each run lasted 14 min, resulting in a total average recording time of 71.32 min. To maximize the informativeness of the landmark, we ensured its stability (the position remained constant across trials), its usefulness for navigation (the position was close to the starting point), and its visual features (it conveyed radial information and was easily visible at different stops; Auger et al., 2012). We used a proximal landmark, similar to that used by Doeller et al. (2008) to ensure that the landmark information dominated over idiothetic information during path integration, as speculated in previous work (Zhao and Warren, 2015). This also allowed us to disregard a complex part of cue combination, namely, the process that allows us to check the stability of a landmark before relying on the conveyed information (Auger et al., 2012; Page and Jeffery, 2018; Fischer et al., 2024).
EEG acquisition and processing
EEG signals were acquired at a sampling rate of 500 Hz using two cascaded eego mylab amplifiers (ANT Neuro) and digitized at 24 bit resolution (Fig. 1C). Wet electrodes were embedded in a 128-channel actively shielded Waveguard cap according to the extended 5% system (Oostenveld and Praamstra, 2001). The online reference electrode was CPz, and the ground electrode was a wired gel droplet. The impedances of all electrodes were carefully kept below 10 kΩ using conductive gel, with most electrodes below 5 kΩ. One electrode was placed below the participant's left/right eye to improve the identification of eye-movement-related activity in the subsequent processing steps. EEG recordings, stimulus presentation, and motion capture were synchronized using Lab Streaming Layer (LSL) software (Kothe et al., 2024) and recorded via the LabRecorder.
EEG data were preprocessed offline (Fig. 5) using Matlab (R2024a; The MathWorks) and the BeMoBIL Pipeline (Klug et al., 2022) working with scripts for the EEGLAB toolbox version 2024.0 (Delorme and Makeig, 2004). First, nonexperimental time segments were removed, and we manually inspected the raw signal to remove excessive artefacts. The data were then downsampled to 250 Hz, before line noise was removed using Zapline Plus (Klug and Kloosterman, 2022) which enables the automatic detection and removal of artifactual peaks (mainly 50 and 90 Hz corresponding to line noise and the refresh rate of the head-mounted display). Bad electrodes were rejected using clean_raw_data with a correlation criterion of 0.8 and a maximum time of broken electrodes of 0.5. These electrodes were then interpolated using spherical spline interpolation, and the data were rereferenced to a common average reference. Subsequently, we cleaned the data in the time domain using ASR with a burst criterion of 30 (Chang et al., 2020), which resulted in the rejection of an average of 3.01 ± 3.76% of the data points. The data were then temporarily high-pass filtered at 1.75 Hz before AMICA decomposition (Klug and Gramann, 2021) was applied to decompose the mobile EEG data into statistically maximally independent components (ICs). Data points that did not fit the ICA model were rejected in 10 iterations using the AMICA algorithm with a sigma threshold of 3 (Klug et al., 2024). Finally, we fitted equivalent dipole models to each resultant IC using DipFit before labeling the identified components using the lite version of ICLabel (Pion-Tonachini et al., 2019). We rejected components labeled as muscle, line noise, eye, and heart, and kept only those classified as at least 30% brain and with a residual variance (RV) of the respective ICs below 15% (Delorme et al., 2012). This resulted in a conservative average retention of 16.69 ± 6.89 components, in line with what has been reported in other studies using similar mobile EEG (Delaux et al., 2021; Gramann et al., 2021). The data were then filtered between 0.3 and 50 Hz for further analysis and epoched from 3 s before the landmark appeared to 4 s after. Finally, epochs containing artifacts >100 μV were removed.
We then extracted the retained ICs and used repeated measures clustering with a region of interest (ROI) constraint. To do this, we first performed a preclustering principal component analysis using event-related spectral perturbations (ERSP, weight = 3), scalp topographies (weight = 1), the power spectrum (weight = 1), and the location of the dipole (weight = 6). These features were used exclusively for preclustering and were not included in subsequent analyses. We then performed 5,000 iterations of k-mean clustering with 17 clusters and an outlier threshold of 3 SDs. We defined a ROI centered around the RSC using [0, −55, 15] as previous work suggested these coordinates to be the most representative (Delaux et al., 2021). Each clustering solution was then scored using a weighted combination of 6 metrics: the number of subjects per cluster (weight = 6), the number of IC per subject in the same cluster (weight = −3), the normalization of the spread (weight = −1), the mean RV (weight = −1), the distance from the ROI (weight = −3), and the Mahalanobis distance from the median of the multivariate distribution (weight = 1). We then combined and scaled these multivariate matrices on a scale from 0 to 1 and finally selected the highest ranked solution. If multiple ICs per subject were present in the cluster, we selected the best IC for each subject (based on dipole position and RV). Finally, we performed a time–frequency decomposition of the activity of the selected ICs in the cluster. To ensure the detection of transient activity, we used the Superlet approach (Moca et al., 2021) implemented in the FieldTrip toolbox (Oostenveld et al., 2011) and performed a single-trial z-score normalization as suggested by Grandchamp and Delorme (2011). For each trial, we thus defined the baseline window from –500 to –200 ms before landmark appearance (a period in which participants were stationary following the confidence validation) and computed the mean and standard deviation of the spectral power at each frequency. All time–frequency estimates were then converted to z-scores relative to their own baseline, ensuring that across-trial averages were not biased by a few high-amplitude epochs.
Statistical analysis
Statistical analyses of the behavioral data were performed using the R statistical software (version 4.4.2, R Foundation for Statistical Computing) with R studio (version 2024.09.1) and linear mixed-effects models from the lme4 package (Bates et al., 2014). To compare the different lmer models (e.g., with and without interaction between fixed effects or with more or less complex random effects structure), we computed the Akaike information criterion (AIC; Akaike, 1974) and selected the model with the lowest AIC value, also ensuring that the final model did not exhibit singularity. To extract p values, we used the anova function and the Satterthwaite method with type III sum of squares. Estimated marginal means (EMMs, referred to as M in the manuscript) were calculated using the emmeans package in R and are the mean values reported hereafter. Finally, post hoc Tukey's honestly significant difference (HSD) tests were performed, and partial eta-squared and Cohen’s d were reported as post hoc effect size measures. To check for normality of residuals and homoscedasticity, both were carefully examined using quantile–quantile plots and box plots, respectively. For metacognitive abilities, we computed a composite score of normalized subjective confidence and normalized error:
For intertrial phase coherence (ITPC), we separated epochs for each subject based on the median of their correction value to ensure an equal number of trials between classes, thereby controlling for potential biases in the estimates (Cohen, 2014). We then computed the Fourier spectrum using wavelets (width = 6; Gaussian width = 3) between 2 and 20 Hz with a 0.2 Hz step and calculated the ITPC using the formula:
For the mean resultant vector length (MRL), we used the following formula:
Results
Data were collected from 28 young adults performing a path integration task in a virtual reality environment combined with high-density mobile EEG (128 electrodes). Each path included a sequence of four stops, after which participants returned to the starting location (Fig. 1A). At each stop along the path, they physically turned to indicate their starting position and rated their confidence in their response. Then, a previously seen proximal landmark was presented at either the second or third stop of the sequence. Participants had to wait without moving for 1 s after presentation of the landmark and then update their prior estimates based on the newly available visual information (Fig. 1B). To disentangle the distinct contributions of translations and rotations to error accumulation during path integration, we systematically manipulated these parameters across three different paths configurations (Fig. 1E).
Experimental protocol and setup. A, Top-down view of the environment (left) and the first-person perspective experienced by participants (right). B, Sequence of the actions performed at a stop during landmark presentation. Participants pointed toward their starting position (left), then the landmark appeared for one second while they remained still (middle) and then rotated to correct their pointing (right). Only absolute errors were analyzed. C, Experimental setup worn by participants, including (1) HTC Vive Pro virtual reality headset, (2) 128-channel ANT Neuro EEG cap, (3) Zotac computer, (4) two cascading ANT Neuro amplifiers, and (5) two lithium batteries. D, The 8 × 12 m dark room, showing a participant at the start of a trial (top) and at the final stop (bottom) before returning to the starting point. E, The three path configurations used in the experiment to selectively manipulate the amount of translation and rotation.
Participants were equipped with a fully mobile virtual reality and EEG setup (Fig. 1C), enabling them to move freely within an 8 × 12 m room devoid of any sensory cues that could serve as landmarks (Fig. 1D). We spatially filtered the EEG data and subsequently source reconstructed the origin of the recorded brain dynamics focusing our analyses on the RSC and including data from 23 participants (Fig. 2).
Presentation of the pipeline used for EEG preprocessing. EEG preprocessing and clustering pipeline. Raw data were downsampled (500→250 Hz), line noise removed (Zapline plus), and bad channels interpolated (clean_raw_data, spherical interpolation). Time-domain cleaning (ASR, burst = 30) was applied, followed by re-referencing and high-pass filtering (1.75 Hz). Independent component analysis (AMICA, 2,000 iterations) was performed, with dipole fitting and component labeling. Preprocessed data were bandpass filtered (0.3–50 Hz), epoched, and outliers removed (<100 µV). ICs with ≥30% brain contribution and residual variance (RV) <0.15 were clustered using features (topographies, ERSP, power spectra, dipoles) and ROI-based repetitive clustering (N = 17, 5,000 iterations). The best clustering solution was selected based on spatial spread, mean RV, and ROI distance.
Landmarks allow for resetting the path integration system
In this initial behavioral analysis, we extracted participants’ pointing errors defined as the deviation between their pointing response to the starting location and the correct homing angle (hereafter referred to as “pointing error” for simplicity). We analyzed this error across the four different stops and three path configurations to assess how the landmark reduced the pointing error and whether this recalibration effect persisted over subsequent path segments (Fig. 3A). After comparing models using the AIC (Akaike, 1974), we applied a linear mixed model with the following parameters:
Behavioral results. A, Homing pointing error at each stop in the sequence and after landmark presentation. Light blue represents landmarks shown at the second stop, and coral represents those shown at the third stop. Lines indicate different path configurations (standard, distance, and rotation). B, Participant confidence ratings on a visual scale from 0 (not confident) to 1 (extremely confident). C, Absolute distance error for the homing segment. D, Mean pointing error for the fourth stop. E, Density plot of homing positions averaged across all participants and experimental conditions. The red star marks the starting position participants were instructed to return to.
In this model, we analyzed the pointing Error including as fixed factors: Stop, representing the stop at which participants performed the pointing task (1st, 2nd, 3rd, or 4th); Landmark_position, indicating the stop where the landmark was presented (2nd or 3rd stop); and Path_configuration, reflecting the navigation condition (standard, long-distance, or increased-rotation). Additionally, Stop was included as a random slope and intercept for each subject.
We found a significant main effect of stop position in the path sequence (F(3,38.55) = 24.51, p < 0.001,
Next, we examined pointing errors at the fourth stop and found that they were significantly larger when the landmark was presented at the second stop (M = 53.1 ± 4.16) than when the landmark was presented at the third stop (M = 30.7 ± 4.16; t(560) = 15.31, p < 0.001, d = 2.32, 95% CI = [1.99 2.66]). This suggests that while landmarks can temporarily reduce accumulated noise, the noise rapidly builds up again, indicating that the corrective effect seems to be short-lived. Considering the effect of path configuration, we reported a significant main effect on the error (F(2,587.99) = 85.18, p < 0.001,
To confirm that the landmark effectively reduced the accumulated error and to examine how the path configuration and the stop at which the landmark was presented (either Stop 2 or 3) influenced this correction, we applied the following model:
RSC theta activity correlates with the extent of landmark-based correction
To explore the neural basis of the observed landmark-based recalibration of the path integration system, we analyzed how theta activity in the RSC regions was modulated by the presentation of a landmark (Fig. 5). Theta activity (4–8 Hz) was examined over the period during which the participants were presented with the landmark and had to remain still for 1 s before rotating to adjust their pointing response.
Analysis of z-scored theta activity in the RSC revealed two distinct peaks (Fig. 4A,B). The first peak occurred ∼300 ms after the landmark was presented, while the second peak appeared at the onset of participants' rotational movements to correct their previous pointing, following the 1 s stationary period. Notably, theta activity in the RSC for the first peak exhibited a linear decrease as a function of correction magnitude (mean beta within the significant window, 340–476 ms: −0.197, pFDR < 0.05; Fig. 4C). In our hierarchical model, path configuration was included as a covariate to control for its potential influence on theta activity, given that we had previously observed an effect of path configuration on pointing error at the behavioral level. This allowed us to isolate the specific relationship between theta activity and correction magnitude. No significant effects of path configuration were observed at any time point following landmark presentation after FDR correction (all pFDR > 0.84), indicating that RSC theta activity was not modulated by path configuration. At this point, it is important to clarify how we interpret the magnitude of correction and how it reflects the degree of alignment between landmark information and self-motion cues. Small corrections suggest that the spatial representation was already well aligned with self-motion cues, requiring only minor adjustments of this representation before anchoring for later retrieval. In contrast, larger corrections indicate that the landmark prompted a substantial readjustment of the spatial representation. In the former case, smooth recalibration to stable values remains possible, a mechanism proposed by computational models to be supported by phase resetting (Monaco et al., 2011), underscoring the critical role of theta phase in spatial coding (e.g., the oscillatory interference model).
EEG results for the power analysis. A, z-scored time–frequency representation (2–40 Hz) aligned to the onset of landmark appearance and the physical rotation required for correction. B, Mean z-scored theta activity (4–8 Hz) within the same time window. C, Beta estimates of the effect of correction on theta activity. Black regions indicate statistically significant clusters identified through permutation testing with FDR correction.
To investigate this possible relationship between theta activity and landmark-based path integration updating, we analyzed ITPC (Makeig et al., 2002) in the RSC. Results of the cluster-based permutation tests indicated that, following landmark presentation, ITPC values were higher when participants made smaller corrections compared with larger corrections (pcluster < 0.05, for the 94–550 time window). This effect was particularly marked for theta activity, with a significant cluster observed between 254 and 512 ms after landmark presentation, while no difference was found during the rotation period (Fig. 5D). These results suggest that when the landmark information aligned with the prediction of participants based on self-motion cues (i.e., their pointing before landmark presentation), RSC theta activity exhibited greater power and phase coherence. In contrast, when participants required to update their spatial representation, theta power and phase coherence were reduced, a result that we can interpret as a decrease in phase resetting (Yeung et al., 2004, 2007; Min et al., 2007). This phase alignment effect, observed when the landmark matched participants' predicted spatial representation, was confirmed by the polar histogram of theta phase (Fig. 5E), with an increased concentration of phases between 200 and 270°. Statistical analyses of the MRL (Fig. 5F) indicated greater phase angle consistency when the landmark elicited small corrections but only in the time window after landmark presentation (F(1,23.22) = 12.757, p < 0.001,
EEG results for the phase analyses. A, ITPC for trials in which the landmark elicited a small correction, indicating that the landmark information was aligned with the path integration-derived representation. B, Same as A, but for trials when there was an important misalignment between each representations. C, Difference between high- and low-correction trials. The black outline indicates cluster-based significant differences. D, Theta band results (2–8 Hz) with standard error. The black bar indicates significant clusters from permutation testing between the two conditions, corrected for false discovery rate (FDR). E, Polar histogram of theta phase distribution relative to correction magnitude, presented across three 100 ms time windows: before and after landmark presentation and at a matched interval after rotation onset. F, Violin plot illustrating the mean resultant length, an index of phase angle consistency. Statistical comparisons for each time window were performed using a linear mixed model.
Our results also revealed a second peak of theta activity occurring when participants initiated the rotation. However, this activity was not related to the magnitude of correction, suggesting that theta activity during physical rotation may originate from a motoric component. This interpretation aligns with the proposed multifaceted role of low-frequency oscillations in spatial cognition, supporting their involvement in both spatial memory processes and the integration of body-related sensory information during movement (Ekstrom and Watrous, 2014). Building on this framework and recent studies in rodents reporting an association between acceleration and theta activity (Kropff et al., 2021), we also extracted both peak speed and peak acceleration from participants within a 500 ms time window following the onset of rotation, along with the corresponding theta activity in the RSC. We then applied linear mixed-effects models to examine the relationships between theta activity and these kinematic parameters. The results indicated that theta activity increased linearly with peak acceleration (β = 0.774, 95% CI = [0.238 1.311], t(849) = 2.832, p = 0.005). In contrast, no such relationship was observed for peak speed (t(861) = −1.24, p = 0.216). Text S2 provides extended results with control analyses supporting the movement-related origin of the second peak of RSC theta activity reported here, comparing also with other frequencies bands.
The degree of confidence modulates the integration of landmark-based information
We aimed to determine whether participants' confidence prior to landmark presentation influenced how they adjusted their pointing (Fig. 2B). We first assessed their ability to accurately judge their own performance using the visual scale, computing confidence-weighted accuracy (CWA which ranges from 0 to 1; see Material and Methods for details). Results showed that participants effectively used subjective confidence measures and accurately judged their performance, although this ability declined over successive stops (see Text S3 for detailed analyses).
We then investigated how the degree of confidence before the landmark presentation influenced the extent to which participants integrated landmark information to update their spatial representation. Specifically, we hypothesized that when participants are highly confident in their spatial representation based on self-motion cues, they would be less likely to update their response in light of added information available. Based on model comparisons using the AIC, we selected a linear mixed model with the following parameters:
Presentation of landmarks improves performance of outbound path, but only for long distances
Although our primary focus in this study was on homing pointing errors, participants also completed a homing path after the fourth stop by walking back to the remembered starting location. Our results (Fig. 2C) indicate that path configuration significantly influenced distance error (F(2,140) = 107.07, p < 0.001,
Discussion
This study investigated the neural dynamics underlying path integration and the integration of visual landmarks during naturalistic human spatial navigation. We found that although landmarks effectively reduced angular homing errors, it accumulated rapidly again following landmark presentation. Critically, while participants successfully exploited allothetic cues, correction efficacy was modulated by confidence in idiothetic spatial representations suggesting that stronger internal priors may down-weight newly available cues. High-density mobile EEG revealed theta-band activity in the RSC as a signature of this recalibration process. When only minor adjustments to the idiothetic representation were required, landmark presentation elicited increased theta power and robust phase resetting. In contrast, when substantial updating was needed, phase resetting was markedly decreased. Finally, we provide evidence for two dissociable mechanisms for tracking translation and rotation, each associated with distinct patterns of landmark-based error correction. Together, these findings reveal a dual role of theta oscillations in the RSC, reflecting both landmark-triggered recalibration of spatial representations and a motor-related response at the onset of rotation.
Differential impacts of rotational and translational noise on path integration
Consistent with our hypothesis, participants successfully corrected their homing responses after landmark presentation. However, error rapidly re-accumulated, resembling the accumulation observed before landmark presentation. This transient correction suggests that most accumulated error stems from noise and velocity gain bias, with minimal contribution from memory decay of the starting position (Stangl et al., 2020). Interestingly, participants showed greater pointing errors with increased rotations and greater homing errors when walking farther. These findings suggest that selectively increasing noise in rotational or translational components of path integration induces corresponding errors, supporting distinct systems for translation and rotation, consistent with vestibular organization (Angelaki and Cullen, 2008). Crucially, landmark reduced both error types, indicating that landmarks conveyed directional and positional information, a property of intramaze landmarks, unlike distal landmarks providing only directional cues (Doeller and Burgess, 2008). Furthermore, landmarks corrected condition-specific errors. With increased distance, homing accuracy improved when the landmark appeared near the homing response, but pointing errors showed no improvement. These findings indicate that translation and rotation are distinct components of path integration, with the RSC prioritizing task-relevant information (Chrastil et al., 2016, 2017).
RSC theta activity supports subtle adjustments of spatial representation by landmarks
Increased theta activity emerged ∼300 ms after landmark presentation in dipoles reconstructed around the RSC, indexing the neural correlates of path integration correction. This finding extends rodent literature to humans, supporting the RSC's role in processing stable environmental features (Fischer et al., 2020, 2024; Campbell et al., 2021). Human electrophysiological evidence remains limited, but intracranial EEG from the hippocampus suggests a role for slow-frequency oscillations in spatial updating during landmark processing (Ekstrom et al., 2005; Watrous et al., 2011). In rodents, invasive recordings showed hippocampal theta oscillations phase-locked to RSC activity (Alexander et al., 2020), underscoring the role of RSC–HPC connectivity in spatial cognition (Subramanian et al., 2024). Crucially, we observed not only increased theta following landmark presentation, but also a scaling of this activity with the magnitude of homing correction. This pattern was accompanied by increased theta phase alignment when landmark information aligned with the self-motion-based spatial representation. This concurrent increase in theta power, ITPC and phase alignment can be interpreted as a phase reset of oscillations in the RSC (Yeung et al., 2004, 2007; Min et al., 2007). In other words, when landmark information aligned with the spatial representation from self-motion cues, RSC theta activity exhibits increased phase resetting. Theta phase is proposed to play a key role in memory and spatial navigation (Lisman and Jensen, 2013; Kunz et al., 2019), with phase resetting creating optimal conditions for synaptic potentiation (McCartney et al., 2004; Hasselmo, 2005). This mechanism may serve to facilitate cumulative error correction through environmental sensory input (Burgess et al., 2007; Hasselmo, 2008), like proposed by attractor models, before anchoring the corrected spatial representation (Kunz et al., 2019; Schonhaut et al., 2024). Supporting this interpretation, Baker and Holroyd (2009, 2013) suggested that rewards elicit a phase resetting of theta rhythm. This was later extended by Lin et al. (2022) and Güth et al. (2025), who proposed that theta phase resetting may periodically realign the theta rhythm (e.g., to a zero-phase reference) to prevent cumulative drift of spatial information, anchoring spatial representations to salient spatial cues. Here, we provide direct evidence, with a correlation between theta phase resetting and correction magnitude of the spatial representation, also confirming fMRI findings that RSC contains spatial representations for both landmarks and self-motion cues (Chen et al., 2025).
From a mechanistic perspective, our findings extend the framework proposed by Campbell et al. (2018) in mice to humans. Here we propose that theta activity in the RSC may reflect the model proposed with the sub-critical and super-critical regimes of attractor network dynamics during navigation. In the sub-critical regime, when discrepancies between path integration and landmark representations are minimal, we observed increased theta-band power and phase resetting compared with larger mismatch cases. This enhanced neural synchronization could reflect attractor network continuous fine adjustments to align internal spatial representations with external cues, supporting stable and accurate navigation. Computational modeling by Monaco et al. (2011) already suggested that a single cue can effectively reset oscillator phases, enabling corrections of the path integrator (Canavier, 2015). However, this phase alignment may occur only when cue mismatch is minimal, as reported in our small corrections trials, and prior reports of smooth head direction resetting (Valerio and Taube, 2012). In cases of larger discrepancies (i.e., super-critical regime), these fine corrections though phase resetting may be insufficient. Thus, landmarks trigger the need for significant reworking of the spatial representation, with a remodeling of the head direction system (Valerio and Taube, 2012) before transferring the corrected information across different brain regions through theta activity (Watrous et al., 2013; Place et al., 2016). In the present results, we observed reduced RSC theta activity and phase alignment, possibly reflecting delayed reprocessing of spatial representations. This remapping may involve increased connectivity between the RSC and the hippocampus, an hypothesis that warrants investigation through intracranial recordings. Taken together, our noninvasive findings provide strong evidence for RSC theta activity supporting landmark-based updating of spatial representations initially formed through path integration in freely moving humans.
RSC theta activity reflects vestibular encoding of rotation beyond landmark processing
Beyond landmark processing, we observed increased RSC theta activity during participants' physical rotation for angular homing adjustments. However, this second theta peak did not scale with correction magnitude but instead varied linearly with peak acceleration These results replicate rodent findings during translational movement (Kropff et al., 2021), extending this to mobile noninvasive recordings in humans. This also aligns with the proposed key role of the RSC to compute heading rotation information. As a central hub in the head direction network, the RSC supports the integration of self-motion cues for maintaining a stable sense of orientation (Chrastil et al., 2016, 2017; Gramann et al., 2021). Thus, our findings support a multifaceted role for low-frequency oscillations (Ekstrom and Watrous, 2014), with theta activity involved in both landmark processing and vestibular encoding of rotational acceleration.
Subjective confidence modulates landmark use and cue integration for spatial updating
Finally, after each homing response, participants reported their subjective confidence. They used the scale effectively to assess performance (Fleming and Lau, 2014), though this ability declined across stops. This suggests that participants can reasonably judge their performance, but this judgment is affected by accumulating noise. Subjective confidence has been proposed to reflect a Bayesian representation of probability (Geurts et al., 2022). The Bayesian framework provides a principled approach for optimally integrating multiple sensory cues, particularly when using noisy sensory input (Ernst and Banks, 2002; Seilheimer et al., 2013). By combining prior knowledge with sensory likelihoods to compute posterior distributions, Bayesian models support optimal perception and reliable spatial estimates despite variable inputs (Cheng et al., 2007; Newman et al., 2023). Within this framework, greater confidence in prior knowledge is thought to reduce the influence of new, discrepant information (Meyniel et al., 2015). Consistent with this idea, we found that when participants were more confident in their prior pointing prior, they relied less on landmark to adjust their response, leading to less effective corrections. Our results provide evidence in spatial navigation that prior confidence influences cue weighting in a Bayesian framework, paving the way for further investigations using this easy-to-use confidence metric.
To conclude, our findings provide among the first noninvasive neuroimaging evidence that theta oscillations support landmark-based correction of path integration during naturalistic human navigation. While participants used landmarks to recalibrate spatial representation, this correction was modulated by prior confidence, highlighting the influence of internal priors on sensory integration. Theta phase resetting in the RSC facilitated spatial correction when only fine adjustments were sufficient, providing support for emerging mechanistic models of spatial updating. We also observed a second theta peak during rotation, driven by acceleration rather than homing correction, highlighting the multifaceted role of theta in human spatial navigation. Overall, our results underscore the potential of mobile EEG to reveal neural mechanisms of dynamic spatial cognition, offering new insights into how the brain flexibly integrates multimodal signals for naturalistic navigation.
Data Availability
All the raw data and the analysis codes generated for the present study are available online on the OSF repository of the study.
Footnotes
This research would not have been possible without the generous help of our volunteer participants, and we are truly grateful for their support. We also thank Catherine Buchanan for her careful reading of the manuscript and her feedback. This work was supported by the French government through the France 2030 investment plan managed by the National Research Agency (ANR), as part of the Initiative of Excellence Université Côte d’Azur under reference number ANR-15-IDEX-01 and, in particular, by the interdisciplinary Institute for Modeling in Neuroscience and Cognition (NeuroMod) of Université Côte d’Azur.
The authors declare no competing financial interests.
- Correspondence should be addressed to Clément Naveilhan at clement.naveilhan{at}univ-cotedazur.fr or Stephen Ramanoël at stephen.ramanoel{at}univ-cotedazur.fr.
