LocoVR: Multiuser Indoor Locomotion Dataset

　　Kojiro Takeyama¹² , Yimeng Liu¹, Misha Sra¹

1: University of California Santa Barbara

2: Toyota Motor North America

Visit our github repo to access the dataset and code.

Visit arxiv to access our paper (Our paper has been accepted to ICLR 2025!)

Abstract

Understanding human locomotion is crucial for AI agents such as robots, particularly in complex indoor home environments. Modeling human trajectories in these spaces requires insight into how individuals maneuver around physical obstacles and manage social navigation dynamics. These dynamics include subtle behaviors influenced by proxemics - the social use of space, such as stepping aside to allow others to pass or choosing longer routes to avoid collisions. Previous research has developed datasets of human motion in indoor scenes, but these are often limited in scale and lack the nuanced social navigation dynamics common in home environments.

To address this, we present LocoVR, a dataset of 7000+ two-person trajectories captured in virtual reality from over 130 different indoor home environments. LocoVR provides full body pose data and precise spatial information, along with rich examples of socially-motivated movement behaviors. For example, the dataset captures instances of individuals navigating around each other in narrow spaces, adjusting paths to respect personal boundaries in living areas, and coordinating movements in high-traffic zones like entryways and kitchens. Our evaluation shows that LocoVR significantly enhances model performance in three practical indoor tasks utilizing human trajectories, and demonstrates predicting socially-aware navigation patterns in home environments.

Figure1. Overview of LocoVR dataset

LocoVR Dataset

Table1 summarizes the statistics of existing human trajectory and motion datasets and LocoVR. Our dataset contains 2500K frames of human trajectories in 131 scenes. The number of scenes surpasses all the real human motion datasets. We collected two-person trajectories that are geometrically and socially aware, which is not included in most of the compared datasets. The number of trajectories is 7071 in total.

LocoVR facilitates the enhancement of task performances from both geometric and social perspectives in unseen, complex, and confined indoor environments. Also, it includes full-body human poses, along with head orientation data in addition to trajectories. These additional observations can facilitate a deeper understanding of human locomotion and enhance model performance.

Table1. Statistics of existing human motion datasets and our LocoVR dataset.

LocoVR_website_distance.mov

Maintain social distance when passing others

LocoVR_website_yield.mov

Yield a path to allow others to pass

LocoVR_website_detour.mov

Take a longer detour to avoid interfering with others

Figure2. Social motion behaviors in indoor scenes

Evaluation

Global path prediction

Figure3 demonstrates the result of global path prediction. Our LocoVR dataset, with its large-scale diversity, enables the prediction of geometry-aware smooth paths, even in complex and previously unseen environments (Figure3 (a)). Furthermore, LocoVR demonstrates its ability to predict social motion behaviors (Figure3 (b)-(d)), attributed to its capacity to learn these behaviors across a wide variety of scenes. In contrast, the trajectories generated by the prior dataset (GIMO) are unstable, lack smoothness, and are unable to handle social motion behaviors due to its limited scale, diversity, and absence of multi-person data.

predicted_geometry.mp4

(a) Navigation along with complex geometry

predicted_yield.mp4

predicted_social.mp4

(b) Maintain social distance when passing through others

predicted_detour.mp4

(d) Take a longer detour to avoid interference with others

Figure3. Results of global path prediction (static and goal-conditioned path prediction)

Trajectory prediction

Figure4 shows the result of trajectory prediction. The model trained on LocoVR is able to predict a trajectory, taking into account both the obstacles and the other person's movement. In contrast, predicted trajectory distribution with the prior dataset (GIMO) is spread to multiple directions, resulting in collisions with other people since it does not include multi-person data.

Trajectory_prediction.mp4

Figure4. Result of trajectory prediction (dynamic and non-goal conditioned path prediction)

Goal prediction

Figure5 shows the results of object prediction. In LocoVR, as the trajectory progresses, the probability distribution of the goal area narrows down near the true goal object. This is due to the model learning from the dataset a policy that narrows down the goal area based on the areas already passed and the current heading direction. On the other hand, the prior dataset (GIMO) does not include a sufficient number of trajectories or scenes to learn a policy applicable to unseen scenes, resulting in the probability distribution of the goal area not being appropriately narrowed down.

Goal_prediction.mp4

Figure5. Result of goal prediction based on past trajectory

Video

VIDEO_ICLR2025_final.mov

Citation

BibTex

@inproceedings{

takeyama2025locovr,

title={Loco{VR}: Multiuser Indoor Locomotion Dataset in Virtual Reality},

author={Kojiro Takeyama and Yimeng Liu and Misha Sra},

booktitle={The Thirteenth International Conference on Learning Representations},

year={2025},

url={https://arxiv.org/abs/2410.06437}

}

Page updated

Google Sites

Report abuse