Long-Term Human Trajectory Prediction using 3D Dynamic Scene Graphs

TL;DR

This paper introduces a novel method for long-term indoor human trajectory prediction using 3D dynamic scene graphs and language models, outperforming existing short-term focused methods by modeling complex human-environment interactions.

Contribution

It combines 3D scene graphs with large language models to predict human interactions and trajectories over 60 seconds, addressing limitations of prior short-term prediction methods.

Findings

Achieves 54% lower negative log-likelihood compared to baselines.

Reduces displacement error by 26.5% over 60 seconds.

Introduces a new semi-synthetic dataset with interaction annotations.

Abstract

We present a novel approach for long-term human trajectory prediction in indoor human-centric environments, which is essential for long-horizon robot planning in these environments. State-of-the-art human trajectory prediction methods are limited by their focus on collision avoidance and short-term planning, and their inability to model complex interactions of humans with the environment. In contrast, our approach overcomes these limitations by predicting sequences of human interactions with the environment and using this information to guide trajectory predictions over a horizon of up to 60s. We leverage Large Language Models (LLMs) to predict interactions with the environment by conditioning the LLM prediction on rich contextual information about the scene. This information is given as a 3D Dynamic Scene Graph that encodes the geometry, semantics, and traversability of the environment into a hierarchical representation. We then ground these interaction sequences into multi-modal spatio-temporal distributions over human positions using a probabilistic approach based on continuous-time Markov Chains. To evaluate our approach, we introduce a new semi-synthetic dataset of long-term human trajectories in complex indoor environments, which also includes annotations of human-object interactions. We show in thorough experimental evaluations that our approach achieves a 54% lower average negative log-likelihood and a 26.5% lower Best-of-20 displacement error compared to the best non-privileged (i.e., evaluated in a zero-shot fashion on the dataset) baselines for a time horizon of 60s.