SoRTS: Learned Tree Search for Long Horizon Social Robot Navigation

TL;DR

This work tackles trustworthy long-horizon social robot navigation by integrating Monte Carlo Tree Search with offline socially-aware motion predictions, a normative reference guide, and a global cost map. The SoRTS framework Biases MCTS through three modules to enable collision-free, socially compliant planning in dynamic multi-agent environments, demonstrated in general aviation with a high-fidelity simulator and a 26-pilot user study. Results show SoRTS achieves performance and trust comparable to competent human pilots and significantly outperforms its ablation, with additional self-play validation across increasingly complex scenarios. The approach is domain-agnostic and supported by a dedicated simulator (X-PlaneROS) and rich datasets (TrajAir), highlighting a practical path toward trustworthy aerial autonomy and potential extension to other shared-space domains.

Abstract

The fast-growing demand for fully autonomous robots in shared spaces calls for the development of trustworthy agents that can safely and seamlessly navigate in crowded environments. Recent models for motion prediction show promise in characterizing social interactions in such environments. Still, adapting them for navigation is challenging as they often suffer from generalization failures. Prompted by this, we propose Social Robot Tree Search (SoRTS), an algorithm for safe robot navigation in social domains. SoRTS aims to augment existing socially aware motion prediction models for long-horizon navigation using Monte Carlo Tree Search. We use social navigation in general aviation as a case study to evaluate our approach and further the research in full-scale aerial autonomy. In doing so, we introduce XPlaneROS, a high-fidelity aerial simulator that enables human-robot interaction. We use XPlaneROS to conduct a first-of-its-kind user study where 26 FAA-certified pilots interact with a human pilot, our algorithm, and its ablation. Our results, supported by statistical evidence, show that SoRTS exhibits a comparable performance to competent human pilots, significantly outperforming its ablation. Finally, we complement these results with a broad set of self-play experiments to showcase our algorithm's performance in scenarios with increasing complexity.

Figures (4)

• Figure 1: Our flight simulator setup and a user study experiment. In each, the User interacts with a Human pilot, our proposed algorithm, SoRTS, and our Ablation. The figure also shows a resulting interaction between a User and SoRTS.
• Figure 2: An overview of SoRTS, a Monte Carlo Tree Search (MCTS)-based planner for social robot navigation which provides long-horizon simulations, collision checking and goal conditioning. Its tree search is biased by three components; a Social Module, a Reference Module and a Cost Map. The social module uses a socially-aware motion prediction model to predict a set of possible future states given the social dynamics of the scene. The reference module provides the agent with a global path that embodies navigation guidelines the agent must follow. The cost map encodes a global visitation to encourage the agent to move toward more desirable regions.
• Figure 3: User study results. Left: Each row shows the resulting trajectories of a User interacting with our Human pilot, SoRTS, and the Ablation; reference paths are shown as solid lines, executed ones as dashed lines. The top row shows a user that successfully followed the expected path. We can observe that the ablation did not follow the reference path as smoothly as SoRTS, and also cut short when approaching the goal. The bottom row shows a user that did not follow the expected path. Here, SoRTS still managed to navigate properly. In contrast, the ablation behaved erratically, unsafely crossing over the runway twice. Right: Box-plots showing per-algorithm results. The top ones show the avg. efficiency (a) and safety (b) scores given by the users. The bottom ones show the avg. RE (a) and LS (b) metrics obtained from the trajectory data.
• Figure 4: Self-play results. The top row shows examples where the Ablation agents fail to resolve social conflicts and end in a loss of separation situation. With the same initial conditions, the bottom row shows that SoRTS agents are able to adjust their paths to properly resolve these situations.