Intention Aware Robot Crowd Navigation with Attention-Based Interaction Graph (2203.01821v4)

Abstract: We study the problem of safe and intention-aware robot navigation in dense and interactive crowds. Most previous reinforcement learning (RL) based methods fail to consider different types of interactions among all agents or ignore the intentions of people, which results in performance degradation. To learn a safe and efficient robot policy, we propose a novel recurrent graph neural network with attention mechanisms to capture heterogeneous interactions among agents through space and time. To encourage longsighted robot behaviors, we infer the intentions of dynamic agents by predicting their future trajectories for several timesteps. The predictions are incorporated into a model-free RL framework to prevent the robot from intruding into the intended paths of other agents. We demonstrate that our method enables the robot to achieve good navigation performance and non-invasiveness in challenging crowd navigation scenarios. We successfully transfer the policy learned in simulation to a real-world TurtleBot 2i. Our code and videos are available at https://sites.google.com/view/intention-aware-crowdnav/home.

Intention-Aware Robot Crowd Navigation with Attention-Based Interaction Graph

The paper "Intention Aware Robot Crowd Navigation with Attention-Based Interaction Graph" proposes an advanced approach for enabling safe and efficient robot navigation in densely populated and dynamic environments. The authors highlight that previous strategies, predominantly based on reinforcement learning (RL), often ignore the diverse and crucial interactions among agents or fail to adequately account for the intentions of human agents, consequently undermining navigation performance. This paper presents a novel solution by integrating attention mechanisms within recurrent graph neural networks, allowing for the effective capture of heterogeneous interactions both spatially and temporally.

Methodology Overview

The paper introduces a Spatio-Temporal Interaction Graph (sti-graph) to represent the crowd navigation scenario, modeling interactions among all visible agents through graphical structures. The agents are represented as nodes, with robot-human (RH) and human-human (HH) interactions being distinctly modeled through separate sets of edges. This distinction is crucial since RH and HH interactions bear different implications for the robot's decision-making process. The integration of temporal functions across successive timesteps enables the model to factor in historical data and anticipated movements, overcoming the limitations of shortsighted, reaction-based methodologies.

To further enhance navigation decisions, the authors incorporate future trajectory predictions of human agents into the RL framework. Trajectory predictors forecast the paths of humans over several time steps ahead, and these forecasts are integrated into the decision-making process via a novel reward function that discourages the robot from infringing upon the predicted trajectories of humans.

Attention Mechanism and Network Architecture

The core of the proposed solution lies in the utilization of attention mechanisms, critical for differentiating the relative importance of interactions among agents in crowded environments. The robot employs separate multi-head attention networks for RH and HH interactions, allowing for context-specific processing. These networks are adeptly combined with a Gated Recurrent Unit (GRU) that processes the spatial-temporal data, culminating in a robust system for real-time crowd navigation.

The network is trained using Proximal Policy Optimization (PPO), a stable and efficient RL algorithm. This setup allows the robot to predict and reason about both immediate and delayed agent movements, promoting socially aware and non-invasive behaviors.

Experimental Evaluation

Two simulated environments were utilized to evaluate the performance of the proposed approach. The methodology was compared against established baselines such as ORCA, Social Force (SF), and decentralized structural RNN (DS-RNN). The results consistently demonstrated superior navigation metrics, including higher success rates and reduced path lengths, when using the proposed prediction-aided models. Notably, by modeling HH interactions and leveraging future trajectory predictions, the robot exhibited improved social compliance and awareness, significantly minimizing intrusion times into human paths.

Moreover, real-world trials conducted with a TurtleBot 2i validated the transferability of the learned policies to physical robots amidst real human interactions. Despite certain limitations associated with environmental constraints, the robot effectively navigated the crowd and adhered to socially compliant behavior.

Implications and Future Directions

This research presents noteworthy implications for the development of autonomous systems functioning in crowded and dynamic environments. By explicitly modeling and utilizing temporal interactions, along with informed trajectory predictions, the proposed framework elevates the level of sophistication achievable in robot navigation tasks. Furthermore, the decoupling of interaction types through differentiated attention mechanisms sets a foundation for future studies to refine crowd navigation algorithms further.

Future work could involve enhancing the realism of simulated environments or the collection and utilization of real-world pedestrian trajectory datasets to refine the RL training process. Addressing the mutual influence dynamics between humans and robots could also serve as a promising direction to expand and improve the overall system's efficacy. Together, these advancements could significantly push the frontier of socially aware autonomous navigation systems.