Embedded Files
Abstract
Abstract
Safe and efficient crowd navigation for mobile robot is a crucial yet challenging task.
Safe and efficient crowd navigation for mobile robot is a crucial yet challenging task.
Previous work has shown the power of deep reinforcement learning frameworks to train efficient policies. However, their performance deteriorates when the crowd size grows. We suggest that this can be addressed by enabling the network to identify and pay attention to the humans in the crowd that are most critical to navigation.
Previous work has shown the power of deep reinforcement learning frameworks to train efficient policies. However, their performance deteriorates when the crowd size grows. We suggest that this can be addressed by enabling the network to identify and pay attention to the humans in the crowd that are most critical to navigation.
We propose a novel network utilizing a graph representation to learn the policy. We first train a graph convolutional network that accurately predicts human attention to different agents in the crowd. Then we incorporate the learned attention into a graph-based reinforcement learning architecture.
We propose a novel network utilizing a graph representation to learn the policy. We first train a graph convolutional network that accurately predicts human attention to different agents in the crowd. Then we incorporate the learned attention into a graph-based reinforcement learning architecture.
The proposed attention mechanism enables the assignment of meaningful weightings to the neighbors of the robot, and has the additional benefit of interpretability.
The proposed attention mechanism enables the assignment of meaningful weightings to the neighbors of the robot, and has the additional benefit of interpretability.
Experiments on real-world dense pedestrian datasets with various crowd sizes demonstrate that our model outperforms state-of-art methods by 18.4\% in task accomplishments and by 16.4\% in time efficiency.
Experiments on real-world dense pedestrian datasets with various crowd sizes demonstrate that our model outperforms state-of-art methods by 18.4\% in task accomplishments and by 16.4\% in time efficiency.
Ablation study to investigate the contribution of graph structure and human attention.
Ablation study to investigate the contribution of graph structure and human attention.
SA-GCNRL represents the model trained and tested with graph representation but attention from SARL. UARL represents the SARL with uniform attention.
SA-GCNRL represents the model trained and tested with graph representation but attention from SARL. UARL represents the SARL with uniform attention.
Figure 1
Figure 1
Performance on real-world pedestrian datasets (GGCNRL VS. SARL)
Performance on real-world pedestrian datasets (GGCNRL VS. SARL)
Below two examples show the performance comparison of our method with state-of-the-art method in two different testing cases.
Below two examples show the performance comparison of our method with state-of-the-art method in two different testing cases.
The left column illustrates good performance of our GGCNRL method. The learned attention weights of our attention network has been marked in the upper area of the pedestrians.
The left column illustrates good performance of our GGCNRL method. The learned attention weights of our attention network has been marked in the upper area of the pedestrians.
The right column shows corresponding performance of SARL method . The weights marked are calculated from self-attention weights in SARL.
The right column shows corresponding performance of SARL method . The weights marked are calculated from self-attention weights in SARL.
GGCNRL (Ours) SARL (Chen et.al, 2019)
More tests with different crowd densities
More tests with different crowd densities
Below show several examples of GGCNRL testing on different datasets, including both successful cases and failure case.
Below show several examples of GGCNRL testing on different datasets, including both successful cases and failure case.
NYC-GC Hotel
Zara2 Zara2 (failure case)
Details about the simulation environments
Details about the simulation environments
We build the simulation environments from different trajectory datasets , as shown below. The average human numbers per frame of each dataset are shown as well. The black triangles show the start positions and goals of the robot and they are 4 meters away from the center black circle.
We build the simulation environments from different trajectory datasets , as shown below. The average human numbers per frame of each dataset are shown as well. The black triangles show the start positions and goals of the robot and they are 4 meters away from the center black circle.
Page updated
Google Sites
Report abuse