Introduction
Methods
Overview
-
The visual SLAM module is the basis of the navigation framework Safe-Nav. On the one hand, it provides the 2D navigation map and robot pose for the global planner module. On the other hand, it generates the feature point maps for the DRL-based local planner module as one of the spatial observation inputs.
-
The global planner module is responsible for planning a local waypoint on the 2D navigation map based on the current robot pose. Then, the relative distance and angle between the current robot pose and the local waypoint are computed as the non-spatial observation inputs for the DRL-based local planner module.
-
The DRL-based local planner module can directly map both non-spatial inputs and spatial inputs from the RGB-D camera and other modules to safe actions towards local waypoints while avoiding collisions and tracking failure.
Visual SLAM module
Global planner module
DRL-based local planner module
Problem formulation
Observation space
Action space
Movement modes
Self-attention modules
Reward function
Training
Parameter description | Parameter | Value |
---|---|---|
Maximum linear velocity | \(\nu _{max}\) | 0.2 |
Minimum linear velocity | \(\nu _{min}\) | 0.1 |
Angular velocity | \(\omega \) | 12 |
Constant for reaching the local waypoint | \(R_{wp}\) | 20 |
Distance threshold within the local waypoint is reached | \(D_{wp}\) | 0.2 |
Weight for the change of relative distance | \(\omega _{1}\) | 50 |
Weight for the change of relative angle | \(\omega _{2}\) | 1 |
Constant for colliding with obstacles | \(R_{CO}\) | 20 |
Constant for tracking failure | \(R_{tf}\) | 20 |
Weight for the change of minimum depth value | \(\omega _{3}\) | 50 |
Weight for the change of average depth value | \(\omega _{4}\) | 60 |
Weight for the change of the number of feature points | \(\omega _{5}\) | 0.1 |
Time steps threshold for continuous back motions | \(D_{back}\) | 3 |
Constant for continuous back motions | \(R_{back}\) | 5 |
Time steps of each local planner episode | k | 15 |
Threshold of the minimum depth value | \(Th_{depth}\) | 0.3 |
Time steps of entering the collision-avoidance mode | \(D_{depth}\) | 2 |
Threshold of the number of feature points | \(Th_{points}\) | 280 |
Time steps of entering the tracking-failure-avoidance mode | \(D_{points}\) | 2 |
Evaluation metrics
Simulation setup
Results
Training analysis
Performance evaluation
Comparison with the existing DRL-based methods
Method | Success rate | SPL | Collision rate |
---|---|---|---|
PPO | 17.92 | 16.88 | 80.83 |
DD-PPO | 6.70 | 6.60 | 93.30 |
Safe-Nav | 55.39 | 51.91 | 19.82 |
Comparison with classic methods
Scene | SR | SPL | CR | TFR | ||||
---|---|---|---|---|---|---|---|---|
DWA | Safe-Nav | DWA | Safe-Nav | DWA | Safe-Nav | DWA | Safe-Nav | |
Ed | 19.35 | 45.07 | 17.19 | 41.08 | 19.35 | 15.49 | 61.29 | 38.03 |
Pa | 52.17 | 56.25 | 42.09 | 55.18 | 39.13 | 30.21 | 8.70 | 13.54 |
Gr | 20.83 | 47.22 | 16.58 | 43.73 | 37.50 | 25.00 | 41.67 | 27.78 |
Ri | 8.33 | 45.95 | 5.24 | 41.05 | 41.67 | 16.62 | 50.00 | 37.84 |
De | 16.67 | 82.35 | 12.96 | 78.51 | 33.33 | 11.76 | 50.00 | 5.88 |
Total | 23.02 | 55.39 | 18.53 | 51.91 | 33.33 | 19.82 | 43.65 | 24.61 |
Scene | SR | SPL | CR | TFR | ||||
---|---|---|---|---|---|---|---|---|
Baseline | Safe- Nav | Baseline | Safe- Nav | Baseline | Safe- Nav | Baseline | Safe- Nav | |
Ed | 37.50 | 45.07 | 37.13 | 41.08 | 18.75 | 15.49 | 43.75 | 38.03 |
Pa | 50.70 | 56.25 | 49.71 | 55.18 | 40.85 | 30.21 | 8.45 | 13.54 |
Gr | 41.46 | 47.22 | 38.27 | 43.73 | 26.83 | 25.00 | 31.71 | 27.78 |
Ri | 26.87 | 45.95 | 25.64 | 41.05 | 28.35 | 16.62 | 44.78 | 37.84 |
De | 41.67 | 82.35 | 40.28 | 78.51 | 30.00 | 11.76 | 28.33 | 5.88 |
Total | 39.64 | 55.39 | 38.21 | 51.91 | 28.96 | 19.82 | 31.40 | 24.61 |