"Smart" Robot Navigation System Proposed by Berkeley team

Figure ｜ BADGR Autonomous planning and navigation in real-time (Source: Berkeley)

Mobile robot navigation is usually considered to be a geometric problem. The goal of the robot is to perceive the geometry of obstacles in the environment to plan a collision-free path to a specified location. At present, from indoor navigation to automatic driving, the main method of autonomous navigation is to let the robot construct a map, position itself on the map, and use the map to plan and execute the action that causes the robot to reach the target. This simultaneous positioning and mapping (SLAM ) And path planning methods have achieved impressive results and are the basis of the most advanced autonomous navigation technology. However, this method still has limitations, such as performance degradation in texture-free scenes. To ensure that the robot's navigation behavior in the environment becomes better and better, it needs to rely on more and more expensive sensors, and in an open environment, mobile robot autonomous navigation will also face more challenges, and sometimes it is difficult to use pure Geometric analysis to solve.

Figure | Off-road environment and town environment (Source: Berkeley)

For example, to reach a destination on foot, you need to walk through lush grass. For humans, it is better to walk directly over. When people push a car, they prefer to walk on a relatively flat road. These seeming judgments that can be made without thinking are very difficult for today ’s autonomous navigation mobile robots, and it is likely that the decision will fail: they will think that the tall grass is the same obstacle as the concrete wall, and they do n’t understand the smooth selection. The difference between rough roads and bumpy roads.

Picture ｜ Clear path Jackal mobile robot platform (Source: Berkeley)

Because most mobile robots think purely based on geometry, from the perspective of semantic understanding, they use human-provided traversability or computer vision methods trained on-road labels to achieve, but traversability, bumpiness, etc. The attributes related to mobility are the physical characteristics of the natural environment. Can the robot infer autonomous navigation capabilities directly from the image? And choose the most suitable path planning like a human to reach the goal. AI researchers from the University of California, Berkeley have developed a solution, which is a completely autonomous, self-improving mobile robot navigation system based on end-to-end learning. Mobile robots can use their own experience in the real world to learn the physical properties of the environment without any simulation or manual supervision, the team called this robot learning system BADGR: Berkeley autonomous driving ground robot.

First look at the hardware configuration. The researchers used a Clear path Jackal mobile robot as a test platform. The size of this robot is 508 mm × 430 mm × 250 mm and weighs 17 kg. It is very suitable for driving in urban and off-road environments. The default sensor kit includes a six-freedom IMU (for measuring linear acceleration and angular velocity), a GPS unit for approximate global position estimation and an encoder for measuring wheel speed. The researchers also added new sensors on top of the robot: two forward-facing 170-degree fields of view 640 × 480 pixel cameras, a 2D lidar, and a compass. The robot is equipped with an NVIDIA Jetson TX2 computer, which is ideal for running deep learning applications. The data is saved to an external SSD, which is large enough and fast enough to store 1.3 GB of sensor data stream per minute. The team remotely monitors experiments, video streaming, and remote operations when necessary via a 4G smartphone installed on top of the robot. 

The focus of the next work is divided into four steps: 1. Collect data autonomously; 2. Automatically label data through self-supervision; 3. Train an image-based neural network prediction model; 4. Use the prediction model to plan and execute the robot to complete the navigation task.

              Figure ｜ BADGR Autonomous planning and navigation in real-time (Source: Berkeley)

Researchers have designed data collection methods that can collect large amounts of diverse data for training with minimal manual intervention. Due to the high cost of collecting data using real-world robotic systems, the team chose to use non-strategic learning algorithms in order to be able to use any control strategy to collect data and train all the data. In addition, the second consideration when designing a data collection strategy is to ensure that the environment is fully explored, while also ensuring that the robot performs the sequence of actions that it actually wants to perform during the test. A simple unified random control strategy is not enough, because the robot will mainly drive straight lines due to the robot's linear and angular velocity interaction interface, which will lead to insufficient exploration and unrealistic test time action sequences. Therefore, the team uses time-dependent random walk control strategies to collect data. Test two, the task of reaching the specified GPS position in the off-road environment. The SLAM + P strategy incorrectly marks the grass as an impenetrable obstacle, so it rotates in place to try to find a traversable path, but after the rotation fails to detect any traversable path, the robot is difficult to move forward. In contrast, the BADGR method has learned from experience, and some tall grass can indeed be traversed, so it can successfully continue to guide the robot to the target without mistakenly thinking that the grass is an obstacle. This is because BADGR has learned from experience. The grass is actually traversable.

In addition to being able to understand the physical properties of the environment, a key aspect of BADGR is its ability to continuously self-supervise and collect more and more data to improve the model in real time. To prove this ability, the researchers conducted a controlled study in which BADGR collected and trained data from an area and then moved to a new target area, where the initial navigation failed, but then the new area was collected and trained.

Figure | As more data is collected, BADGR's intelligence continues to improve (Source: Berkeley).

This experiment not only proves that BADGR can improve after collecting more data, but when BADGR encounters a new environment, the previously collected experience can actually accelerate learning. As BADGR automatically collects data in more and more environments, the time required to successfully learn navigation in each new environment will be less and less.