Better than Spot? Science's cover article reveals how the robot dog ANYmal trots "by feel"

Author | Fu Jing

This is Spot, a robot dog worth RMB 527,300 from Boston Dynamics in the United States. It can dance, act as a security guard, herd sheep, and conduct patrol inspections. It can be said that it is proficient in everything.

Recently, just as a Spot traveled to the Chernobyl Nuclear Power Plant in Ukraine to survey the distribution of radiation contamination, a robot dog from a Swiss company ANYbotics also made its debut.

This red dog is also said to be good at walking freely in complex terrain.

The name of this dog that is more like an animal is ANYmal. It also appeared on the cover of the latest issue of Science Robotics, a subsidiary of Science and a top robotics journal.

On October 21, 2020, a paper revealing how the ANYmal robot dog can navigate complex terrain with ease was published in Science Robotics, titled Learning quadrupedal locomotion over challenging terrain.

The research team for this study comes from the Robotics Systems Laboratory of ETH Zurich, the Robotics and Artificial Intelligence Laboratory of the Korea Advanced Institute of Science and Technology, and the Intel Intelligent Systems Laboratory.

Gao Feng, professor at Shanghai Jiao Tong University and chief scientist of the National 973 Program, once said during the 2018 World Robot Conference:

Wheeled robots will eventually move towards legged robots, and quadrupedal robots will have better mobility.

As the paper describes, legged locomotion can extend the range of a robot's activities into some extremely challenging environments. However, traditional legged locomotion controllers rely on "state machines" that explicitly trigger motion primitives and reflexes.

It can be said that this design ensures that the robot can perform quadrupedal movement in various complex scenarios, but it does not reach the level of movement of animals in nature. After all, highly irregular contours, diverse terrains, slippery surfaces, and obstacles are all potential obstacles. In this case, the robot may lose balance due to slipping, or even cause catastrophic failure.

What’s more serious is that the robot cannot obtain accurate information about the physical properties of the terrain - receptive sensors including cameras and lidar cannot reliably measure physical properties (such as friction and flexibility).

To address this situation, researchers have proposed a robust controller that incorporates proprioceptive feedback into motion control. This feedback comes from two of the most durable and reliable sensors for legged robots: joint encoders and inertial measurement units (IMUs).

Previously, a research team has successfully realized the application of legged robots from simulation to physical environment through the following two achievements:

1. Modeling of physical systems including brakes;

2. Physical parameter randomization: Parameters can differ between simulation and real-world environments, making the controller robust to a wide range of situations without having to model them accurately up front.

However, after training the controller in a simulation environment, researchers found that ANYmal was still not capable of navigating rough terrain with ease.

For this reason, researchers introduced three more steps (as shown below).

1. Temporal Convolutional Network (TCN): can generate actuation based on an extended history of proprioceptive states.

2. Privileged learning: The training process is broken down into two stages. First, a "teacher" is trained that can perceive the terrain and the contact between the robot and the ground. Then the "teacher" guides the purely proprioceptive "student" (i.e., the controller) to learn.

3. Adaptively “synthesize” terrain: Based on the controller’s performance at different stages of the training process, an accurate perception of the terrain’s physical features is formed, resulting in a highly flexible controller.

The researchers said:

Our research shows that ANYmal’s controller can tame the incredibly complex terrain of the physical world without requiring precise modeling or a series of dangerous and costly field trials.

It is understood that driven by this controller, the two generations of ANYmal robot dogs can trot in various environments including mud, sand, gravel, dense vegetation, snow, flowing water, etc., and this "test environment" has exceeded the research scope of all previous legged robots.

It looks quite stable indeed!

In the indoor test, ANYmal was walking up a staircase that was 16.8 cm high, which is higher than the clearance for a dog's legs when walking normally on flat terrain.

In fact, ANYmal also participated in the DARPA (Defense Advanced Research Projects Agency) Robot Underground Challenge.

The goal of the competition is to develop robotic systems that can quickly map, navigate, and search complex underground environments, including tunnels, urban subsurfaces, and cave networks. During the competition, human operators are not allowed to assist the robots, and only remote operation is allowed. Therefore, this requires the above controllers to perform without failure for a long period of time.

And the performance of ANYmal and its controller did not disappoint: the controller controlled two ANYmal-B robot dogs in four tasks lasting 60 minutes, and the controller had zero failure rate throughout the competition.

The picture below shows the robot dog traversing a steep staircase during the competition.

At the end of the paper, the researchers admitted that ANYmal only exhibited a trotting gait, which is indeed narrower than the range of gait patterns of quadrupeds in nature.

However, researchers also stated:

The gait pattern is constrained to a certain extent by the robot's kinematics and dynamics. In the future, ANYmal may have multiple gait capabilities, and diverse training programs and goals can stimulate these capabilities.

I look forward to ANYmal becoming even stronger in the future.

Source:

• https://robotics.sciencemag.org/content/5/47/eabc5986

• https://www.youtube.com/watch?v=P6y_dhFTgik