This generation of Atlas does not complete the entire route through pre-programming as before, but relies on its own algorithm to perceive the environment and act autonomously.
Through this algorithm, Atlas can plan out the route that he will walk in real time, and at the same time make adjustments in advance according to the distance of obstacles to be encountered.
◎Our reporter Zhai Dongdong
Recently, Boston Dynamics released the latest sports video of its biped robot Atlas. Its parkour skills are impressive. Many netizens said that the robot has become "excellent".
In the video, Atlas shows difficult movements such as jumping, balance beam, and horse vaulting, and can smoothly pass various obstacles such as slopes and single-plank bridges to complete complex parkour sports.
The Internet celebrity robot Atlas is not the first time that it has entered the public eye.
In 2016, Boston Dynamics released a walking video of Atlas, showing people a robot that truly has the ability to move in multiple scenes indoors and outdoors.
In the past six years, the Internet celebrity robot has continuously upgraded its balance control and route planning, showing the potential of humanoid robot technology time and time again.
From a toddler to being able to parkour
Although for the current Atlas, running, jumping, backflip and other difficult movements are not a problem, but in 2016, it is still in the toddler stage.
In the video released by Boston Dynamics in 2016, it is not difficult to find that although the appearance of the Atlas at that time was almost the same as the current parkour version, it still walked cautiously and staggered, and the action was not very consistent. , Some netizens said that it "looks like a drunk man going home late at night".
Even so, the 2016 version of Atlas is still an important breakthrough for humanoid robots.
Researchers from Boston Dynamics said: "The main energy of our robotics project is focused on how to make the biped robot traverse a variety of complex terrain stably, and its biggest breakthrough is that the robot does not need to use integrated sensors to model the terrain. An algorithm Everything can be done, just like an autonomous vehicle has got rid of lidar and various sensors."
One year later, Boston Dynamics released another Atlas related video. This time Atlas' movements became smoother, learning to jump between steps, and even a backflip. In 2018, Atlas learned to run and was able to run while running. Cross the logs and jump up the steps; by 2019, Atlas has mastered the "gymnastics" skills, able to jump, roll, and stand upside down continuously; at the end of 2020, Boston Dynamics released a video of Atlas dancing.
In this video, two Atlas robots dance to the music, and their movements are so smooth and natural that many netizens consider it to be CG special effects; in the recently released video, Atlas has achieved a new parkour achievement, and its dexterity is getting more and more. Close to humans.
Environmental awareness continues to improve
According to reports, this generation of Atlas does not complete the entire route through pre-programming as before, but relies on its own algorithm to perceive the environment and act autonomously.
Through this algorithm, Atlas can plan out the route that he will walk in real time, and at the same time make adjustments in advance according to the distance of obstacles to be encountered.
In order to carry out parkour course training, Atlas will be imported into a high-level map, this high-level map is a brief description of the location of obstacles and some main actions.
Atlas will use this brief information to navigate, while using real-time sensory data to respond to real situations.
For example, Atlas knows that a box will appear on the map, and its action is to jump on it.
This box may be far from the label on the map, or even moved 0.5 meters to the side, but Atlas can still rely on the perception algorithm to correctly find this box and jump on it.
Scott Kundesma, head of the Atlas team at Boston Dynamics, wrote in a blog post: “Atlas’ actions are now driven by perception, and they weren’t before. For example, in the previous video, the robot’s control The system still needs to make a lot of key adjustments in the execution of the task to keep its balance and posture goals, and the robot does not perceive and react to its environment. "In this parkour, Atlas is based on what it sees. To adjust its movements.
This means that engineers do not need to pre-program jumping movements for all platforms and gaps that the robot may encounter.
Atlas' judgment of the environment relies on algorithms, but to complete parkour, it needs to design in advance to make jumps, running, jumping, backflips and other actions, and put these actions into the template library.
In the design, these actions will be adjusted according to the trajectory optimization technology, and then selected from the template according to the specific action goals.
In action, Atlas’ “brain”, MPC (Model Predictive Controller), will calculate the best action needed at the moment, and predict the best way of action based on the passage of time.
Although Atlas’ excellent performance shows its strong athletic ability, a new blog recently published by Boston Dynamics-"Built it Break it Fix it" (Built it Break it Fix it) shows a lot of Atlas. Picture of wrestling.
In this blog, the robot no longer "flies over the wall", but shows all kinds of wrestling, rollover, and loss of balance. The engineers are like Atlas's nanny, using hydraulic machines to lift the fallen robot and using pliers to reset the body of Atlas. Bolt, repeatedly debug the hardware and software problems behind a small failed action.
According to Boston Dynamics, Atlas wrestles much more frequently than in the video. Only by continuously providing detailed repairs can it be possible to use these "artificial" to build up the intelligence that humans can see.