Exploring the World of Robotics

Building Robots with Boston Dynamics CEO Robert Playter

Boston Dynamics has been creating robots for over 30 years, including humanoid robot Atlas and robot dog Spot.
The team aimed to create a natural-looking gait for their robots, which took over 10 years to achieve.
Robert Playter is the CEO of Boston Dynamics and has been with the company from the beginning, including its roots at MIT.
Playter received his PhD in aeronautical engineering from MIT in 1994, where he wrote his thesis on robot gymnastics.
Playter was inspired to pursue robotics after seeing Marc Raibert's robot perform a somersault.
Raibert's focus on simplifying a problem to its core essence and solving the fundamental problems is a value that Boston Dynamics still holds.
Boston Dynamics' researchers work on developing software and hardware at a principled level to control robots in real-time.
The company values breaking things, repairing them, and testing them to make progress.
Playter believes that identifying with a robot's physical interaction and pursuing something you love are sources of happiness for roboticists.
Legged locomotion is a multi-decade problem that Boston Dynamics has been working on solving.
The company's approach to controlling robots is changing and becoming more broadly available.
Playter believes that understanding the dynamics of motion and feedback-control principles are essential in building robots.Challenges of Developing Natural-Looking Humanoid Robots
The humanoid form allows for intuition about how the robot should look while moving, which is an art that requires tapping into a deep knowledge of the body.
A coach's ability to see and articulate small changes in an athlete's movement is important in achieving elegant movement.
Getting the physics of movement right in a robot results in more efficient and stable movement that is perceived as lifelike, elegant, and natural-looking.
Developing a natural-looking gait in humanoid robots is challenging and took around 10-15 years to achieve with newer techniques.
Challenges in developing natural-looking walking include dealing with singular configurations, underactuation, and the complexity of the humanoid form.
Picking up and throwing heavy objects with the humanoid robot requires modeling the weight distribution, predicting how the robot will manage its body plus the object, and maintaining balance.
Modeling new objects that the robot might need to pick up is possible with the development of new tools that allow for quicker creation of new behaviors.
Generalizing the shape of objects to focus on the center of mass is important in picking up and throwing objects.
Atlas, the humanoid robot, can jump onto a box that is around a meter high.
Developing humanoid robots that can carry other robots is a challenge that has not been achieved yet.
The ability to model the full physics of the robot and the objects it interacts with is important in achieving natural-looking movement.
The complexity of the humanoid form, with its heavy upper body and legs, makes maintaining balance and dealing with inertia a challenge.Building Robots That Can Jump and Backflip: Lessons Learned from the DARPA Robotics Challenge
Boston Dynamics is currently working on the next generation of Atlas, which will have stronger actuators to perform even more challenging tasks.
The development of jumping and backflipping robots has evolved rapidly in recent years, thanks to model-predictive control techniques that allow the robot to project ahead and explore options to get from A to B.
Early experiments required a lot of manual iteration to change the trajectory of the arms or legs to achieve a successful backflip, but more recent techniques allow for much more natural and optimal trajectories.
The robot can think ahead for the next second or two about how its motion will transpire, and solve for optimal trajectories to get from A to B in real-time.
Calculating physics using Newton's laws and model-predictive control techniques allow the robot to adjust on the fly in the air and stabilize itself during tricks.
The DARPA Robotics Challenge in 2015 tasked humanoid robots to do various tasks such as driving a car, opening a door, identifying and shutting valves, using a tool to cut a hole in a surface, crawling over stairs, and rough terrain.
Boston Dynamics provided the Atlas robot and baseline software to various teams that competed to see who could get through the maze the fastest.
The challenge was humbling as the tasks were not all that hard for a person, but it was hard to get the robots to do it due to their general-purpose nature and the need to deal with uncertainty.
Robots need to be able to deal with unexpected situations to be useful in the future, which is the goal of a general-purpose robot.
Building successful robots for specific tasks has been the wisdom up until now, but the goal is to build general-purpose robots that can be used for a wide variety of activities.
The wisdom in building successful robots up until this point has been to "go build a robot for a specific task, and it'll do it very well," but this approach won't work for the future.
By breaking the robot repeatedly, Boston Dynamics finds the weak points, redesigns it to not break so easily next time, and learns lessons to improve not just how to make the backflip work but also how to build the machine better.Challenges and Advancements in Robotics Engineering: A Conversation with Boston Dynamics' Founder
Boston Dynamics' robots require fast feedback loops and updates for accurate predictions in real-world scenarios.
Multiple layers of calculations are necessary to run simultaneously, ranging from 1,000 hertz to 100 hertz, depending on the level of detail and complexity of the robot's movements.
The software engineering pipeline requires continuous development, simulation tools, and the ability to run the same code on both simulation and hardware.
Physics-based simulation tools are used to test the robots in simulation before putting them on the actual robot, and the same code is used for both.
The challenge of simulating contact events, such as foot-ground contact or manipulation, requires capturing the important parts of the interaction while remaining computationally feasible and efficient.
Boston Dynamics has about 40 developers working on their four Atlas robots, meaning that sharing resources and using software pipelines is essential.
In the early days of developing BigDog, loose rock and foot placement were the epitome of a hard walking surface, leading Boston Dynamics to build rock piles and boxes of rocks for testing.
Boston Dynamics does not put strong boundaries around anthropomorphizing their robots and believes it can enhance interaction between humans and robots.
The company is beginning to pay attention to physical indications and communication from the robots, such as when they need help or are about to turn, to make them more useful and understandable for their customers.
Boston Dynamics has about 1,000 robots out with customers currently, and this layer of physical indication is becoming increasingly important.
Robots require a level of intelligence that they do not yet possess, and people often attribute more intelligence to them than is appropriate.
While the robots are not yet capable of sipping beer on the beach, they are capable of walking on soft surfaces and navigating complex facilities, taking measurements and communicating their actions with people.