Boston Dynamics' Atlas does parkour now

I need to talk about the Atlas parkour video.

Boston Dynamics posted it yesterday. Two Atlas robots running through an obstacle course. Jumping over beams. Vaulting up platforms. Running up stairs. And then, at the end, a backflip off a raised beam with a stuck landing.

A robot did a backflip off a beam and stuck the landing.

I’ve watched it fifteen times. Each viewing reveals something new about the smoothness of the movements, the balance corrections happening in real time, the way the robots adjust their gait between obstacles.

The DARPA comparison

Six years ago, the DARPA Robotics Challenge asked humanoid robots to do basic tasks. Open a door. Turn a valve. Walk across rubble. Drive a car.

The results were a comedy reel. Robots face-planting through doorways. Tipping over on flat ground. Falling off ladders. The compilation video of DRC robots falling became one of the most-watched robotics videos on the internet. It was funny because the robots were so bad at things a toddler could do.

Six years later, Atlas does parkour.

The speed of improvement here is the thing that matters most. It’s not linear. In 2015, robots couldn’t walk through a door reliably. In 2018, Atlas could do a standing backflip. In 2021, it’s doing parkour courses with fluid, adaptive movement.

That trajectory, from “can’t open a door” to “backflip off a beam” in six years, is the kind of exponential progress that sneaks up on you. Each incremental improvement seems modest. Then you compare the starting point to the current point and realize how far the line has moved.

What’s actually hard

The parkour isn’t the impressive part. Not really.

The impressive part is the real-time adaptation. When Atlas runs up to a vault, it has to calculate its approach speed, jump timing, hand placement, and landing position in milliseconds. If the surface is slightly different from expected, slightly higher or lower or angled, the robot has to adjust mid-motion.

Human athletes do this subconsciously. Thousands of years of evolution built us a sensorimotor system that handles terrain variation without conscious thought. We don’t think about how to run up stairs. We just do it.

Atlas has to do it with math. Every movement is a solution to an optimization problem. Joint angles, torques, center of mass, ground reaction forces. All computed in real time. And the solutions are getting good enough that the movements look natural.

That’s the uncanny part. Watching Atlas run, it doesn’t look robotic anymore. It looks like a person in a suit. The gait is smooth. The arm swing is natural. The landing absorption after a jump uses the knees the way a human’s would.

The gap between “robot” and “moving like a human” is closing. Not slowly. Fast.

Where this goes

Boston Dynamics isn’t making Atlas for parkour. They’re demonstrating capability. The same control systems that let Atlas vault over a beam will let a future version carry boxes in a warehouse, assist in search and rescue, or operate in environments too dangerous for humans.

The physical capability is arriving. The question I keep asking is: when does the physical capability meet the cognitive capability? When does a robot that can do parkour also understand what it’s looking at, make decisions about what to do, and operate autonomously in unstructured environments?

Atlas still runs pre-programmed courses. It can adapt within the course, but the course itself is designed. The jump from “adaptive within a known environment” to “autonomous in an unknown environment” is enormous.

But six years ago, Atlas couldn’t open a door.

I’ve learned not to bet against exponentials.

Related thinking: