Boston Dynamics' Atlas parkour — what the video doesn't tell you

The viral Atlas parkour video — the one where Atlas runs up uneven boxes, jumps a gap, and does a backflip — is not what it looks like. It's both more impressive AND less impressive than you think.

Here's the real story.

What you see

A two-legged humanoid robot runs up a series of boxes of varying heights, jumps a gap between two platforms, vaults sideways over a beam, and lands a clean backflip. The whole thing is one continuous take, 41 seconds long, set to dramatic music.

The video has been watched over 50 million times. It is, by a wide margin, the most successful piece of robotics PR ever made.

Why it's MORE impressive than it looks

Most viewers assume the robot is following pre-recorded motions — that someone programmed each move and Atlas plays them back. That's NOT what's happening.

The robot is doing real-time whole-body model-predictive control. At every moment — roughly 500 times per second — it's:

1. Sensing its joint angles, foot pressures, and orientation.
2. Simulating, in its onboard computer, what would happen for the next half-second if it applied various combinations of joint torques.
3. Picking the combination that gets closest to the goal (e.g., "get the right foot onto that box without falling").
4. Executing the first millisecond of that plan.
5. Re-doing the whole thing.

This is the same math used to land rockets. Atlas is solving an optimisation problem in real time, on a moving body, in a changing environment. That's why it can recover from a slip — the math just re-plans from wherever it is.

The course was new to the robot. Not pre-rehearsed.

Why it's LESS impressive than it looks

It took many takes. Boston Dynamics has been candid about this in interviews. Atlas fell, mis-stepped, and aborted runs many times before the perfect take was captured.

A human gymnast would also need many takes. But a human gymnast can also do parkour on a course they've literally never seen before. Atlas — at the time of the parkour video — could only run pre-mapped courses. The boxes and gap had been measured and surveyed by the robot before the run. The route was planned in advance, even if the moment-to-moment execution was real-time.

That's a real limitation. As of 2025, Atlas couldn't walk into an unfamiliar warehouse and improvise. It can now. The parkour video was a milestone, not the destination.

How it was built

The hydraulic Atlas (the one in the video) had 28 hydraulically-actuated joints, driven by an onboard 5-litre/minute pump. Hydraulics are powerful but messy — they leak, they're heavy, they're noisy. That's why Boston Dynamics retired the hydraulic Atlas in April 2024 and replaced it with an all-electric version.

The new electric Atlas can do most of the same moves — it has rotating joints that the hydraulic one couldn't — but isn't as raw-powerful per kilo. The trade-off is acceptable because the electric one is quieter, cleaner, lighter, and manufacturable.

What this video really proves

It proves that dynamic balance — the engineering principle Marc Raibert bet his career on in the 1980s — has won. Static-balance robots (which most early industrial humanoids used) cannot do this. Dynamic-balance robots, with whole-body model-predictive control, can.

That's the unlock. Every humanoid built after this video — Optimus, Figure 03, 1X Neo — has dynamic balance baked in from day one. The parkour video was, in a sense, the moment robotics shifted definitively to a new paradigm.

---

Read about Atlas (the robot) for the full history, including the 2024 transition to electric.