A video of a backflip shows the result of a control problem. It doesn't directly show the muscle forces, joint torques or balance adjustments that made the movement possible. Giving the frames to a simulated character therefore cannot be the same as giving it an animation clip and pressing play.
Berkeley's Skills from Videos work separates the task into pose estimation, motion reconstruction and motion imitation. The final stage learns a controller in a physics simulation. That separation is the part I find most useful: it distinguishes seeing a movement from producing one.
Pose estimation interprets the person in each image. The estimates can be ambiguous, especially when a limb is hidden or the viewpoint makes several poses look similar. Treating every frame independently can also produce jitter. A knee that jumps between two plausible positions may be tolerable as a noisy detection, but it is a poor instruction for a physical body.
Motion reconstruction turns those estimates into a more coherent reference sequence. The reference still serves as a target, not a guarantee that a simulated character with different proportions can follow every position exactly. The imitation stage then learns actions that bring the character's motion toward that target while interacting with the simulated world.
Consider the difference between drawing a foot above the floor and supporting a body on it. An animation system can place the foot wherever a keyframe says. A physical controller has to manage contact, momentum and balance. If it applies the wrong action during landing, the character falls even if the desired pose looks correct.
That gives the simulator a constructive role. It evaluates whether the proposed behavior works under its dynamics. The controller can discover corrections that were not explicitly visible in the source clip because the objective concerns reproducing the skill, not copying every pixel.
There are two kinds of generalization to keep separate. A pose estimator may recognize a new person or camera view. A controller may handle a variation in the physical situation. Success at one doesn't imply success at the other. An accurate pose sequence can still lead to an unstable controller, and a robust controller cannot fully repair a reference that describes the wrong movement.
For a game-development workflow, I would want to inspect the intermediate stages rather than only the final character. The estimated poses reveal perception failures. The reconstructed motion reveals smoothing or alignment problems. The simulated rollout reveals contact and control failures. Each stage suggests a different fix.
The research doesn't turn arbitrary online footage into universally reliable behavior. It does show a practical way to connect visual evidence with physical learning. That is more interesting to me than a perfect-looking imitation by itself: the character has to learn how to make the motion happen in its own body.