Yes. A humanoid robot can follow a recorded human movement with only a few centimetres of average error and still make that movement smaller, flatter or less energetic. A robotics study submitted on 14 October 2026 measured this directly. It found that the controller variant with the lowest positional error was not the one that best preserved a movement's energy and posture. Preserving expression can mean accepting slightly less positional precision.

Two bows that land in the same place

Picture two people bowing to an audience. One folds forward deeply and slowly, lets the head drop, and pauses before rising. The other makes a quick, shallow dip, as if acknowledging a colleague across a room. If you track only where the head, hands and feet end up at each moment, the two bows may differ by surprisingly little. To anyone watching, they communicate completely different things: gratitude and ceremony on one hand, polite efficiency on the other.

Dancers and movement teachers work with this difference every day. The same path through space can be performed with more or less extension, with more or less effort, close to or far from the body. Movement analysts have developed vocabularies for these qualities, such as Laban's Effort framework, which the Lab has discussed in a practitioner guide to prompting AI with Effort vocabulary.

Robots learning human movement face the same distinction. The question is whether their training and evaluation can see it.

How robots are usually scored

Humanoid robots increasingly learn whole-body movement by tracking: a controller receives a recording of a human motion and tries to make the robot's body follow it while staying balanced. The standard measure of success is positional. Researchers compare the positions of key points on the robot, such as hands, feet and pelvis, with the same points in the reference motion, and average the distances. A smaller number means a closer copy.

That average is useful. It tells you whether the robot is broadly doing the right thing. It does not tell you how the error is distributed. A robot that makes every part of a movement slightly smaller and smoother can keep each point close to its target while shrinking the movement's overall reach and vigour. Small, systematic losses can pass unnoticed in an average that was never designed to look for them.

What the new study measured

In Beyond Kinematic Tracking, Lingfan Bao and colleagues tested three existing whole-body trackers on a Unitree G1 humanoid. They used eight everyday motion types: walking, kicking, punching, crouching, waving, clapping, crossing the arms and bowing. Alongside positional error, they measured five characteristics of each movement:

  • amplitude, how far the hands, ankles and torso travel;
  • energy, a measure of how intensely the joints change angle (the authors stress this is a movement-intensity proxy, not mechanical energy);
  • body expansion, how far the arms extend from the chest;
  • body lean, how far the torso tilts;
  • pelvis height, how high the body is carried above the feet.

The authors' central finding is that lower average positional error does not ensure these characteristics are preserved. In one simulated comparison, a tracker variant reduced average keypoint error from 2.84 cm to 2.18 cm. Yet its error in movement energy rose by about a third, and its error in torso lean also increased. By the standard score it was the most faithful copy. By two of the measures closest to how the movement looks and feels to an observer, it had moved further away.

Correcting for expression, and what it costs

The researchers then trained an additional "residual" policy, a small correction layered on top of each existing tracker. It nudges the robot to restore the five characteristics, concentrating on the body parts that are most active in a given movement and restricting changes elsewhere to limit interference with the original controller.

In simulation, this reduced errors in all five characteristics for all three trackers, while positional error stayed similar: for one tracker it rose slightly, from 2.84 cm to 2.91 cm, as energy and lean errors fell substantially. On the physical robot, across 48 executions per tracker, the corrected versions again reduced characteristic errors. For one tracker, torso-lean error fell from 4.07 to 2.54 in the study's units, and amplitude error fell from 14.55 to 10.53 in the study's normalised units.

The authors are candid about limits. Five descriptors do not capture all expressive qualities. Averages over a whole movement can hide differences in timing. Improving one characteristic can worsen another, and in one case, crouching on one tracker, the energy error grew after correction. This is a step toward measuring expression, not a solution to it.

Why this matters beyond robotics

The study gives engineering form to a point movement practitioners have long made: what you measure decides what survives. The Lab's explainer on how movement should be graded argued that someone has to decide what "good" movement means before an AI can be trained to produce it. A digest in August noted that a low average pose error can conceal what happens through time. This paper shows the same problem inside the control loop of a real robot.

For artists and researchers working with motion capture, retargeting or AI-generated movement, the practical lesson is to check expressive qualities directly rather than assume they come free with accuracy. If a performer's movement is transferred to an avatar or a robot, ask whether its size, intensity, openness, lean and level have survived, not only whether the joints are in roughly the right places.

There is also a limit worth keeping in view. Measuring five kinematic qualities is not the same as preserving intention, relationship or felt experience. A robot that leans exactly as far as a dancer has matched a shape; whether it has matched a meaning is a separate question, and one that the Lab's earlier explainer on whole-body control and body sense keeps open.

The takeaway

Precision and expression are related but not identical. A copy can be accurate on average and still lose what made the original movement distinctive. Measuring expressive qualities explicitly, and accepting small trade-offs in positional accuracy to protect them, is how machines will begin to move with character rather than merely in the right place.

References

  • Bao, L., Zhu, H., Pan, Y., Wang, J., Yu, J., Peng, T., Jia, Y., Kanoulas, D., & Zhou, C. (2026). Beyond kinematic tracking: Preserving expressive characteristics in humanoid whole-body control. arXiv. https://arxiv.org/abs/2610.19393