It sounds backwards: why would a robot learn from a flat camera view instead of a richer 3D skeleton? Because the 3D skeleton is often an estimate, while the camera image is what the robot can actually observe. A recent model called UCAG-P makes this trade explicit. It describes actions through camera-visible anchor motion, then translates that shared description into the controls of a particular robot.

The hidden cost of 3D

A monocular camera records colour and brightness on a flat sensor. To recover depth, software must infer it from perspective, body proportions, occlusion and learned priors. The result can be useful, but it is not a direct measurement. An arm moving toward the camera may look shorter; a hand behind an object may disappear; a twist can be ambiguous.

Robotic datasets add another problem: every robot has different joints, coordinate systems and action commands. A policy trained to move one arm cannot simply copy its motor values to a humanoid or a human demonstration.

What UCAG-P changes

The authors of One Policy, Many Embodiments represent manipulation through anchor motion in image and camera-frame coordinates. “Move the gripper toward the cup” becomes a geometric relation visible from the camera, rather than a platform-specific list of joint commands. A geometry-conditioned translator then converts the shared prediction into the target embodiment’s controls.

The paper reports training on thousands of hours of robot, simulation and human-demonstration data, with one checkpoint evaluated across several manipulation benchmarks. Those figures are author-reported preprint results. The important idea is not that 2D contains more information than 3D. It is that a representation can be more transferable when it matches what different embodiments and cameras have in common.

A concrete example

Imagine teaching a robot to pick up a mug. A 3D pipeline might first estimate the mug’s depth, reconstruct the hand and solve inverse kinematics. A camera-centric pipeline can begin with a simpler relation: the hand anchor moves rightward and downward until it aligns with the mug’s visible handle. The robot still needs depth, force and collision checks to execute safely. But the shared policy can learn the visible action geometry before embodiment-specific control fills in the missing variables.

This is similar to giving a dancer a pathway rather than prescribing every joint angle. A pathway leaves room for bodies and contexts to differ. It also leaves something important unresolved: two movements can share a visible path while differing in effort, balance or touch.

What the representation leaves out

Camera-centred action geometry is not a neutral window. It privileges what the camera sees. It can miss hidden contact, weight transfer, internal preparation and depth ambiguities. A system that succeeds on a well-lit benchmark may still fail when clothing, occlusion, camera placement or body proportions change.

The archive’s discussion of 2D motion interfaces and the limits of motion-capture skeletons make the same point from complementary angles: accessibility improves when the input is easier to obtain, but inference should never be presented as capture.

The useful question

Instead of asking whether 2D or 3D is “better,” ask which variables must be shared and which must remain local. Camera-visible geometry may be the right common language for transferring an action across bodies. Force, support and lived quality may need separate sensors or human evaluation. A good embodied system can use a simple shared representation without pretending it is a complete body.

References

Xiaomi Embodied Intelligence Team, Xu, S., Li, F., Zhan, G., et al. (2026). One policy, many embodiments: Unified camera-centric action geometry pre-training for heterogeneous embodied manipulation. arXiv:2608.26058. https://arxiv.org/abs/2608.26058

Yokoyama, K., & Ukita, N. (2026). A plug-and-play 2D motion interface for real-world motion language models. arXiv:2608.15984. https://arxiv.org/abs/2608.15984