RESEARCH / BLOG

Research,
Explained.

Read the research question, follow the authors’ methods and explore the evidence behind the article.

Research Blog

Automation · Research insights

Robots learn a predictive body model by watching their own movements

A Nature Machine Intelligence study builds self-models from brief video observations, supporting motion planning and recovery from tested changes.

By NPT Research2 min read

Study published: 25 February 2025 · Nature Machine Intelligence. Original research commentary by NPT Research.

Before moving through a narrow gap, a robot benefits from predicting where its whole body will go. Conventional simulation requires a carefully specified model. Researchers investigated whether a robot could learn a useful representation of itself from images of its own movement.

Latest breakthrough

The study introduced a self-supervised framework that learns morphology and motion from brief raw video data without supplied kinematic priors. The learned simulation supported motion planning and demonstrated detection of abnormalities and recovery from damage in the tested robotic systems.

Read the original study: Teaching robots to build simulations of themselves.

How the study worked

The team recorded robot movements and learned the relationship between motor commands, joint configurations and occupied space. Experiments compared predicted geometry with real movement and used the learned model for planning. Damage-related tests examined whether the representation could help adapt to a changed body.

The idea in everyday terms

A person checking whether a backpack will fit through a doorway uses an internal picture of their current shape. The robot’s learned model serves a narrower computational purpose: predicting its occupied space and movement. Calling this a self-model does not establish consciousness or human-like self-awareness.

Why it matters

Models that can be learned or updated from observation could reduce the effort of bringing unusual robot designs into operation. Camera-based Jacobian learning tackles a closely related control problem through local motion sensitivity rather than the same predictive representation.

What the evidence does not establish

Successful adaptation in the reported tests is not proof of automatic recovery from arbitrary damage. Observability, data quality, model error and physical capability still constrain what the robot can do. A simulator also needs appropriate checking before it is trusted near people or valuable equipment.

Explore the device

Use the robotic-arm model to inspect the volume occupied by each link as the joints move. That whole-body geometry is important beyond the position of the gripper. Trajectory-based slip prevention shows another reason that predicting motion can improve manipulation.

Explore the Industrial robotic arm model and its description, or read our earlier device-focused research explanation.

Primary citation: Yuhang Hu, Jiong Lin, Hod Lipson. Teaching robots to build simulations of themselves. Nature Machine Intelligence (2025). 10.1038/s42256-025-01006-w.

Sources and further reading

  1. Teaching robots to build simulations of themselves Published 25 Feb 2025