#computer vision

Published in Science Robotics, the system uses a camera mounted on a robot’s hand to watch its own movements. From this visual data, a neural network learns a self-model that can be used to simulate physical interactions, navigate space, and plan movements more effectively.

This approach bypasses the need for predefined physics models or manual calibration. The result is a more adaptive, resilient robot that can function in dynamic or unfamiliar environments—potentially accelerating development in autonomous systems and soft robotics.

The team says this method could lead to robots that are easier to train and more capable of adjusting to physical wear, damage, or manufacturing variability.

Image: Qualitative results demonstrating robustness in out-of-distribution scenarios. Credit: Nature (2025). DOI: 10.1038/s41586-025-09170-0.

Header image: A 3D-printed robotic arm grips a pencil while training through random motions and a single camera, using a system called Neural Jacobian Fields (NJF). Instead of depending on sensors or manually coded models, NJF enables robots to learn how their bodies respond to motor commands through visual observation alone—paving the way for more adaptable, cost-effective, and self-aware machines. Credit: Image courtesy of the researchers.

Scientific paper: Sizhe Lester Li et al, Controlling diverse robots by inferring Jacobian fields with deep networks, Nature (2025). DOI: 10.1038/s41586-025-09170-0

Read more at MIT News

Related Content

Artificial intelligence enhances inspection in poultry processing industry

Other Recent News

Firefly Aerospace eyes $5.5B IPO as Blue Ghost lander preps for annual moon missions.

Apple’s iOS 18.5 rolls out with satellite messaging, new wallpapers, and gesture-based alerts.

Dubai’s DIFC hits 1,388 innovation firms, bolstering its global fintech leadership.

NASA spots double supernova, giving astronomers a rare glimpse into cosmic death.

Scientists revive the 1918 flu virus to study its pandemic potential.