A new artificial intelligence technique allows robots to learn from each other's experiences in real time, without sending sensitive camera footage to a central server. Researchers at Ulsan National Institute of Science and Technology (UNIST) announced the breakthrough on September 3rd, introducing a federated learning technology called FedDepth that could reshape how industrial robots are trained and deployed across factory floors worldwide.

What Problem Does FedDepth Actually Solve?

Monocular depth estimation, which is the ability to judge distance using a single camera, serves as the "eyes" of robots like Boston Dynamics' Spot and Atlas. Traditionally, robots collect visual data from their cameras and send it back to a central location for training, a process that raises privacy concerns and creates bottlenecks in real-world deployments. FedDepth changes this equation by allowing robots working in the field to learn collaboratively without pooling their raw video data in one place.

The innovation addresses a fundamental challenge in federated learning: data heterogeneity. When robots operate in different environments, they capture video with different lighting conditions, angles, and object types. Mixing all this dissimilar data at once can confuse the learning process. FedDepth solves this by using "soft clustering," a technique that groups robots with similar visual characteristics together while allowing them to overlap across multiple groups.

"By overlapping data across multiple groups, it prevents excessively dissimilar data from being mixed all at once, while still leveraging all the shared characteristics among multiple robots for learning," explained Professor Kyungdon Joo's team at UNIST.

Professor Kyungdon Joo, Graduate School of Artificial Intelligence at UNIST

How Much Better Does FedDepth Perform?

The results speak for themselves. When researchers trained three different types of depth estimation AI models using FedDepth, their relative distance estimation error decreased by 17 to 32 percent compared with conventional federated learning approaches. That improvement matters in real-world scenarios: a robot that can judge distances more accurately is less likely to drop objects, collide with obstacles, or fail at precision tasks like assembly work.

The breakthrough arrives at a critical moment for industrial robotics. Companies like Boston Dynamics have been pushing humanoid and quadruped robots into manufacturing environments, but their success depends on reliable perception systems. A robot that can't accurately gauge distance or spatial relationships becomes a liability on a factory floor, no matter how strong its actuators are.

How Robots Can Benefit From Federated Learning

• Privacy Protection: Raw camera footage never leaves individual robots or local facilities, addressing security and intellectual property concerns that manufacturers face when deploying AI systems.
• Faster Adaptation: Robots can learn from the collective experience of other robots in the field without waiting for centralized model retraining, allowing them to adapt to new environments and tasks more quickly.
• Reduced Bandwidth Requirements: Only model updates are shared between robots, not gigabytes of video data, making deployment feasible in factories with limited network infrastructure.
• Improved Accuracy: By leveraging experiences from multiple robots while respecting data differences, FedDepth produces more robust depth estimation models than robots trained in isolation.

The timing of this announcement also connects to broader momentum in physical AI. Earlier this month, a humanoid robot called Moby, built by Noble Machines and powered by NVIDIA's Isaac GR00T framework, demonstrated the ability to handle heavy industrial tasks like lifting steel beams and pipes in semiconductor manufacturing facilities. Moby's success relied on simulation-to-deployment pipelines and vision-language-action models, but robots like Moby still need robust depth perception to operate safely around humans and equipment.

FedDepth represents a different but complementary approach to the same problem: making robots smarter and more reliable without centralizing all their data. As factories increasingly deploy multiple robots working in tandem, the ability for those robots to learn from each other's experiences becomes a competitive advantage. A warehouse with ten robots using FedDepth effectively has the learning capacity of ten robots combined, without the privacy and infrastructure headaches of traditional centralized training.

The research team's work suggests that federated learning, long considered a niche technique for privacy-sensitive applications like healthcare, may become standard practice in industrial robotics. As more companies deploy robots in real-world environments, the pressure to keep data local while still improving performance will only grow. FedDepth shows that it's possible to have both.