In a giant leap for multipurpose robotics, MIT engineers have whipped up a new training method that boosts a robot's ability to handle a diverse array of tasks – from swinging hammers to turning wrenches. This breakthrough, courtesy of a technique called Policy Composition (PoCo), taps into an advanced form of artificial intelligence known as diffusion models to merge data from numerous sources. Unlike other methods that depend solely on narrow datasets, PoCo promises robots a well-rounded education, leading to a 20 percent spike in performance.
Traditionally, robots trained on single-dataset diet struggle when confronting new tasks or unfamiliar environments. These limited data sets are like pieces of a puzzle that never come together. But with PoCo, robots embrace a more holistic view, a mashup of learning strategies derived from various tasks and domains – whether that's mimicking human repairs or swiping moves from simulated environments. “Addressing heterogeneity in robotic datasets is like a chicken-egg problem. If we want to use a lot of data to train general robot policies, then we first need deployable robots to get all this data,” Lirui Wang, an MIT electrical engineering and computer science graduate student and the paper's lead author, explained in a statement obtained by MIT News.
MIT's approach has the potential to crack the code on robotics modularity — each new dataset or domain can be slathered on like an additional layer of learning, without the need to reboot the entire training process. By assembling policies that excel in various skills, the resulting robot can adapt on the fly, showing off its dexterity and problem-solving in real-world tests.
A peculiar side effect of this new process is the individual strength of each training session. By combining the policies, robots output maneuvers that are sharper and more precise than any single policy could produce. “The striking thing was that when we finished tuning and visualized it, we can clearly see that the composed trajectory looks much better than either one of them individually,” Wang told MIT News.
With big names like Amazon and the National Science Foundation funding the project, it's clear there's real interest in getting these jack-of-all-trades robots off the ground. NVIDIA's AI Agents Initiative lead Jim Fan, not linked with the MIT project, summed up the potential impact: “We will need all three kinds of data to succeed for robotics: internet data, simulation data, and real robot data. How to combine them effectively will be the million-dollar question. PoCo is a solid step on the right track,” according to MIT News.
