Reward models are reshaping how machines learn to manipulate the world. SARM2 can grade dozens of tasks and turn raw demonstrations into policies that keep improving on their own.
Folding a T-shirt is one of the hardest things a robot can do. A crumpled shirt has no fixed shape, hides its own corners, and rearranges itself every time it is touched. Tasks like these stretch over minutes and dozens of coordinated, two-handed moves, with countless ways to go wrong along the way.
Most robots learn manipulation by imitation. A person tele-operates the robot through a task many times, and the robot is trained to copy what it saw. The trouble is quality. A large dataset gathered by many people varies enormously, and the only cheap proxy for quality is how long a demo took. The mediocre examples get copied alongside the good ones, dragging performance down.
SARM, our first stage-aware reward model, watches the video and estimates progress. It judges which stage the task is in and estimates how far along things are on a scale from zero to one. SARM2 removes the per-task labeling requirement with a vocabulary of just 21 action primitives that cover the vast majority of manipulation tasks, using a mixture of experts to grade roughly ten very different tasks densely and accurately.
A good judge is immediately useful. Reward-Aligned Behavior Cloning weights each snippet of training data by how much real progress it makes. The same dataset, read more carefully, teaches far more. The ambitious use is letting the robot practice on its own. With a judge it can trust, the robot attempts a task, grades its own attempts, and uses reinforcement learning to refine its behavior. We call this loop SPIRAL.
That accuracy is the whole game. A self-improving robot is only ever as good as the judge grading it. The bet behind this work at XDOF is to treat robot data not as a commodity but as something to be understood deeply, with the pipelines and models to extract every bit of learning signal it holds.