Researchers led by Yufei Ding have presented UniLatent, a method for reusing robot demonstrations across arms with different shapes, looks and kinematics, in a paper accepted to CoRL 2026.
UniLatent renders matching views of the source and target robots in simulation and aligns the source robot's visual encoder to the target's frozen encoder, so a policy trained on the source demonstrations can run on the target robot without converting images pixel by pixel.
Across three simulated tasks and two target arms it reached up to 93% success without further training, more than 30 points above the best pixel-translation method on average. Moving from an xArm to a Franka arm in the real world, it succeeded in 27 of 30 attempts after fine-tuning on 10 demonstrations per task.
Background
Yufei Ding
Robotics researcher
Yufei Ding is first author of the UniLatent paper on transferring robot skills between robot arms.
UniLatent
Robot learning method
UniLatent is a framework for transferring robot policies between robot arms through a shared latent space.
