INSAIT announces a landmark achievement: for the first time in history, a Bulgarian institution has a paper accepted at the Conference on Robot Learning (CoRL), a top research conference in robotics. The paper, titled “Generalist Robot Manipulation Beyond Action Labeled Data”, will be presented at CoRL 2025, taking place September 27–30 in Seoul, South Korea.
The work introduces MotoVLA, a novel approach that enables robots to acquire manipulation skills from unlabeled human and robot videos, drastically reducing the need for massive
action-labeled datasets. By predicting 3D dynamics directly from video and aligning them to actions with only a small labeled set, MotoVLA unlocks out-of-action generalization—allowing robots to perform completely new tasks without ever seeing their action labels during training.
In rigorous testing, MotoVLA delivered state-of-the-art results, outperforming prior baselines by up to +14.1% on key benchmarks, both in simulation and on real robotic platforms. This breakthrough underscores the potential of combining 3D vision, weak supervision, and robotics to push forward the development of truly generalist robotic systems.
The research team includes INSAIT researchers Alexander Spiridonov, Dr. Jan-Nico Zaech, Nikolay Nikolov, Prof. Luc Van Gool, and Dr. Danda Paudel.