MMLab@NTU Proposes NEO-VLA: A Natively Embodied Model That Maps Pixels Directly to Actions
5 hour ago / Read about 0 minute
Author:小编   

A team from Nanyang Technological University (NTU) and collaborators have proposed NEO-VLA, which brings a native vision-language model into robot action learning. Its visual pathway eliminates the need for a separately pretrained visual encoder, and the action module can read representations directly from different depths of the backbone—better exploiting the multi-level visual information required for fine manipulation. The researchers validated performance through multiple controlled experiments, combining simulation with real-robot testing. Trained on a large-scale multi-robot dataset of over 2.23 million trajectories, NEO-VLA delivers strong results across multiple benchmarks, outperforms comparable models on real-robot tasks, and achieves low inference latency on a single RTX 3090. The study further confirms that retaining multimodal understanding capabilities during action training can significantly improve policy robustness.