SEVO boosts robot manipulation robustness to 85% in new environments

Vision-Language-Action (VLA) and imitation-learning policies trained on low-cost hardware often fail when deployed outside their training environment—a well-known sim-to-real and lab-to-home gap. The SEVO paper demonstrates that this failure is not primarily a model architecture problem but a data and observation design problem. The authors introduce three complementary mechanisms: (1) body-fixed cameras whose combined fields of view cover the full manipulation workspace, (2) active red-spectrum illumination that physically normalizes object appearance across lighting conditions, and (3) a real-time YOLO segmentation overlay that provides a background-invariant semantic cue. Crucially, they show that a diversified data collection protocol—systematically varying lighting, backgrounds, and distractors during teleoperation—is the single most important factor for generalization.

In controlled real-robot trials across two mobile platforms performing a pick-and-place task using transparent water bottles and visually blending objects, SEVO boosted ACT grasp success from 75% to 95% in the training environment and from 30-35% to 85% in novel environments. SmolVLA improved from 70% to 83% in training and from 35% to 75% in transfer. Without SEVO, policies collapsed to 30-35% in new settings. These results are a clear win for data-centric AI over model scaling—showing that low-cost robots can operate reliably in everyday household environments with the right observation pipeline and diversified data collection.