EgoSteer presents a full-stack approach for steerable dexterous manipulation, spanning egocentric-video pre-training, robot teleoperation, human-in-the-loop correction, and real-robot post-training. Its EgoSmith pipeline reportedly curates 9.6K hours of high-quality data with 9x higher throughput and better accuracy than prior state of the art. The world-model-enhanced VLA is evaluated on more than 40 diverse tasks, including failure recovery and generalization. The paper also reports few-shot adaptation to long-horizon tasks such as box folding on two embodiments, exceeding 75% success. The abstract does not provide detailed protocols or baselines.
No heat snapshots are available in the last 24 hours.