Scientists show predictable training can outperform complex robot learning data

Researchers at New York University Tandon School of Engineering and the Robotics and AI Institute have demonstrated that robots trained with structured, predictable demonstrations achieve better dexterous manipulation than those trained on highly variable data. The team focused on imitation learning, where robots learn by copying demonstrations, but found that traditional motion-planning algorithms like rapidly exploring random trees (RRTs) produce inconsistent solutions, complicating learning. To address this, the researchers developed alternative planning methods that generate more consistent training examples. One approach emphasized steady progress toward goals, while another used predefined motion libraries to reduce variation. These methods reduced high-entropy data, which had previously hindered learning. The team tested their approach on two complex tasks: rotating a cylinder 180 degrees with two robotic arms while adjusting grips, and manipulating a cube inside a robotic hand. Robots trained on consistent data achieved near-perfect success in the dual-arm task with just 100 demonstrations and maintained high performance—90% success in real-world trials—without retraining. The dexterous hand completed 62% of its tasks. The findings suggest that structured training data can outperform large volumes of inconsistent examples, challenging the assumption that more data always improves learning. The study also bridges traditional motion planning and machine learning by using planning algorithms to generate training data for learning systems. The research reinforces a broader trend in robotics, where combining structured approaches with AI enhances performance in complex tasks. By prioritizing consistency over randomness, the team achieved significant improvements in real-world robotic manipulation.