New AI Model Teaches Robots to Reason Through Task Progress Like Human Supervisors

Understanding whether a robot is making real progress toward completing a task remains difficult for AI systems trained in traditional ways. A research team has developed a novel approach that fundamentally changes how video-based AI models evaluate robotic manipulation by shifting them from merely recognizing what's happening to actively assessing whether actions move closer to the goal.

The Challenge with Current Approaches#

Video-based machine learning models trained using standard supervised fine-tuning behave as passive observers—they identify events unfolding in a video but lack the reasoning needed to evaluate whether those events constitute meaningful progress. This limitation becomes particularly acute for long-horizon robotic tasks involving multiple steps over extended sequences, where distinguishing productive motion from wasted effort requires contextual understanding tied to the end objective.

PRIMO R1: A New Framework#

The researchers introduced PRIMO R1 (Process Reasoning Induced Monitoring), a 7-billion-parameter model that fundamentally alters this dynamic. By applying reinforcement learning techniques, the framework incentivizes the model to generate detailed step-by-step reasoning about progress estimation—an ability absent in its predecessors. The architecture also structures video inputs in a novel way: it explicitly anchors each sequence between initial and current state images, providing crucial temporal context for comparison and judgment.

Performance and Benchmarks#

Testing revealed substantial improvements across multiple evaluation scenarios. The 7B PRIMO R1 model reduced mean absolute error by 50% compared to smaller reasoning-focused baselines, while significantly outperforming larger general-purpose models at the 72B-parameter scale. On the RoboFail benchmark—designed to detect when tasks go wrong—PRIMO R1 achieved 67.0% accuracy, exceeding closed-source models like OpenAI's o1 by 6 percentage points.

Real-World Applicability#

The framework demonstrates robust zero-shot performance on challenging failure detection tasks, suggesting it can generalize beyond scenarios it encountered during training. Testing spanned controlled laboratory environments alongside real-world humanoid robot scenarios, indicating the approach moves beyond narrow academic validation toward practical deployment potential.