SUN Programs and Kuafu system unify symbolic control and machine learning to train robots without human demonstrations

Training robots to perform complex, long-horizon tasks has remained constrained by a technical gap between two tracks: control systems built on physical models that execute precisely defined geometric objectives, and machine-learning-based policies that turn those behaviors into fast reactive responses. This gap has long led to the loss of the original task semantics, forced reliance on complex manual engineering of reward functions, and caused a gradual drift in robot behavior away from the planned control trajectories.

A new study presents the SUN Programs framework and the Kuafu system to address this split by formulating semantically unified executable programs.This approach allows the geometric and contact relationships to be defined only once, then automatically transformed and aggregated into costs compatible with model predictive control (MPC), task-completion constraints, reinforcement-learning rewards, stage-transition barriers, and precise diagnostic tools.

The Kuafu system relies on vision and generative-language modules to automatically compose SUN programs from linguistic commands and surrounding scene semantics, then evaluates their kinematic feasibility through predictive control while preserving the original semantics throughout the conditional-policy training stages. In tests covering nine diverse robotic tasks, the system achieved an overall success rate of 82.03 %, far surpassing traditional baselines such as scattered-reward models at 35.67 % and stage-wise behavioral cloning at 24.75 %.

Regarding scalability and productivity, the system, when run in parallel across up to 8,192 trajectories, generated successful-trajectory time per hour at a rate about 10.57 times higher than remote human operation. Using only 500 trajectories per task, data generated by Kuafu successfully trained DP3 policies to achieve a 46.0 % success rate in simulation and about 34.7 % when transferred and tested directly on physical robot arms of the Franco and Kinova models, without requiring human demonstration videos or dense manual rewards.

This shift touches the core cost-and-operation equation in industrial and logistics automation projects across the Arab region.Traditional reliance on human operators to record thousands of demonstration hours via remote-control platforms has been the biggest financial and temporal barrier to deploying robots in warehouses and advanced assembly factories in the Gulf and Egypt. Shifting to software production lines that combine symbolic planning with automatic data generation enables local engineering teams to train robot arms on production lines at modest hardware cost, turning automation expertise from arduous field data collection into the tuning of semantic task descriptions within simulated environments.