Real-world robot manipulation learning from simulation control and real images
ReDOA: Real-World Distillation via Observation Alignment for Image-Based Manipulation Policy Learning
Robotics
Summary
Teaching robots to do tasks using images is hard because collecting real examples is expensive and simulated training often doesn't work perfectly on real robots. The authors propose ReDOA, which uses a smart way to transfer what a robot learns from simulation using detailed state information into a simpler, camera-based robot controller in the real world. This avoids needing costly human demonstrations and helps the robot learn tasks more efficiently and effectively. Their experiments showed better success and faster task completion than traditional methods on real robots.
What this means in practice
- •For robotics engineers: Train robots to perform manipulation tasks more efficiently with fewer real-world demonstrations by transferring control knowledge from simulation to real robot vision.
- •For automation system integrators: Develop industrial robotic systems that can quickly adapt image-based manipulation policies without costly human demonstrations by using real robot data aligned with simulation control.$Commercial implications: Enables development of adaptable robotic manipulation products that reduce demonstration costs and improve deployment speed in manufacturing and logistics.
Authors
Kotaro Kajitsuka, Masaki Murooka, Fumio Kanehiro
Abstract
Data-driven learning for general-purpose robotic manipulation presents a fundamental trade-off between data-collection cost and adaptability to real-world environments. Imitation learning (IL) requires costly human demonstrations, whereas simulation-based learning is vulnerable to sim-to-real gaps in dynamics and sensor observations. We propose ReDOA (Real-World Distillation via Observation Alignment), a framework that decouples control learning from perception learning to learn image-based policies in the real world. A state-based teacher policy trained in simulation is executed on a real robot, and its control knowledge is distilled into an image-based student policy using the collected trajectories. Observation Alignment resolves observation mismatches between data collection and deployment. This framework collects successful, efficiently executed trajectories paired with real-world visual observations, without requiring demonstrations from human experts. In real-robot experiments on three non-prehensile manipulation tasks, including bimanual manipulation, our method achieves better overall performance when jointly considering task success and execution efficiency than both IL based on human demonstrations and policy distillation in simulation. https://isri-aist.github.io/redoa-project-page/