ReDOA: Real-World Distillation via Observation Alignment for Image-Based Manipulation Policy Learning

Kotaro Kajitsuka1,2, Masaki Murooka2, Fumio Kanehiro1,2

1University of Tsukuba   2AIST

Conceptual overview of the ReDOA framework

Learning image-based manipulation policies from real-world experience, without expert demonstrations.

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 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 uses image compositing to remove visual artifacts caused by temporary object trackers, reducing the observation mismatch between data collection and tracker-free deployment. In real-robot experiments on three non-prehensile manipulation tasks, our method achieves a strong balance between task success and execution efficiency.

Project Video

An overview of ReDOA and its real-robot experiments.

Method

ReDOA learning pipeline

Observation Alignment

Trackers make it possible to run the state-based teacher on the real robot, but they appear in the collected images. ReDOA removes these marker regions before training the image-based student.

Observation Alignment Pipeline

Detect the temporary marker region, combine it with the object mask, and inpaint it to create an aligned observation.

Image compositing removes temporary object markers from real-robot observations

Real-World Rollout Evaluation

Representative policy rollouts on the three manipulation tasks. The clips are grouped by task and identify the policy or demonstration source; result overlays indicate success or failure.

Comparison of ReDOA and baseline learning pipelines

ReDOA is compared with policy distillation in simulation and imitation learning from human demonstrations. Evaluation considers both success rate and execution efficiency.

Box rotation task

Box Rotation

Rotate the box to the target orientation.

ReDOA · Success

Human Demonstration · Success

Simulation Distillation · Failure

Drawer opening task

Drawer Opening

Open a drawer from varied initial positions.

ReDOA · Success

Human Demonstration · Success

Cylinder rolling task

Cylinder Rolling

Roll the tapered cylinder to the target orientation.

ReDOA · Success

Human Demonstration · Success

BibTeX

Preprint forthcoming.

@misc{kajitsuka2026redoa,
  title  = {ReDOA: Real-World Distillation via Observation Alignment for Image-Based Manipulation Policy Learning},
  author = {Kajitsuka, Kotaro and Murooka, Masaki and Kanehiro, Fumio},
  year   = {2026},
  note   = {Preprint forthcoming}
}