RoboLineage: Agent-Native Data Lifecycle Governance Across Robot Policy Iterations

Making policy iteration legible across data, training, evaluation, and recollection.

Qian Luo1,2, Wentao Guo1,3, Zhennan Qin1, Nanchun Guo1, Yunhan Zhao1, Yi Ma1,2, Yanchao Yang1,2

1Transcengram 2The University of Hong Kong 3Beijing Institute of Technology

arXiv PDF Code Videos
Manual policy iteration compared with agent-governed lifecycle artifacts.
The three robot platforms used by RoboLineage: ARX, Realman, and GALBOT G1.

Existing robot policy iteration leaves much of the work between training runs to expert reconstruction: which evidence mattered, why data changed, and what should be collected next. RoboLineage turns those transitions into linked lifecycle artifacts that remain consistent as robots, data formats, and policy learners change. Across diverse real-robot workflows, this lifecycle reduces routine human effort while preserving policy quality.

Abstract

Modern robot policies improve through repeated data collection, review, retraining, evaluation, and release decisions, but the evidence connecting these steps is often scattered across local tools, scripts, and expert memory. RoboLineage makes this lifecycle explicit by representing rollouts, reviews, dataset decisions, training runs, policy metadata, evaluations, deployment recommendations, and next-collection plans as typed lineage artifacts.

Agents interpret embodied rollout evidence, adapt accepted data to existing training stacks, maintain data health, and summarize cross-iteration state under explicit artifact boundaries. In real-robot manipulation workflows, RoboLineage makes routine policy iteration faster and more auditable while maintaining downstream policy performance.

Lifecycle artifacts

Rollouts, reviews, datasets, training records, evaluations, and recollection plans remain linked across iterations.

Agent-native governance

Agents prepare semantic evidence and summaries while validated artifacts carry lifecycle state.

Portable iteration

The interface sits behind different robots and existing policy learners.

Lifecycle Overview

RoboLineage follows the policy iteration loop from robot onboarding and semantic capture to review, dataset admission, training integration, evaluation, and next collection.

RoboLineage lifecycle overview across robot onboarding, rollout review, training, evaluation, deployment, and next collection.

Real-Robot Task Coverage

Selected deployment clips span visual placement, transfer, tool use, sorting, wide-workspace manipulation, and contact-rich articulation.

ARX ACT-style action chunking

Red Cube on Blue

Block placement.

Pour Cherries

Container transfer.

Realman Diffusion Policy workflow

Drawer Open and Close

Contact-rich articulation.

Spoon Stir

Tool use.

GALBOT G1 VLA / LeRobot-style workflow

Sort Balls

Tennis-ball and golf-ball sorting.

Wide Workspace Apple

Far-to-near transfer.

Visual Evidence and Review Artifacts

Raw capture remains the source of truth, online VSA writes sparse semantic anchors during collection, and asynchronous review turns evidence packets into dataset decisions and failure notes. The plates below show how task phase, outcome, uncertainty, and failure evidence are surfaced across representative tasks.

Three-lane design for raw rollout capture, online VSA, and asynchronous post-rollout review.

Pick up the red block and place it on the blue block

VSA result for red block on blue block.

Stack the two paper cups together

VSA result for stacking two paper cups.

Swap the positions of the watermelon and the mango

VSA result for swapping watermelon and mango.

Pull the tissue out and place it on the table

VSA result for tissue extraction.

Pour the cherries from one bowl into another bowl

VSA result for cherry transfer.

Right the fallen bottle

VSA result for righting a bottle.

Lineage-Guided Recollection

Failure analysis becomes data-collection action when review evidence is connected to dataset state, evaluation outcomes, and prior policy iterations. In the stacking case, repeated failures become targeted requests for out-of-distribution grasp locations, centered pre-release placement, and bounded-dwell release examples.

Closed-loop stack red-on-blue cube case study with failure evidence and recollection recommendations.

Paper and Citation

Read the preprint on arXiv:2606.22142. 

@article{luo2026robolineage,
  title   = {RoboLineage: Agent-Native Data Lifecycle Governance Across Robot Policy Iterations},
  author  = {Luo, Qian and Guo, Wentao and Qin, Zhennan and Guo, Nanchun and Zhao, Yunhan and Ma, Yi and Yang, Yanchao},
  journal = {arXiv preprint arXiv:2606.22142},
  year    = {2026}
}