RoboLineage: Agent-Native Data Lifecycle Governance Across Robot Policy Iterations
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "RoboLineage: Agent-Native Data Lifecycle Governance Across Robot Policy Iterations".
Dev: Here's the gist of the paper. RoboLineage is an agent-native data lifecycle governance system designed for robot policy iteration, addressing the scattered nature of evidence connecting repeated data collection, review, retraining, evaluation, and release decisions in modern robot policies.
Rosa: The core concept is to make this lifecycle explicit by treating consequential transitions—such as rollouts, reviews, dataset decisions, training runs, policy metadata evaluations, and deployment recommendations—as typed lineage artifacts.
Dev: Agents interpret embodied rollout evidence to summarize state and adapt accepted data within existing training stacks while strictly maintaining data health boundaries defined by these artifacts.
Rosa: The system formulates robot policy iteration as a data lifecycle governance problem by defining a typed lineage graph that tracks the flow of evidence, artifacts, and decisions across iterations.
Paper summary: Rosa: So, we're seeing how RoboLineage tackles the scattered evidence problem in robot policy iteration. The main idea is to make that entire lifecycle explicit by treating rollouts reviews dataset decisions training runs and deployment recommendations as typed lineage artifacts.>
Dev: They claim this system allows agents to interpret embodied rollout evidence and adapt accepted data within existing training stacks while strictly maintaining data health boundaries defined by those artifacts.>
Taro: It’s about turning the messy, embodied data ecosystem behind robot policy iteration into something that is explicit, auditable, and partially automated.>
Rosa: They introduce these typed artifacts like visual anchors post-rollout reviews dataset decisions and next collection briefs to make the intermediate work of policy iteration inspectable.>
Dev: The system tracks dependencies between things like robot profiles rollouts dataset decisions and policy metadata across different iterations.>
Taro: So, if we think about this for the listener, it means we can see precisely why a specific rollout was admitted or which data update caused a change in the dataset.>
Rosa: This paper focuses on how agent work enters the lifecycle through these artifacts which makes things inspectable and governed.>
Dev: It’s about linking every policy artifact back to the robot task rollouts reviews and data rules that produced it, emphasizing traceability.>
Conclusion: Rosa: So looking at RoboLineage by Luo et al., the core idea is taking the messy reality of robot policy iteration and formalizing it through an agent-native governance structure.>
Dev: It’s about moving from scattered local tools to a system where agents produce typed artifacts that manage the entire lifecycle of policy improvement.>
Taro: The implication for us in the real world is that routine policy iteration could become significantly faster and much more auditable without losing the quality of our resulting policies.>
Rosa: It's about creating a concrete record that answers why a specific policy exists by linking it to all the preceding data collection and review decisions.>
Dev: This system aims to accelerate routine policy iteration while ensuring that there's clear governance over what data is used at every stage of improvement.>
Qian Luo, Wentao Guo, Zhennan Qin, Nanchun Guo, Yunhan Zhao, Yi Ma, Yanchao Yang
Transcengram University of Hong Kong Institute Beijing Institute of Technology
cs.RO
Submitted: 2026-06-20
Updated: 2026-11-07
Comments: Accepted to CoRL2026
Project page: https://robolineage.github.io
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: RoboLineage is an agent-native data lifecycle governance system designed for robot policy iteration, addressing the scattered nature of evidence connecting repeated data collection…
Key concepts
- Typed Lineage Graph
- A structured map that tracks the flow of every piece of evidence, artifact, and decision throughout a robot's policy development process. It explicitly links things like rollouts to dataset decisions and policy metadata evaluations, making the entire lifecycle visible and traceable.
- Agent-Native Governance Architecture
- A system where specialized agents act as workers that produce and manage governance states. These agents interpret task-grounded judgments to create bounded, auditable lifecycle states, allowing for autonomous yet controlled management of data artifacts.
- Artifacts (e.g., Rollout Records)
- Specific, versioned records stored as files with content hashes that capture key events like raw data paths from rollouts or decisions about which dataset to use. These artifacts form the backbone of the lineage graph, providing concrete evidence for policy decisions.
- Policy Iteration as Data Lifecycle Governance
- Framing the process of repeatedly improving a robot's policy as a governance problem. The system applies data lifecycle principles—tracking evidence, managing quality, and routing data—to this iterative cycle to ensure policies are built on sound, traceable foundations.
Terminology
Summary
RoboLineage is an agent-native data lifecycle governance system designed for robot policy iteration, addressing the scattered nature of evidence connecting repeated data collection, review, retraining, evaluation, and release decisions in modern robot policies. p. 0 The core concept is to make this lifecycle explicit by treating consequential transitions—such as rollouts, reviews, dataset decisions, training runs, policy metadata evaluations, and deployment recommendations—as typed lineage artifacts. p. 0 Agents interpret embodied rollout evidence to summarize state and adapt accepted data within existing training stacks while strictly maintaining data health boundaries defined by these artifacts. p. 0
The system formulates robot policy iteration as a data lifecycle governance problem by defining a typed lineage graph that tracks the flow of evidence, artifacts, and decisions across iterations. p. 1 It introduces an agent-native governance architecture where agents function as artifact-producing lifecycle workers, translating task-grounded semantic and integration judgments into bounded, auditable lifecycle states. p. 1 Specific agents include the Robot Onboarding Agent (normalizing robot facts), Task Config Agent (defining operational rules), Online Visual Snapshot Agent (converting sparse events to semantic evidence), Post-Rollout Review Agent (aggregating multimodal evidence into annotations), Data Governance and Data Health Agents (managing rollout routing to training, review, or failure pools), and a Master Agent for cross-iteration synthesis. p. 3
The system is implemented as a lightweight backend with an operator frontend, exposing the lifecycle through various agent views while agents maintain governance state. p. 1 The lineage graph tracks dependencies between nodes like robot profiles, rollouts, dataset decisions, and policy metadata. p. 3 The design emphasizes traceability—linking every policy artifact back to the robot, task, rollouts, reviews, and data rules that produced it—and agent autonomy through explicit state management. p. 3
Key experimental results demonstrate that Full RoboLineage achieves superior automated outcomes and admission accuracy compared to baselines while exhibiting the lowest bad-data leakage. p. 6 Furthermore, it maintains policy quality comparable to expert-managed workflows while reducing routine lifecycle labor by approximately 2. p. 54 times in routine cycles. p. 0 Closed-loop case studies show that lineage-guided recommendations outperform random recollection or no-lineage memory planners in driving targeted collection improvements, as lineage artifacts make the current dominant failure explicit and turn it into a concrete collection target for the next policy iteration. p. 7
The system operates on typed, versioned provenance records stored as ordinary files with content hashes. p. 5 Minimal schemas define key artifacts: rollout records (including raw data paths and hashes), dataset decisions (detailing eligibility, review priority, and routing), and policy schemas (linking to parent policies, training runs, and metadata). p. 5 The Master Agent outputs compact state summaries (master.state.json, master.review.json, etc.) and a typed next collection brief recommending specific actions (collect, review, hold, train) grounded in existing lineage artifacts to suggest targeted improvements like collecting more data for a weak phase or rerunning framework discovery.
The design rationale is built upon five principles: Traceability (linking artifacts to all preceding decisions), Agent Autonomy (allowing agents to interpret and commit only typed artifacts), Portability (using profiles and adapters to absorb robot/framework differences), Operator Efficiency (removing routine review labor while preserving human oversight for ambiguity), and a concrete record that answers why does this policy exist?
The system's goal is to accelerate routine policy iteration while ensuring auditable governance, functioning as an open-source, lightweight lifecycle layer adaptable across different robot embodiments and training families. Future work includes extending capabilities through richer collection modes (e. p. 4g., reference correction), lifelong robot memory to summarize iterations, learning governance decisions (to improve review priors), and establishing community standards for rollout manifests and semantic review artifacts to enhance interoperability across frameworks. p. 1
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Agent-native governance architecture allows agents to operate as
artifact-producing lifecycle workers
that translatetask-grounded semantic and integration judgments into bounded, auditable lifecycle state.
This enables agents to perform substantial semantic work while their outputs remainversioned lifecycle state.
-
The system can reduce routine policy iteration labor by automating the review and integration steps, as the effort study shows a
2.4x reduction
in routine-cycle effort through agent support before training begins. -
Lineage-guided recommendations can drive targeted recollection, as shown in closed-loop case studies where RoboLineage outperforms other planners by shifting the
dominant failure mode rather than simply adding more data.
-
The system preserves policy quality while reducing pre-training labor, as experiments show that RoboLineage remains within
0.4 percentage points of the expert-managed workflow on average
while reducing routine lifecycle labor from 90 to 14 active minutes per cycle. -
The system enables robust multi-robot adaptation by using a
Robot Onboarding Agent
to normalize robot-specific facts into acommon profile,
ensuring downstream modules have thesame contract across robot types.
-
It provides explicit traceability for policy quality, as it ensures that
every policy artifact links back to the robot, task, rollouts, reviews, and data rules that produced it.
Sources
- Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success
- RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation