Evolving Skill Modules under a Fixed Planner: Versioning, Rollback, and Runtime Governance for Long-Lived Robot Systems
This paper presents a software-lifecycle framework for long-lived robot systems that decouples a fixed decision layer from versioned embodied capability modules and a runtime governance layer. Evaluating the system on six robosuite tasks, the authors report three negative results—including instability in peak task success across random seeds, performance limitations of time-sliced multi-policy rotation, and paralysis caused by an overly aggressive action-level safety shield—alongside measurements of promotion gates and rollback machinery.
Summary and content fields are drawn from the document text and are pending editorial review.
Who issued it
Who it governs
Robotics researchers and engineers looking to understand the lifecycle management, versioning, and runtime governance challenges of evolving learned robot skill modules.
What the document contains
The document addresses the software lifecycle, versioning, rollback, runtime governance, and evaluation of modular capability updates for long-lived robotic systems using simulated manipulation tasks.
The text does not provide formal interface compatibility proofs, operational deployment on physical hardware, or language-model code synthesis implementations.
- Peak task success varies significantly across random seeds within a single method, indicating single-run peaks cannot reliably rank methods.
- A fixed-schedule rotation of whole-task policies loses reached goal states due to unconditional handoffs, but retention improves significantly when completion states are held.
- An action-level safety shield can achieve high violation reduction by discarding actions entirely, reducing task success to zero if acceptance criteria omit task completion.
- Version gating and rollback machinery can effectively screen out injected capability regressions, though with measured clean-candidate rejection costs.
- Maintaining a fixed decision layer provides structural stability and clear provenance without sacrificing task performance compared to mutable dispatchers.
1. Introduction 2. Related Work 3. Method 4. Experiments 5. Discussion 6. Conclusion
How it scores
Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.
Accountability
4/5Clear responsibility for AI outcomes
holding the decision layer fixed is a standard design choice in hierarchical control rather than a new one, and we make no novelty claim for it; its role here is to make attribution exact
1. Introduction
screening score · 2026-10-23
Fairness
0/5Non-discrimination and equitable treatment
screening score · 2026-10-23
Human Dignity
0/5Respect for the inherent worth of persons
screening score · 2026-10-23
Human Oversight
2/5Humans remain in control of consequential decisions
Robot learning on the job places deployment-time retraining behind human oversight that decides how much autonomy each policy version receives
1. Introduction
screening score · 2026-10-23
Privacy
0/5Protection of personal data and private life
screening score · 2026-10-23
Enforcement
4/5Consequences and compliance mechanisms with real force
a gate that decides whether a new version may be promoted, a way back when it should not have been, and an enforcement boundary
1. Introduction
screening score · 2026-10-23
Impact Assessment
4/5Required assessment of an AI system's risks/impacts before or during use
a promotion gate that admits a candidate only if it holds up on held-out evaluation, and a rollback path when a promoted version regresses.
1. Introduction
screening score · 2026-10-23