Part 1-2

Skill-Policy Compatibility

LoadingHyperFrames composition

Why compatibility is the problem

What breaks when the skill library updates?

An agent learns from at least two streams: user-specific tasks and general behavior. Behavior is too data-hungry for users to train, so each user adapts the deployed checkpoint of that moment and stays pinned to it.

01Users adapt a frozen pretrained checkpoint
02An update shifts the representation and breaks them
03Freezing or retraining both fail at deployment scale

Key messageThe open question is how skills can keep updating without every adapted user retraining.

[9]Daehee Lee, Dongsu Lee, TaeYoon Kwack, Wonje Choi, and Honguk Woo. "Policy Compatible Skill Incremental Learning via Lazy Learning Interface." NeurIPS, 2025.

24

Skill Memory for Lifelong Embodied Agents · Submission · Part 1 · Part 1-2.2 · The ProblemSkill-Policy Compatibility

Current 2424 / 38

1TitleSkill Memory for Lifelong Embodied Agents

2Part 1Lifelong Embodied AgentsLifelong Embodied Agents

Intro

3Research IntroductionAbout Me
41.1Lifelong AgentsLifelong Embodied Agents
51.2Skill DefinitionLifelong Embodied Agents
61.3Long HorizonLifelong Embodied Agents
71.4Definition of CRLLifelong Embodied Agents
81.5After DeploymentLifelong Embodied Agents

Part 1-1

9LocalityLocality
101-1.1IsCiLLocality
111-1.2IsCiL IILocality
121-1.3IsCiL ProblemLocality
131-1.4MethodSkill Memory
141-1.5IsCiL Fig. ISkill Memory
151-1.6IsCiL Fig. IIISkill Memory
161-1.7CompleteTask-wise Selective Adaptation
171-1.8Semi-completeTask-wise Selective Adaptation
181-1.9IncompleteTask-wise Selective Adaptation
191-1.10Training OverviewTask-wise Selective Adaptation
201-1.11UnlearningTask-wise Selective Adaptation
211-1.12Summary MessageTask-wise Selective Adaptation

Part 1-2

22CompatibilityCompatibility
231-2.1SIL-CSkill-Policy Compatibility
241-2.2The ProblemSkill-Policy Compatibility
251-2.3SIL-C Fig. ISkill-Policy Compatibility
261-2.4MethodBilateral Lazy Learning Interface
271-2.5The MechanismBilateral Lazy Learning Interface
281-2.6SIL-C Fig. IIIBilateral Lazy Learning Interface
291-2.7SIL-C Fig. IIForward & Backward Skill Compatibility
301-2.8Emergent SILForward & Backward Skill Compatibility
311-2.9Explicit SILForward & Backward Skill Compatibility
321-2.10Across SettingsForward & Backward Skill Compatibility
331-2.11Interface AblationForward & Backward Skill Compatibility
341-2.12Summary MessageForward & Backward Skill Compatibility

35Part 2DiscussionPart 2 · Discussion

Section 1

36Real-Time AdaptationReal-Time Adaptation for Real-World Robots

Section 2

37Q&AQuestions & Answers

Section 3

38ReferencesReferences