Part 1-2

Bilateral Lazy Learning Interface

LoadingHyperFrames composition

How the proxy runs

Trajectory matching with goal and initial state

Task state-goal pairs form one space and skill start and goal distributions form another. Matching projects between the two.

01Task pairs are clustered by state while retaining their associated goals
02The retrieved goal is projected into the skill state space
03A skill is accepted when the target lies within its goal distribution and rejected otherwise
04Candidate skills are ranked by initial state distance

Key messageA lazy learning interface resolves many-to-many start–goal relations at runtime.

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

27

Skill Memory for Lifelong Embodied Agents · Submission · Part 1 · Part 1-2.5 · The MechanismBilateral Lazy Learning Interface

Current 2727 / 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