Skip to main content

# Learning from demonstration

2 related articles

Subscribe to updates

arXiv Jun 17, 2026
GRID disentangles a general reward from heterogeneous demonstrators, enabling generalist pretraining that outperforms standard learning-from-demonstration baselines
The work introduces General Reward Inference and Disentanglement (GRID), a social learning method that uses an information bottleneck to decompose each agent's reward function into a general reward shared across all agents and specific rewards capturing individual preferences, so that training solely on the general reward yields a generalist agent internalizing universal environmental competencies such as safety and basic task proficiency; experiments on a synthetic basis function decomposition, multi-agent Craftax, MuJoCo Gym continuous control, and the Highway-Env driving simulator confirm semantically meaningful reward disentanglement, outperformance of standard learning-from-demonstration baselines, and more efficient and stable specialization.
The work introduces General Reward Inference and Disentanglement (GRID), a social learning method that uses an information bottleneck to decompose each agent's reward function into a general reward shared across all agents and specific rewards capturing individual preferences, so that training solely on the general reward yields a generalist agent internalizing universal environmental competencies such as safety and basic task proficiency; experiments on a synthetic basis function decomposition, multi-agent Craftax, MuJoCo Gym continuous control, and the Highway-Env driving simulator confirm semantically meaningful reward disentanglement, outperformance of standard learning-from-demonstration baselines, and more efficient and stable specialization.
University of Washington
Caleb Chang（First author）
General Reward Inference and Disentanglement (GRID)Learning from demonstration Information Bottleneck
arXiv Sep 30, 2026
A 100-page, 412-reference survey splits robot in-context learning into four interfaces and argues for evaluating 'did it infer the teaching' separately from 'can it execute after objects change'
Organized around how contextual evidence connects to execution, this survey sorts the robot in-context learning (ICL) literature into four families — context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution — compares their transfer assumptions and the roles of training, correspondence, and memory across manipulation and navigation, and proposes evaluation practices that separate responsiveness to teaching, physical transfer, and benefits from retained experience, together with an agenda linking compositional task acquisition, faithful transfer, and physical recursive self-improvement.
Organized around how contextual evidence connects to execution, this survey sorts the robot in-context learning (ICL) literature into four families — context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution — compares their transfer assumptions and the roles of training, correspondence, and memory across manipulation and navigation, and proposes evaluation practices that separate responsiveness to teaching, physical transfer, and benefits from retained experience, together with an agenda linking compositional task acquisition, faithful transfer, and physical recursive self-improvement.
Haojian Huang（First author）
In-Context Learning Learning from demonstration Robot adaptation survey

You’re all caught up