← News

Agents & MCPSources: arXiv · 16 Nov 2026

EmbodiExp composes subgoal experience so embodied agents can transfer lessons across tasks

The arXiv paper reports a framework that stores verified subgoal outcomes from past tasks and recombines them for new ones, with reported success gains on two embodied benchmarks.

In plain words

Embodied agents are AI systems that act in simulated or physical environments. The EmbodiExp framework saves records of how past subgoals were completed, including when each record applies and what result must be verified, then chains these records together in new tasks. The authors report higher task success rates than the same agent running without stored experience.

What was announced

The paper argues that tasks with different overall goals often share subgoals, giving a basis for cross-task transfer. Each stored experience describes how to pursue a subgoal, when the guidance applies, and which outcome must be verified for completion. Because a verified outcome from one experience can satisfy the conditions for applying another, experiences drawn from different past tasks can be composed online as a new task unfolds.

Learning works at the subgoal level: the framework keeps experience from successful subgoals and treats explicit action failures as signals about which actions to avoid under specific conditions. In experiments on the EmbodiedBench and VIGIL benchmarks, with stored experience held fixed during held-out evaluation, the authors report success rates of 68.0% on EB-Habitat and 68.3% on EB-ALFRED, against 50.0% and 54.7% for the same agent without stored experience. The figures are the authors' own reported results and cover only the abstract's summary of these two benchmark families.

What it does not show

  • Evidence is limited to the paper's abstract; the full method, experimental setup and baselines are not available for verification.
  • The reported success rates are the authors' own results on two benchmarks and have not been independently confirmed.
  • The abstract does not state whether the approach has been tested on physical robots outside simulation.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.