Extract
Turn successful and failed trajectories into reusable procedural knowledge.
Continual skill learning for embodied agents
Learn not only new skills, but how to evolve them.
PRACTICE trains a dedicated skill learner to grow and refine a persistent skill library from interaction trajectories—while keeping the embodied task executor frozen.
Framework overview01 / 05
49.7%EB-ALFRED success
58.3%EB-Habitat success
+9.7points over prior experience methods
2benchmarks, one reusable learner
The central idea
The library evolves.
The update policy learns.
Existing experience-augmented agents often rely on fixed prompting workflows to create or edit skills. PRACTICE replaces that hand-crafted update procedure with a trained model that reasons over the current library and fresh evidence together.
Consolidate
Merge batch-level edits into a compact, persistent, and internally consistent library.
Revise
Add, refine, merge, or remove skills as new interaction evidence arrives.
Training curriculum
Three stages.
One co-evolving system.
Oracle-grounded supervision establishes the editing language; diverse failures broaden coverage; online skill-edit distillation aligns the learner on its own edit distribution.
- Stage 0Oracle-Grounded SFTCold-start skill generation and maintenance.
- Stage 1Failure-Aware SFTLearn invalid conditions and recovery strategies.
- Stage 2Skill OPDMatch teacher distributions on student-generated prefixes.
Main results
Stronger skills.
Same frozen executor.
Under the same Qwen3-VL-32B executor, PRACTICE delivers the strongest average performance among the reported experience-augmented approaches on both benchmarks.
EB-ALFRED49.7%
+9.7 points over the strongest prior experience-augmented baseline.
EB-Habitat58.3%
+2.6 points with particularly strong gains on Common, Complex, and Visual splits.
29 learned skills
EB-ALFRED develops longer manipulation routines: 13 primitive and 16 composite skills.
15 learned skills
EB-Habitat favors compact search-and-transfer abstractions: 5 primitive and 10 composite skills.
Frozen executors
Improvement comes from external procedural knowledge—not repeated executor fine-tuning.
What changes in execution?
Skills recover failures without erasing flexibility.
Paired episode analysis reveals where a learned library helps—and where grounding and state tracking still dominate.
Stable planning structure
Reusable skills compress recurring multi-step behaviors and preserve completed subgoals after local errors.
Environment-specific abstractions
ALFRED learns manipulation routines; Habitat learns repeated object search and transport patterns.
Grounding remains essential
Source inference, object identity, holding state, and long-horizon consistency remain the primary bottlenecks.
Qualitative view
From reusable plans
to grounded action.
Representative trajectories show successful recovery, long-horizon composition, and the boundary between skill structure and online execution.
Citation
Build agents that learn how to improve.
If PRACTICE supports your research, please cite the manuscript.
@article{bai2026PRACTICE,
title = {{PRACTICE}: From Experience to Expertise in Self-Evolving Embodied Agents},
author = {Bai, Ziyi and Li, Siqi and Huang, Tinglei and Karlsson, Börje F.},
year = {2026},
journal = {arXiv preprint arXiv:2608.30760},
url = {https://arxiv.org/abs/2608.30760}
}