Continual Learning for Embodied AI

A Comprehensive Review and Perspective

Zirui Li1,† (equal contribution), Zhiqi Mao1,† (equal contribution), Haohan Yang1, Yuxin Cai1, Shuo Yang2, Heye Huang3, Jindou Jia1, Yunlong Lin4, Yangang Ren1, Yuan Ji1, Ziying Song1, Dongyue Huang5, Yanjun Huang2, Andrea Cavallaro6, Ziwei Liu7, Chen Lv1,* (corresponding author)

  1. 1NTU School of Mechanical and Aerospace Engineering
  2. 2Tongji University School of Automotive Studies
  3. 3KAIST Cho Chun Shik Graduate School of Mobility
  4. 4Beijing Institute of Technology School of Mechanical Engineering
  5. 5NTU School of Electrical and Electronic Engineering
  6. 6EPFL
  7. 7NTU College of Computing and Data Science

Zirui Li and Zhiqi Mao contributed equally. Corresponding author: Chen Lv (lyuchen@ntu.edu.sg).

BibTeX Library

Manuscript version of 28 October 2026 · literature reviewed to 6 October 2026

Illustration: a robot arm at a kitchen counter, a quadruped on rocky ground, a car on a city road and a humanoid robot in a room, connected by one path.

Abstract

Section 1 and Figure 1

The field at a glance

Figure 1a

Continual learning titles in DBLP, by year

Figure 1b

Forgetting under plain sequential fine-tuning

Values from each source; not comparable across rows.

Section 2

Problem formulation

Operational definition, Section 2.1

Figure 2, Sections 3 to 9

Taxonomy of the reviewed literature

Section 1.3 and the evidence tables

Evidence explorer

Literature cut-off: 6 October 2026.

Section 1.4

Seven propositions

Section 3 and Table 4

Algorithmic mechanisms

Section 4

Foundation models and self-improving agents

Studies with foundation-model learners

Coded studies whose learner builds on a pretrained VLA, VLM or LLM, or that Table 5 lists. Table 5 covers those discussed in Section 4; the others appear in Tables 6 to 11. Each entry names its table.

Self-improving agents

Sections 5 to 8

Tasks, environments, embodiments, and feedback

One tab per section of the review: its opening argument, the authors’ assessment, its figure where there is one, and a profile of the coded studies filed under that axis or, for Section 8, under feedback.

Sections 9.1 and 9.2, Tables 12 and 13

Benchmarks and resources

Sections 9.3 to 9.6

A difficulty hierarchy for evaluation

Sections 10 and 11, Table 14

Open problems and research programs

Conclusion

Figures

The figures of the review. Select one to open it at full resolution with its caption; the arrow keys move between figures.

Library

The works cited in the full version of the review, with DOI or arXiv links where the bibliography records them, and BibTeX entries.

Citation

If you use this review, please cite it as follows.

BibTeX 
@misc{li2026continual,
  title  = {{Continual Learning for Embodied AI: A Comprehensive Review
            and Perspective}},
  author = {Li, Zirui and Mao, Zhiqi and Yang, Haohan and Cai, Yuxin and
            Yang, Shuo and Huang, Heye and Jia, Jindou and
            Lin, Yunlong and Ren, Yangang and Ji, Yuan and
            Song, Ziying and Huang, Dongyue and Huang, Yanjun and
            Cavallaro, Andrea and Liu, Ziwei and Lv, Chen},
  year   = {2026},
  url    = {https://lzrbit.github.io/cl-eai/},
  note   = {Preprint}
}