![[2602.18025] Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18025] Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets
Summary
This article presents a novel approach to offline reinforcement learning by integrating cross-embodiment learning to enhance robot policy pre-training across diverse robot datasets.
Why It Matters
The research addresses the challenge of collecting high-quality demonstrations for various robot platforms, proposing a method that utilizes offline reinforcement learning and cross-embodiment techniques to improve efficiency and effectiveness in robot training. This has significant implications for advancing robotics and AI applications.
Key Takeaways
- Combines offline reinforcement learning with cross-embodiment learning for improved robot training.
- Utilizes heterogeneous robot trajectories to develop universal control priors.
- Introduces an embodiment-based grouping strategy to reduce gradient conflicts during learning.
- Demonstrates superior performance in pre-training compared to traditional behavior cloning methods.
- Highlights the limitations of increased suboptimal data and robot types on learning effectiveness.
Computer Science > Artificial Intelligence arXiv:2602.18025 (cs) [Submitted on 20 Feb 2026] Title:Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets Authors:Haruki Abe, Takayuki Osa, Yusuke Mukuta, Tatsuya Harada View a PDF of the paper titled Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets, by Haruki Abe and 3 other authors View PDF HTML (experimental) Abstract:Scalable robot policy pre-training has been hindered by the high cost of collecting high-quality demonstrations for each platform. In this study, we address this issue by uniting offline reinforcement learning (offline RL) with cross-embodiment learning. Offline RL leverages both expert and abundant suboptimal data, and cross-embodiment learning aggregates heterogeneous robot trajectories across diverse morphologies to acquire universal control priors. We perform a systematic analysis of this offline RL and cross-embodiment paradigm, providing a principled understanding of its strengths and limitations. To evaluate this offline RL and cross-embodiment paradigm, we construct a suite of locomotion datasets spanning 16 distinct robot platforms. Our experiments confirm that this combined approach excels at pre-training with datasets rich in suboptimal trajectories, outperforming pure behavior cloning. However, as the proportion of suboptimal data and the number of robot types increase, we observe that conflicting gradients across morphologies begin to imped...
