# What Matters in Learning from Offline Human Demonstrations for Robot Manipulation

**Authors:** Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, Roberto Martín-Martín  
**arXiv:** [2108.03298](https://arxiv.org/abs/2108.03298)  
**Date:** 2021-08-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2108.03298v2/

## Key points

- Challenges and opportunities of using human demonstrations for robot manipulation
- Lack of open-source human datasets and reproducible learning methods
- Extensive study using six offline learning algorithms for robot manipulation
- Evaluation on five simulated and three real-world multi-stage manipulation tasks
- Sensitivity of algorithmic design choices and variability in training and evaluation objectives
- Dependence on demonstration quality for successful learning from human datasets
- Ability to learn proficient policies for challenging multi-stage tasks beyond current reinforcement learning methods
- Scaling to natural real-world manipulation scenarios with only raw sensory signals
- Open-sourced datasets, algorithm implementations, codebases, trained models available at https://arise-initiative.github.io/robomimic-web/
- Contribution to advancing the field and fostering further research in utilizing offline human demonstrations for robot manipulation.

## Summary

In their paper titled "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation," authors Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu and Roberto Martín-Martín address the challenges and opportunities of using human demonstrations to teach robots manipulation skills. They highlight that while imitation learning and batch reinforcement learning have shown promise in this area, the lack of open-source human datasets and reproducible learning methods makes it difficult to assess the current state of the field. To address this gap, the authors conduct an extensive study using six offline learning algorithms for robot manipulation. They evaluate these algorithms on five simulated and three real-world multi-stage manipulation tasks with varying levels of complexity and dataset quality. The study aims to analyze the critical challenges associated with learning from offline human data for manipulation. Based on their findings, the authors derive several lessons. They emphasize the sensitivity of algorithmic design choices and demonstrate how different stopping criteria can lead to variability in training and evaluation objectives. Additionally, they highlight the dependence on demonstration quality as a crucial factor in successful learning from human datasets. The authors also identify opportunities presented by learning from human datasets. They showcase the ability to learn proficient policies for challenging multi-stage tasks beyond what current reinforcement learning methods can achieve. Furthermore, they emphasize that this approach enables scaling to natural real-world manipulation scenarios where only raw sensory signals are available. To facilitate future research and fair comparisons in learning from human demonstration data, the authors have open-sourced their datasets and all algorithm implementations. Interested researchers can access codebases, datasets trained models and more at https://arise-initiative.github.io/robomimic-web/. Overall this paper provides valuable insights into the challenges and potential benefits of utilizing offline human demonstrations for robot manipulation. The authors' extensive study and open sourced resources contribute to advancing the field and fostering further research in this area.