# Learning to Shift Attention for Motion Generation

**Authors:** You Zhou, Jianfeng Gao, Tamim Asfour  
**arXiv:** [2102.12141](https://arxiv.org/abs/2102.12141)  
**Date:** 2021-02-24  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2102.12141v1/

## Key points

- Generating motion trajectories accurately is a major challenge in robot learning from demonstration
- Prior approaches have failed to capture all modes or averaged them out, resulting in invalid trajectories
- Limited number of demonstrations available may not cover the entire working space
- Researchers propose a novel motion generation model with extrapolation ability that can shift attention between local frames during task execution
- The proposed model uses real-valued non-volume preserving (RealNVP) transformations to learn local latent representations of motion trajectories to address the problem of multiple modes
- The density estimation method provides powerful and stably invertible transformations that can capture all modes present in human demonstrations
- To improve extrapolation ability, the model shifts attention between different local frames during task execution
- Experiments on the docking problem showed that their proposed method outperformed other approaches and has potential applications in various domains such as robotics and automation.

## Summary

In the field of robot learning from demonstration, one major challenge is generating motion trajectories that accurately capture the multiple modes present in human demonstrations for a given task query. Prior approaches have either failed to capture all modes or averaged them out, resulting in invalid trajectories. Another issue is the limited number of demonstrations available, which may not cover the entire working space. To address these problems, researchers propose a novel motion generation model with extrapolation ability that can shift attention between local frames during task execution. The proposed model addresses the problem of multiple modes by using real-valued non-volume preserving (RealNVP) transformations to learn local latent representations of motion trajectories. This density estimation method provides powerful and stably invertible transformations that can capture all modes present in human demonstrations. Additionally, to improve extrapolation ability, the model shifts attention between different local frames during task execution. To evaluate their approach, researchers conducted experiments on the docking problem used in previous works where a trajectory must be generated to connect two dockers without collision. They increased the complexity of this task and demonstrated that their proposed method outperformed other approaches. Furthermore, they evaluated their approach in real robot experiments. Overall, this research proposes a promising solution for generating accurate motion trajectories using robot learning from demonstration techniques by addressing issues related to multiple modes and limited demonstrations. The proposed model has potential applications in various domains such as robotics and automation where accurate motion generation is crucial.