# Residual Learning from Demonstration

**Authors:** Todor Davchev, Kevin Sebastian Luck, Michael Burke, Franziska Meier, Stefan Schaal, Subramanian Ramamoorthy  
**arXiv:** [2008.07682](https://arxiv.org/abs/2008.07682)  
**Date:** 2020-08-18  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2008.07682v1/

## Key points

- Contacts and friction are fundamental factors in robotic manipulation tasks
- Researchers propose a framework called residual learning from demonstration (rLfD)
- rLfD combines dynamic movement primitives (DMP) with reinforcement learning (RL) based residual correction policy
- rLfD operates in task space, utilizing the full pose of the robot
- rLfD demonstrates superior performance compared to alternative methods
- rLfD enhances the generalization abilities of DMPs
- The researchers train an agent to successfully perform insertions of pegs, gears, and plugs into sockets in simulated and real-world environments
- The study contributes to advancing robotic manipulation capabilities by introducing a novel framework combining DMPs with RL-based residual correction
- Results show improved performance and generalization abilities in contact-rich tasks involving insertion.

## Summary

In the study of robotic manipulation tasks, contacts and friction are fundamental factors that robots must contend with. To address this challenge, researchers propose a framework called residual learning from demonstration (rLfD). This framework combines dynamic movement primitives (DMP) with a reinforcement learning (RL) based residual correction policy. The key feature of rLfD is its application in task space, where it operates on the full pose of the robot. By directly addressing the task space and utilizing the complete robot pose information, rLfD demonstrates superior performance compared to alternative methods. Additionally, rLfD enhances the generalization abilities of DMPs. To evaluate their approach, the researchers train an agent to successfully perform insertions of pegs, gears, and plugs into their respective sockets in both simulated and real-world environments. Overall, this study contributes to advancing robotic manipulation capabilities by introducing a novel framework that combines DMPs with RL-based residual correction. The results demonstrate improved performance and generalization abilities in handling contact-rich tasks involving insertion.