# IRIS: Implicit Reinforcement without Interaction at Scale for Learning Control from Offline Robot Manipulation Data

**Authors:** Ajay Mandlekar, Fabio Ramos, Byron Boots, Li Fei-Fei, Animesh Garg, Dieter Fox  
**arXiv:** [1911.05321](https://arxiv.org/abs/1911.05321)  
**Date:** 2019-11-13  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1911.05321v1/

## Key points

- Learning from offline task demonstrations in robotics is a significant area of interest
- Challenges arise with larger datasets and longer-horizon tasks that exhibit greater variations
- The IRIS framework addresses these challenges by breaking down the control problem into a goal-conditioned low-level controller and a high-level goal selection mechanism
- IRIS has been evaluated across three datasets, including the RoboTurk Cans dataset collected through crowdsourcing
- The effectiveness of IRIS enables learning of performant policies solely through offline learning methods

## Summary

, , , , 

In the field of robotics, learning from offline task demonstrations is a topic of significant interest. Offline learning from a small set of demonstrations can effectively solve simple short-horizon manipulation tasks with limited variation in task instances. However, challenges arise when dealing with larger datasets and longer-horizon tasks that exhibit greater variations. In such cases, the data may showcase diverse scenarios and include suboptimal solution approaches. To address these challenges, a novel framework called Implicit Reinforcement without Interaction at Scale (IRIS) has been introduced. This framework breaks down the control problem into two components: a goal-conditioned low-level controller that mimics short demonstration sequences and a high-level goal selection mechanism that establishes goals for the low-level controller while selectively combining elements of suboptimal solutions to enhance task completion success rates. The effectiveness of IRIS has been evaluated across three datasets, including the RoboTurk Cans dataset collected through crowdsourcing by humans. The results demonstrate that IRIS enables the learning of performant policies solely through offline learning methods. For further information and visual demonstrations, additional results and videos can be accessed at https://stanfordvl.github.io/iris/. Authors involved in this research include Ajay Mandlekar, Fabio Ramos, Byron Boots, Li Fei-Fei, Animesh Garg, and Dieter Fox. Their work titled "IRIS: Implicit Reinforcement without Interaction at Scale for Learning Control from Offline Robot Manipulation Data" sheds light on innovative approaches to tackling challenges in robotic manipulation tasks through large-scale demonstration datasets.