# Deep Imitative Models for Flexible Inference, Planning, and Control

**Authors:** Nicholas Rhinehart, Rowan McAllister, Sergey Levine  
**arXiv:** [1810.06544](https://arxiv.org/abs/1810.06544)  
**Date:** 2018-10-15  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.06544v4/

## Key points

- Authors introduce a novel approach combining Imitation Learning (IL) and goal-directed planning
- Development of Imitative Models - probabilistic predictive models of desirable behavior for expert-like trajectories
- Flexible goal objectives include constrained goal regions, unconstrained goal sets, and energy-based goals
- Outperforms six IL approaches and a planning-based approach in dynamic simulated autonomous driving task experiments
- Efficiently learns from expert demonstrations without requiring online data collection
- Robust to poorly specified goals such as those located on the wrong side of the road

## Summary

, , , , 


In their paper "Deep Imitative Models for Flexible Inference, Planning, and Control," authors Nicholas Rhinehart, Rowan McAllister, and Sergey Levine introduce a novel approach that combines the benefits of Imitation Learning (IL) and goal-directed planning. IL is known for its ability to learn desirable autonomous behavior, but directing IL towards arbitrary goals can be challenging. On the other hand, planning-based algorithms leverage dynamics models and reward functions to achieve goals; however, specifying reward functions that elicit desirable behavior is often complex. The proposed solution in this paper is the development of Imitative Models - probabilistic predictive models of desirable behavior capable of planning expert-like trajectories to reach specified goals. These models enable the successful direction of behavior towards desired outcomes by deriving various families of flexible goal objectives such as constrained goal regions, unconstrained goal sets, and energy-based goals. Through experiments in a dynamic simulated autonomous driving task, the authors demonstrate that their approach significantly outperforms six IL approaches and a planning-based approach. Importantly, the method efficiently learns from expert demonstrations without requiring online data collection. Furthermore, the authors highlight the robustness of their approach to poorly specified goals such as those located on the wrong side of the road. Overall, this research contributes valuable insights into enhancing autonomous behavior learning through a combination of imitation learning and goal-directed planning using imitative models. The results showcase promising advancements in achieving specified goals effectively while maintaining interpretability and efficiency in autonomous systems.