# NAVINACT: Combining Navigation and Imitation Learning for Bootstrapping Reinforcement Learning

**Authors:** Amisha Bhaskar, Zahiruddin Mahammad, Sachin R Jadhav, Pratap Tokekar  
**arXiv:** [2408.04054](https://arxiv.org/abs/2408.04054)  
**Date:** 2024-08-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.04054v1/

## Key points

- NAVINACT is a framework that combines Reinforcement Learning (RL) and Imitation Learning (IL) to bootstrap RL in real-world robotic tasks.
- It addresses challenges in exploration and generalization by dynamically switching between motion planning-based navigation and learning policies based on the task at hand.
- NAVINACT incorporates imitation data to enhance efficiency in navigating complex environments.
- Its multi-head architecture includes ModeNet for mode classification, NavNet for waypoint prediction, and InteractNet for manipulation tasks.
- NAVINACT improves sample efficiency, addresses distribution shift issues, and ensures robust task execution through its approach.
- Extensive evaluations show superior performance in adaptability, efficiency, and generalization compared to existing methods.
- In simulated scenarios, NAVINACT outperforms baseline approaches by 10-15% in training success rates at 30k samples and by 30-40% during evaluation phases.
- In real-world settings, it achieves a 30-40% higher success rate on simpler tasks compared to baselines while excelling in complex two-stage manipulation tasks.

## Summary

, , , , 

NAVINACT is a cutting-edge framework that combines Reinforcement Learning (RL) and Imitation Learning (IL) to effectively bootstrap RL in real-world robotic tasks. It addresses challenges in exploration and generalization by dynamically switching between classical motion planning-based navigation and learning policies based on the task at hand. By incorporating imitation data, NAVINACT enhances efficiency in navigating complex environments. Its multi-head architecture comprises ModeNet for mode classification, NavNet for waypoint prediction, and InteractNet for manipulation tasks. Through this approach, NAVINACT improves sample efficiency and addresses distribution shift issues, ensuring robust task execution. Extensive evaluations across simulation environments and real-world tasks demonstrate its superior performance in terms of adaptability, efficiency, and generalization compared to existing methods. In simulated scenarios, NAVINACT outperforms baseline approaches by 10-15% in training success rates at 30k samples and by 30-40% during evaluation phases. In real-world settings, it achieves a 30-40% higher success rate on simpler tasks compared to baselines while excelling in complex two-stage manipulation tasks. Inspired by HYDRA's concept of switching between sparse (classical motion planning) and dense (learning) modes [29], NAVINACT refines this idea to optimize RL outcomes through efficient task execution. Overall, NAVINACT represents a dynamic solution that combines navigation and imitation learning to bootstrap reinforcement learning effectively for diverse environments while maintaining robust performance. For further details including datasets and supplementary materials, visit our website at {https://raaslab.org/projects/NAVINACT/}.