# Privileged Sensing Scaffolds Reinforcement Learning

**Authors:** Edward S. Hu, James Springer, Oleh Rybkin, Dinesh Jayaraman  
**arXiv:** [2405.14853](https://arxiv.org/abs/2405.14853)  
**Date:** 2024-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.14853v1/

## Key points

- Research paper titled "Privileged Sensing Scaffolds Reinforcement Learning" explores sensory scaffolding for novice learners
- "Scaffolder" approach uses reinforcement learning with privileged sensor access to enhance artificial agent performance
- Evaluation on diverse robotic tasks (S3 suite) shows Scaffolder outperforms prior baselines
- Sensory scaffolding enhances artificial agents' learning capabilities in practical scenarios
- Project website demonstrates behaviors learned by Scaffolder, such as spiral search and run-and-jump maneuvers
- Published at ICLR 2024, the research provides insights into leveraging privileged sensing for reinforcement learning and highlights benefits of sensory scaffolding in training artificial agents

## Summary

The research paper titled "Privileged Sensing Scaffolds Reinforcement Learning" delves into the concept of sensory scaffolding and its potential impact on novice learners. The study focuses on training artificial agents using privileged access to expensive and informative sensors during training to improve performance. This approach, called "Scaffolder," utilizes reinforcement learning to effectively exploit privileged sensing in various auxiliary components. To evaluate its effectiveness, the researchers designed a suite of ten diverse simulated robotic tasks known as the "S3" suite. The results show that Scaffolder outperforms relevant prior baselines and highlights how sensory scaffolding can significantly enhance the learning capabilities of artificial agents in practical scenarios. The project website showcases behaviors learned by Scaffolder, including information-gathering strategies like spiral search in Blind Picking and robust behaviors like run-and-jump maneuvers in Blind Locomotion. Overall, this research contributes valuable insights into leveraging privileged sensing for reinforcement learning tasks and demonstrates the potential benefits of incorporating sensory scaffolding in training artificial agents. These findings were published as a conference paper at ICLR 2024 and provide a foundation for further advancements in machine learning algorithms and robotics applications.