Summary of Unsupervised Zero-shot Reinforcement Learning Via Functional Reward Encodings, by Kevin Frans et al.
Unsupervised Zero-Shot Reinforcement Learning via Functional Reward Encodings
by Kevin Frans, Seohong Park, Pieter Abbeel, Sergey Levine
First submitted to arxiv on: 27 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper presents a novel approach to zero-shot reinforcement learning (RL) by pre-training a generalist agent on unlabeled offline trajectories. The proposed functional reward encoding (FRE) method learns representations of arbitrary tasks using a transformer-based variational auto-encoder, enabling agents to be adapted to new downstream tasks in a zero-shot manner. Empirical results show that FRE agents can generalize to solve novel tasks in simulated robotic benchmarks, often outperforming previous methods. The proposed solution is scalable and provides a way to solve any new task given a small number of reward-annotated samples. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Can we teach an artificial agent to learn from experience without needing to be told what’s right or wrong? This paper shows that it’s possible! Researchers developed a special kind of learning, called functional reward encoding (FRE), which lets agents learn how to solve new problems without any extra training. They tested this idea by having the agent learn from lots of different “rewards” and then use that knowledge to solve new tasks. The results are exciting – the agent was able to figure out how to do things it had never seen before, often better than previous methods. |
Keywords
* Artificial intelligence * Encoder * Reinforcement learning * Transformer * Zero shot
