Unsupervised Zero-Shot Reinforcement Learning via Functional Reward Encodings

Authors: Kevin Frans, Seohong Park, Pieter Abbeel, Sergey Levine

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that FRE agents trained on diverse random unsupervised reward functions can generalize to solve novel tasks in a range of simulated robotic benchmarks, often outperforming previous zero-shot RL and offline RL methods.
Researcher Affiliation Academia Kevin Frans 1 Seohong Park 1 Pieter Abbeel 1 Sergey Levine 1 1 University of California, Berkeley kvfrans@berkeley.edu
Pseudocode Yes Algorithm 1 Functional Reward Encodings (FRE)
Open Source Code Yes Code for this project is provided at: github.com/kvfrans/fre.
Open Datasets Yes We utilize the antmaze-large-diverse-v2 dataset from D4RL (Fu et al., 2020). [...] The Ex ORL dataset is a standard collection of offline data for RL, consisting of trajectories sampled by an exploratory policy on Deep Mind Control Suite (Tassa et al., 2018) tasks.
Dataset Splits No The paper discusses training procedures and data sampling for encoding/decoding, but does not provide specific train/validation/test dataset splits with percentages or counts for reproduction.
Hardware Specification No This research used the Savio computational cluster resource provided by the Berkeley Research Computing program at UC Berkeley. This mentions a cluster but lacks specific hardware details like CPU/GPU models.
Software Dependencies No Appendix A lists the optimizer as "Adam" and mentions "IQL Expectile", but does not specify version numbers for any software, libraries, or frameworks used (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes Appendix A. Hyperparameters [lists specific values for Batch Size, Training Steps, Learning Rate, Network Layers, etc.]