Reproduce “Learning Generalizable Robotic Reward Functions from "In-The-Wild" Human Videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Annie S. Chen, Suraj Nair, Chelsea Finn
Stanford University researchers developed the Domain-agnostic Video Discriminator (DVD), a method for learning generalizable multi-task reward functions by combining limited robot interaction data with extensive "in-the-wild" human videos. This approach enabled a robot to achieve up to double the success rate of baseline methods on unseen tasks and environments in both simulation and real-world experiments, specified by a single human video demonstration.
Key insights
- Combining limited robot data with diverse "in-the-wild" human videos effectively bridges the human-to-robot domain gap for learning generalizable reward functions.
- Training a video discriminator to identify functional task similarity across heterogeneous video sources leads the model to capture abstract functional understanding over superficial visual correlations.
- Leveraging human videos from a greater number of unrelated tasks significantly improves the generalization capabilities of the learned reward function to unseen tasks and environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.