Reproduce “One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning” 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
Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, Sergey Levine
Researchers at UC Berkeley developed Domain-Adaptive Meta-Learning (DAML), a framework enabling robots to acquire manipulation skills from a single human video demonstration. The method uses a learned temporal adaptation objective to implicitly bridge the visual and morphological domain gap between humans and robots, achieving high success rates on tasks like placing (93.8%), pushing (88.9%), and pick-and-place (80%) even with significant cross-perspective domain shifts.
Key insights
- Implicit domain adaptation is achievable through a data-driven meta-learning approach, allowing robots to learn from raw human video despite significant differences in embodiment and viewpoint.
- Incorporating temporal information via a learned adaptation objective is crucial for effective learning from human video demonstrations, outperforming simpler per-timestep adaptation methods.
- Robots can acquire complex manipulation skills from a single, unlabelled human video demonstration after an initial meta-training phase, addressing data efficiency and simplifying skill acquisition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.