Reproduce “One-Shot Visual Imitation Learning via 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
Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, Sergey Levine
The paper presents a meta-imitation learning method that enables robots to acquire new skills from a single visual demonstration. This approach, extending Model-Agnostic Meta-Learning (MAML) with novel architectural components, achieved an 85.8% success rate in simulated pushing and 90% in real-world visual placing tasks, outperforming prior methods.
Key insights
- Gradient-based meta-learning, specifically MAML, can effectively enable one-shot visual imitation from raw pixel inputs for complex robotic tasks.
- Architectural modifications, such as the bias transformation and a two-head structure, are crucial for improving the stability and effectiveness of meta-learning for visuomotor policies.
- The cost of expert action collection can be amortized during meta-training, allowing new skills to be acquired from observation-only demonstrations at test time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.