Reproduce “Keypoint Action Tokens Enable In-Context Imitation Learning in Robotics” 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
Norman Di Palo, Edward Johns
The Keypoint Action Tokens (KAT) framework enables off-the-shelf, text-pretrained Large Language Models to perform few-shot visual imitation learning in robotics without any additional training. It achieves performance comparable to or surpassing state-of-the-art methods in low-data regimes across diverse manipulation tasks, demonstrating a "free lunch" effect where advancements in general LLMs directly benefit robotics.
Key insights
- Off-the-shelf, text-pretrained LLMs can serve as effective "general pattern machines" for low-level visual imitation learning in robotics, extending their in-context learning capabilities beyond linguistic tasks.
- The choice of visual and action representations is crucial, with 3D keypoints from DINO-ViTs and 3D point triplets for poses proving highly effective for LLM processing.
- A "free robotics lunch" effect exists, where advancements in general-purpose text-pretrained LLMs directly translate into improved robot learning performance without domain-specific training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.