Point Policy does not need robot data, motion planning, simulators, or interactions. While the robotics community has been spending millions of $$$ on collecting huge robot datasets, we show that with just 30 videos of a human performing a task (less than 15 mins of human effort), we can teach robots to perform complex tasks in unseen scenarios with new objects.

Here is a thread on X (prev Twitter) with visuals - https://x.com/haldar_siddhant/status/1895518805151138274

Would love to hear the community's questions, thoughts, and comments! We have open-sourced the entire framework and encourage robotics researchers, engineers and enthusiasts to give it a try on their robots.

Paper: https://arxiv.org/abs/2502.20391 More videos and code: https://point-policy.github.io/