Summary of Learning An Actionable Discrete Diffusion Policy Via Large-scale Actionless Video Pre-training, by Haoran He et al.
Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training
by Haoran He, Chenjia Bai, Ling Pan, Weinan Zhang, Bin Zhao, Xuelong Li
First submitted to arxiv on: 22 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes a novel framework to pre-train and fine-tune robotic agents using human videos. The approach addresses the scarcity of action-labeled robotic datasets by leveraging a unified discrete diffusion model that combines generative pre-training on human videos with policy fine-tuning on a small number of action-labeled robot videos. The method starts by compressing both human and robot videos into unified video tokens, which are then used to predict future video tokens in the latent space during the pre-training stage. In the fine-tuning stage, the imagined future videos guide low-level action learning with a limited set of robot data. Experimental results show that this approach generates high-fidelity future videos for planning and enhances the fine-tuned policies compared to previous state-of-the-art approaches. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps robots learn new tasks by using lots of videos of humans doing different things. The problem is that these videos don’t have labels, which makes it hard to teach the robot what each action means. To fix this, researchers came up with a way to use these videos to pre-train the robot’s learning system. They then fine-tune the robot’s actions by showing it a few examples of what the actions look like in real life. This approach worked really well and was able to teach the robot new tasks faster than before. |
Keywords
* Artificial intelligence * Diffusion model * Fine tuning * Latent space
