Geometric Prior Pretraining Improves Imitation Learning Efficiency
What happened
Shogo Iwakata and other authors constructed a geometric visual pretraining dataset: each scene contains only planes, objects, and a block hand, with no textures or backgrounds, and trajectories are generated automatically, allowing the policy to first encounter the geometric relationships between hand and objects; representing the hand as a block avoids binding the dataset to a specific robot morphology. The authors evaluated with ACT on five manipulation tasks each for three simulated robots and on three real-robot tasks. Results show that in most combinations, fine-tuning from the geometric prior achieves higher success rates in early training than training from scratch, and requires only a small number of task demonstrations; based on this, the authors claim that highly simplified geometric scenes can also provide an initialization that transfers across robots and to real tasks.
AI 根据报道生成 · 2 days ago更新
Reporting timeline
Follow the reporting to explore different perspectives.
- arXiv cs.ROGeometric Prior Pretraining Improves Manipulation Imitation Learning Efficiency
The study constructs a geometric visual pretraining dataset containing only a plane, an object, and a cube hand, with trajectories generated automatically and scenes without textures or backgrounds, so that the policy is first exposed to the geometric relationship between the hand and the object. The authors evaluate ACT on three simulated robots with five manipulation tasks each and on three real-world robot tasks; in most combinations, fine-tuning from the geometric prior yields higher success rates in early training than training from scratch, and requires only a small number of task demonstrations. The results indicate that highly simplified geometric scenes can also provide an initialization that transfers across robots and to real-world tasks.
Activity for this event
- Current
- 3
- Peak
- 10Nov 9 14:00
- Past 24 hours
- -50%
The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.