POD-RoPE: Rotary Position Encoding over Physical Observation Distributions for Embodied AI
POD-RoPE: Rotary Position Encoding over Physical Observation Distributions for Embodied AI
POD-RoPE is a new positional encoding method for embodied AI that represents observation footprints as distributions over camera centers, viewing directions, and sampling times. It enables geometry-aware attention by encoding relative relationships between distributions and supports efficient information selection and content aggregation. Evaluated across multiple robot policy architectures and real-world tasks, POD-RoPE improves success rates and achieves significant speedups over existing methods.
This paper introduces POD-RoPE, a novel positional encoding method for embodied AI that represents observation footprints as distributions over camera centers, viewing directions, and sampling times. The approach improves performance in robot policy architectures and real-world tasks, with notable gains in success rates and computational efficiency.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org