Event

POD-RoPE: Physical Observation Distribution Positional Encoding for Embodied AI

1 reports · 1 sources

Xiaodi Huang and co-authors propose the POD-RoPE positional encoding method, which represents a token's observation footprint as a distribution over camera centers, viewing angles, and sampling times, encodes relative relationships between distributions using expected rotations, and reduces to per-point RoPE when the distribution degenerates to a single point. The authors report that after integrating POD-RoPE into Fast-WAM, the average success rate across all 50 tasks in RoboTwin 2.0 increased from 91.83% to 92.88%, and two real-robot tasks on Sharpa North increased from 32.5% to 52.5%. In the H20 model-block benchmark, the method's median speedup for forward plus backward is 1.91× that of P-RoPE and 5.34× that of RayRoPE, with a median latency overhead of 17.4% relative to standard RoPE.