ROBOTERA's VPP2 Tops RoboDojo Robot Manipulation Benchmark

ROBOTERA said in a press release that its VPP2 World Action Model ranked first on the RoboDojo robot manipulation benchmark without extra data or reinforcement based agent self improvement. The model recorded a 32.26% average success rate and an average score of 39.26, with first place results in generalization, precision and memory.

RoboDojo evaluates general purpose robot manipulation across simulated and physical tasks. VPP2 combines video prediction with action generation to predict how actions affect an environment and convert instructions into robot movements. It can also use a vision language model to divide complex instructions into smaller actions.

VPP2 achieved a 58.5% average success rate across 10 task categories on the ALOHA platform and 45% on LIBERO Pro. Adding high level task planning raised its average success rate across five task groups from 27.6% to 57.6%. ROBOTERA has released the model as open source software on GitHub.