PrimeSim is a robot world model meant to act as a simulator: it generates multiview video that follows commanded actions and supports long rollouts and policy training.
Training: It starts from a pretrained video generator, then midtrains on 30,110 hours of robot demonstrations across 22 platforms.
Results: It runs at 10.1 FPS, leads four of five WorldSimProbe suites, and posts the best overall EWMBench score, though fine-grained local calibration still trails the field.
Policy learning: The submitter's note says GRPO training inside PrimeSim raises LIBERO success from 61.7% to 73.2%.
Limits: All experiments are in simulation with no physical robot. The abstract says weights and code are released, and the note gives November 6.
Sign in to comment.
TL;DR of the paper, generated automatically.
PrimeSim is a robot world model meant to act as a simulator: it generates multiview video that follows commanded actions and supports long rollouts and policy training.