World model simulator works across 22 robot platforms, enables policy learning

PrimeSim generates coherent multiview rollouts for rigid and deformable manipulation, achieves state-of-the-art on two world-model benchmarks, and supports GRPO-based policy optimization from model rollouts.

Independent

Travis Davies · Wentao Ma · Lihao Guo · Tengfei Ma · Xiaowen Chang · Jiafeng Xu · +7 more

Research Digest·4 November 2026·2 min read

Illustration: World model simulator works across 22 robot platforms, enables policy learning

In one line

PrimeSim turns a pretrained video generator into an action-conditioned, multiview world model that simulates robot manipulation across embodiments at 10.1 FPS.

Before you read this

Three ideas this paper builds on

world model for agentsA predictive environment model that agents use to simulate outcomes, enabling planning, reasoning, and long-horizon task execution.102 papers →diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.170 papers →group relative policy optimizationA reinforcement learning method that optimizes policy by computing advantages based on group-level reward statistics.171 papers →

Read it as

Robots learn by trial and error on hardware, which is slow, costly and risky. A computer model that predicts what happens next could act as a safe practice arena, but such a model must follow a robot's commands, not just produce pretty video.

PrimeSim trains a video generator on a large collection of demonstrations from many robots. It uses one common action format, keeps several camera views consistent, and generates from its own past predictions so simulations go far beyond the training window.

In tests against several existing commandable video models, PrimeSim led most quality checks and stayed coherent over long manipulation tasks. That brings learned simulators close to useful for robot practice, though fine-grained local detail still needs work.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

world model for agents long-horizon robot manipulation cross-embodiment learning multiview consistency diffusion model group relative policy optimization