GSWorld generates photo-realistic simulation data with multiple camera viewpoints. The videos below showcase our simulation capabilities across different manipulation tasks and robot platforms.

Our zero-shot sim2real approach enables policies trained entirely in GSWorld simulation to perform manipulation tasks directly on real robots without any real-world fine-tuning. The videos demonstrate successful task execution across four manipulation tasks.

DAgger iteratively evaluate policy in GSWorld, record failure cases, recover the scene before failure, and collect corrective data, thus improving polices in a closed loop.

Instead of starting from a randomly initialized policy, we start from a policy trained with real-world teleoperation data.

We benchmark state-of-the-art manipulation policies (ACT and Pi0) in our GSWorld simulation environment. The comparison shows how different methods perform across various manipulation tasks, providing a standardized evaluation platform for robotic manipulation research.

GSWorld enables virtual teleoperation where human operators can control simulated robots in photo-realistic environments. This capability allows for efficient data collection and policy demonstration without requiring physical robot hardware.

We demonstrate zero-shot sim2real transfer for visual reinforcement learning (VRL) tasks. Policies trained in GSWorld simulation using SAC (Soft Actor-Critic) are deployed directly on real robots, showing superior performance compared to training in other simulation environments.