DRIS trains robots to catch with flat plates using parallel physics simulations
Researchers propose Domain-Randomized Instance Set (DRIS), a sim-to-real training technique that runs 10+ physics simulations per timestep to teach robots dexterous manipulation without real-world fine-tuning.
Domain-Randomized Instance Set (DRIS) is a sim-to-real training method that runs multiple randomized physics simulations in parallel during policy learning. Conventional domain randomization samples one set of physics parameters per episode, giving policies limited exposure to the range of dynamics they will encounter on real hardware. DRIS instead maintains a set of 10 or more instances at each timestep, each with different friction, mass, and contact properties, forcing the policy to learn actions that succeed across multiple possible outcomes simultaneously.
The researchers demonstrate the technique on a reactive catching task using a flat plate end-effector—a setup that offers no passive stabilization and requires the robot to react within milliseconds to incoming objects. The trained policies transfer to real hardware without additional fine-tuning, catching objects reliably despite sensor noise and unmodeled dynamics. Theoretical analysis shows that DRIS converges to more robust value functions than single-instance domain randomization, and ablation studies confirm that even 10 instances outperform conventional methods. The flat-plate catching task represents a harder test case than prior work using curved or enclosing grippers, which mechanically stabilize the object on contact.