Termsslug: domain-randomization

Domain Randomization

域随机化

Randomizing simulation parameters to improve model generalization to reality

Curated Intelligence

4 items · Page 1/1
OpenAI 官方博客T1·Industry
42 pts
OpenAI Researcher Lilian Weng on AGI Future and Robotics Milestones

OpenAI Applied AI Researcher Lilian Weng outlines her vision for AGI, emphasizing productivity boosts, accelerated scientific discovery, and safe physical interaction. She reflects on the team's milestone achievement of teaching a single robotic hand to solve a Rubik's cube using deep reinforcement learning and domain randomization. The interview highlights the strategic shift from digital intelligence to embodied AI and research assistance.

OpenAI 官方博客T1·Agents
68 pts
OpenAI Dactyl: Sim-to-Real Transfer for Robotic Dexterity

OpenAI introduced Dactyl, a robotic system trained entirely in simulation using general-purpose reinforcement learning to control a human-like hand. The system achieved zero-shot sim-to-real transfer, manipulating objects with unprecedented dexterity without requiring physically accurate world modeling. This demonstrates that large-scale simulation combined with domain randomization can solve complex real-world manipulation tasks.

OpenAI 官方博客T1·Agents
45 pts
2017 Paper: Sim-to-Real Transfer via Dynamics Randomization

This study proposes dynamics randomization to enable robotic control policies trained in simulation to transfer directly to the real world. The core method involves perturbing physical parameters like mass and friction during simulation, forcing the agent to learn robust policies tolerant of model errors. It represents a key early methodology for bridging the Sim-to-Real gap.

OpenAI 官方博客T1·Agents
68 pts
OpenAI Uses Domain Randomization to Boost Robotic Grasping Generalization

OpenAI researchers propose a data generation pipeline combining domain randomization with autoregressive models for robotic grasping. By training on millions of procedurally generated unrealistic objects, the model achieves over 90% success rate on unseen realistic objects in simulation, addressing generalization challenges caused by limited training data diversity.