Faster physics in Python
OpenAI open-sources a high-performance Python library for robotic simulation using the MuJoCo engine, developed from a year of robotics research.
4 articles
OpenAI open-sources a high-performance Python library for robotic simulation using the MuJoCo engine, developed from a year of robotics research.
OpenAI and DeepMind develop a preference learning algorithm to infer human objectives without explicit reward functions, reducing risks of undesirable AI behaviors.
OpenAI explores multiagent environments where agents compete for resources as stepping stones toward AGI. These environments provide natural curriculum (difficulty matched to competitor skill) and no stable equilibrium, creating constant pressure for improvement.
OpenAI introduces an uncertainty-based exploration method (UCB) using Q-ensembles for reinforcement learning. The technique improves exploration by estimating uncertainty through multiple Q-estimators, enabling better exploration-exploitation trade-off.