RoboWorld
A simulation benchmark for robot reinforcement learning, comprising multiple simulation environments and RL test and training environments, used to evaluate model performance on long-horizon tasks under physical constraints.
Timeline (last 90 days)
The RoboWorld dataset contains 25 simulations, totaling 25 test and 175 training reinforcement learning environments, with difficulty distributed across multiple levels.
On the RoboWorld benchmark, the current strongest frontier models fail on every test task, with an average reward of only 0.06.
The RoboWorld paper proposes four paths to improve model capabilities: in-context learning across repeated attempts, a privileged teacher scheme, self-assisted learning, and code-policy methods.
All Facts (3)
The RoboWorld dataset contains 25 simulations, totaling 25 test and 175 training reinforcement learning environments, with difficulty distributed across multiple levels.
50%UnverifiedOn the RoboWorld benchmark, the current strongest frontier models fail on every test task, with an average reward of only 0.06.
50%UnverifiedThe RoboWorld paper proposes four paths to improve model capabilities: in-context learning across repeated attempts, a privileged teacher scheme, self-assisted learning, and code-policy methods.
50%