Reward Function Design
A core concept in reinforcement learning that defines the objective an agent should pursue and guides policy optimization.
Timeline (last 90 days)
In reinforcement-learning locomotion research, if the reward function is designed too specifically, all robots converge to similar gaits; if it is too loose, training may fail to converge or learn odd behaviors that exploit loopholes.
The reward function for legged locomotion control typically needs to simultaneously encode multiple competing objectives such as velocity tracking, energy efficiency, posture stability, and contact force constraints, and weight deviations can lead to infeasible behaviors on the real machine.
All Facts (2)
In reinforcement-learning locomotion research, if the reward function is designed too specifically, all robots converge to similar gaits; if it is too loose, training may fail to converge or learn odd behaviors that exploit loopholes.
50%UnverifiedThe reward function for legged locomotion control typically needs to simultaneously encode multiple competing objectives such as velocity tracking, energy efficiency, posture stability, and contact force constraints, and weight deviations can lead to infeasible behaviors on the real machine.
50%