PhySA: Decoupling Aesthetic and Physical Rewards for Humanoid Locomotion
PhySA: Decoupling Aesthetic and Physical Rewards for Humanoid Locomotion
This paper introduces PhySA, a method that separates aesthetic scoring from physical smoothness objectives in humanoid locomotion policies. By using vision-language models for aesthetic rewards and adding a phase-triggered correction channel for physical penalties, the approach reduces joint acceleration by 26-32% on MuJoCo Humanoid-v5 compared to baseline methods like PPO, Text2Reward, and Eureka, though it results in shorter episodes. The method shows limited transferability to other simulators or hardware without rescaling.
The study demonstrates that decoupling aesthetic rewards from physical smoothness constraints in humanoid locomotion training reduces joint acceleration by up to 59.5% on MuJoCo Humanoid-v5, though gains do not transfer to IsaacLab or Unitree G1 without rescaling.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
training steps
2,600,000
View original evidence
mulator state. Ablations attribute the smoothness gain to the correction channel. On MuJoCo Humanoid-v5 at 2.6M steps, PhySA lowers joint acceleration by 26 to 32% relative to vanilla PPO, Text2Reward, and Eureka (three seeds), at the cost of shorter episodes. In a single extended run it lowers joint acceleration by 59.5% relative to Prompt2Policy, while a three-seed replication shows a smaller, non-significant reduc
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
mulator state. Ablations attribute the smoothness gain to the correction channel. On MuJoCo Humanoid-v5 at 2.6M steps, PhySA lowers joint acceleration by 26 to 32% relative to vanilla PPO, Text2Reward, and Eureka (three seeds), at the cost of shorter episodes. In a single extended run it lowers joint acceleration by 59.5% relative to Prompt2Policy, while a three-seed replication shows a smaller, non-significant reduc
Open source S4
tion (7.2%). The gain does not transfer to IsaacLab or the Unitree G1, even after rescaling the smoothness term.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org