Description
We are deploying a trained policy on the UnitreeGo2 quadruped robot. We've identified that scaling (by a factor of 0.25) and offset adjustments (constants) are applied to the real sensor data before feeding it into the policy. Despite this, the robot is not able to walk as expected.
The model was trained using rsl_rl in the IsaacSim Orbit environment. We suspect that there might be additional transformations required that we're unaware of.
Question
Are there any other transformations or preprocessing steps needed when deploying the policy on the UnitreeGo2?
Description We are deploying a trained policy on the UnitreeGo2 quadruped robot. We've identified that scaling (by a factor of 0.25) and offset adjustments (constants) are applied to the real sensor data before feeding it into the policy. Despite this, the robot is not able to walk as expected.
The model was trained using rsl_rl in the IsaacSim Orbit environment. We suspect that there might be additional transformations required that we're unaware of.
Question Are there any other transformations or preprocessing steps needed when deploying the policy on the UnitreeGo2?