Removed accelerometer bias

[kristof]

Removed accelerometer[0] += 1.3 from mujoco_infer.py to match what your policy was actually trained with.

Short Summary

JAX Bug (joystick.py):

accelerometer.at[0].set(accelerometer[0] + 1.3)  # Returns new array, but result is discarded

JAX arrays are immutable — .at[].set() creates and returns a new array. Without assigning it back (accelerometer = ...), the original array is unchanged. The +1.3 offset was never applied during training.

NumPy Code (mujoco_infer.py):

accelerometer[0] += 1.3  # Works — NumPy arrays are mutable

NumPy arrays are mutable — in-place modification works. The +1.3 offset IS applied during inference.

Result: Policy trained with accelerometer[0] = X, but inference sees accelerometer[0] = X + 1.3. This observation mismatch can degrade policy performance.