Experiment: one policy per direction

Three policies were trained separately for forward, sideways and turning commands, to be blended at run time. None of them learned to walk, with 334 M or with 1 B steps each, so the idea stops at the training stage.

• Forward command 0.6 m/s: 0.002 m/s reached (one joint policy: 0.437 m/s)
• Sideways 0.4 m/s: 0.002 m/s (joint 0.284 m/s); turning 1.6 rad/s: 0.039 rad/s (joint 1.579 rad/s)
• The turning policy twists its body on planted feet: 11° of heading in 8 s (joint 448°)
• Blending the joint policy's own single-direction actions gives 2.2× its speed error and 5.3× its turning error on combined commands