Behaviour commands: in the robot's control software

The 1 B-step behaviour policy ran every command as trained in the robot's own control software, in simulation with the motor's joint-side constants. It was driven at the 3 cm stairs twice with each command.

• Default walk, crouch, high step and 10° pitch: 2 of 2 crossings; 4 Hz steps 1 of 2 (the other stopped at the first step)
• The crouch walks at 13.2–13.3 cm in the robot's software; standing and standing still hold
• Not the motor's speed limit: the 4 Hz steps crossed 2 of 4 in the robot's software both with the motor's torque-speed curve and with the training simulator's flat 3 N·m limit (30 of 30 in the training simulator)
• Not the riser shape either: on vertical risers in the training simulator the 4 Hz steps pass the 3 cm stairs 29 of 30
• Cause: the contact settings and the ±0.01 rad gear backlash of the robot-software simulation, together: 20–49 s per crossing; either one set as in training, 8–14 s; its whole model as in training, 5 of 6 crossings in 4–7 s
• Next: 200 M more training steps with that gear backlash in the training simulator, against the same steps without it