Behaviour commands: where the terrain loss comes from

The default walk of the behaviour runs loses the harder terrain. A control and a measurement trace about half of the loss to the behaviour reward terms charging the default walk for what terrain makes it do.

• After 200 M steps the multi-terrain run without behaviour terms passes the 4 cm stairs 80 times in 100 and the 20° slope 99; the behaviour runs pass 3–8 and 10–17, so the loss comes with behaviour training and not with the restart
• The posture term charges the default walk 0.04 per step on flat ground, 0.30 on the 20° slope and 0.50 on the 4 cm stairs
• On the 4 cm stairs the trained default walk is charged about 1.8 per step by the posture, gait and swing terms together and stops there (0.00 m/s; 0.18 m/s without the terms)
• Probe with those terms off for the default walk, 200 M steps: 4 cm stairs 3 → 32, 20° slope 10 → 46, 7 cm curb 19 → 44; the swing command still works (6.9 cm at 8 cm)
• Withdrawn from the Oct 13 entry: the falling stair level in training is a result of the loss, not its cause