Behaviour commands: where commanded behaviours are practised

Two probes changed how much of training the default walk gets and where commanded behaviours are practised. Neither brings back the 4 cm stairs; keeping commanded behaviours off terrain helps, giving the default walk more of training does not.

• Default walk, 100 trials, 4 cm stairs / 20° slope: commanded behaviours on flat and rough ground only 56 / 65, the same 84 % default share with commanded behaviours everywhere 12 / 34, previous probe (25 % default) 48 / 66, multi-terrain run without behaviours 82 / 100
• At the same share, keeping commanded behaviours off terrain is worth 44 passes on the 4 cm stairs
• More standing-still practice (17 % of segments instead of 5 %): standing robots no longer tuck a leg to the joint limit (0 of 30 ended, against 29 of 30), but stop more often on the 20° slope (31 against 20 of 100)
• With commanded behaviours in only 16 % of training, they are followed less: high step 5.3 cm at an 8 cm command (previous 6.8), 4 Hz fast gait comes out at 2 Hz
• Common to every behaviour run and absent without behaviours: the policy's exploration noise grows (largest std 6–8 against at most 2.7). Next probe: a lower entropy coefficient