Behaviour commands: holding the default walk

Three ways to stop the falls during behaviour training changed nothing that matters. Holding the default walk to the multi-terrain policy brought its terrain skill back, but then the commands were not learned.

• Lower entropy coefficient: the policy's noise grew more (largest std 24.9 against 7.5). Noise capped at 2.0: falls as before, and 54 of 100 stops in front of the 4 cm stairs. Bounded updates (adaptive rate, clipping): no spikes until late, falls as before
• Default walk held to the multi-terrain policy by a KL term on default-walk steps: falls on flat ground in training 2 % (previous probe 23 %, multi-terrain run 3 %)
• Default walk, 100 trials, 4 cm stairs / 20° slope / 7 cm curb: held 83 / 95 / 66, previous probe 49 / 70 / 39, multi-terrain policy 83 / 99 / 45
• But the commands are not followed: a 4 Hz gait command gives 1.73 Hz, an 8 cm high step 5.2 cm (previous probe 6.7 cm)
• Standing robots no longer end at the joint limit (0 of 30)
• Next: weaker holds, 0.1 and 0.01 of the weight, to find a point that keeps both