Behaviour commands: first run

One policy is to take commands for gait, step frequency, body height, posture and swing height besides velocity, on the multi-terrain field. The first run stopped at 0.9 B of 1 B steps on a GPU memory error and is measured at its last checkpoint.

• Followed on flat ground: body height (13 cm commanded, 14.4 cm measured), step frequency (4 Hz, 4.0 Hz) and body pitch (10°, 3.2° of error against 2.5° when level)
• Not followed: gait (pace, bound and pronk stay a trot) and swing height (8 cm commanded, 5.1 cm measured)
• A 10° pitch command passes 28–30 of 30 trials on every test lane, where the earlier policies fall on all 30
• Test course with the default behaviour: 74.0 % of the trained tests, against 83.0 % for the policy it started from
• Standing still for 20 s: 28 of 30 episodes end at a joint limit