Multi-terrain training: field, test course, curriculum

One blind policy is to cover flat ground, rough ground, slopes, stairs and boxes. The training field, the test course and the training setup are built, and the existing policies are measured on the test course as the baseline.

• Field: 7 terrain families × 14 difficulty levels on 2 m tiles
• Test course: 34 tests on 38 lanes, 14 of them terrain that training never shows
• Baseline pass share on trained / held-out tests: our stair policy 84 % / 86 %, Stanford's climbing policy 62 % / 56 %, our flat policy 47 % / 32 %
• Our stair policy on 5 cm stairs: 15 of 100 on 80 mm treads, 92–94 on 160–240 mm treads
• Level up after 3 crossings in a row, down after one failure (settles near a 79 % pass rate)
• The critic sees 126 extra simulator values; the policy on the robot is unchanged