Survey: training one policy for many terrains

How published work trains a single policy for stairs, slopes and rough ground, and what that means at Pupper's size. The standard recipe is one PPO policy on a tiled terrain field with a curriculum per robot and privileged information given to the critic; the multi-terrain policy followed this recipe.

• legged_gym's field is the template: 8 m tiles with pyramid stairs of 5–21 cm; a robot moves up a level after walking half a tile and down when it covers less than half its commanded distance
• Privileged terrain and state information moved from teacher–student training into the critic and estimator targets. HIMLoco's A1 climbed 16 × 27 cm stairs in 20 of 20 trials, where RMA managed 60 % and Walk These Ways and the vendor controller 0 %
• Sensing the terrain pays on gaps, stepping stones and steps near the robot's limit: on 17 cm stairs an A1's blind policy failed every time where its depth policy succeeded every time
• Pupper's stair tests have no published counterpart: 3 and 4 cm rises are 0.17–0.23 of its 0.1725 m leg against about 0.40 for blind A1 policies, and the printed course's 80 mm treads are 0.53 of its hip spacing against 0.74–0.85 in published stair tests