What a conventional parkour teacher does on this A1

Is this “the same as Extreme Parkour”?

In method, yes: delayed joint-position targets, EP's reward set, proprioception plus 132 privileged height scandots, EP's actuator identity, the same PPO configuration. It is built by _a1_real, which extends _a1_ep_faithful. Three deliberate deviations:

1. The body is ours, not EP's. Measured A1 at 14.5 kg all-in, against EP's battery-less 4.713 kg trunk. Roughly triple the trunk mass changes what is feasible.
2. Symmetrised trunk with randomised COM. EP disables COM randomisation, which makes the model's own asymmetry load-bearing — with one fixed offset the optimal policy is asymmetric and RL reliably finds the limp. Training against a symmetric nominal and randomising around it asks for a gait that tolerates the real robot's asymmetry instead of one that depends on the simulator's.
3. The terrain curriculum is extended. Two obstacle types EP never had make up 30% of it, so 2.103 is a slightly harder number than the same policy would post on stock EP.

Why this baseline exists

Every motion-library comparison run so far has been invalid. Four of five collapsed to zero-length episodes, and the one survivor was training through a reward blackout: the reward manager clips the per-step total at zero, and a gait-diversity penalty of 0.075/step against a 0.042/step positive ceiling made the reward identically zero — not small, zero, and so action-independent — across most of the state space. That arm received roughly 0.0010 reward per step where this baseline received 0.0139.

This baseline was untouched by that bug: it carries none of the APT reward terms. So while every library-vs-library comparison had to be thrown out, this measurement stands, and it is the reference the repaired runs are being judged against.