clobot's G1 walking model in the teleop-walking-benchmark: it walks fine on its own; the harness taking the arms is what stops it
hi; i ran an independent audit of the G1 ports in rhoyn/teleop-walking-benchmark with gaitkeeper, a tool i wrote that reads each port's adapter and runs the policy one change at a time; yours is one where the score undersells the policy
under your env.yaml and with its own arms it stands and tracks for the full 90 s on all 10 seeds; once the benchmark's harness holds the 14 arm joints it lasts about 2 s; so the 2 s score there is about the arms being taken away from a policy that drives all 29 joints, not the walking; the port's gains match your config to within 0.4%
per-run numbers and the stage-by-stage breakdown: https://dundysm.github.io/gaitkeeper/results/#clobot
this is a behavioral finding in my runner (MuJoCo on a CPU, 10 seeds, each seed drawing the benchmark's own run), not a verdict on the policy; if anything looks off against what you see in training or on hardware, i'd like to know
the tool is open source if you want to run it on a newer checkpoint: https://github.com/dundysm/gaitkeeper