With no term on foot attitude the robot stood on the edges of its feet (ankle roll -26 deg) from R0.5 to V2.5b; pricing it fixed that and turned stance width into the next free variable - the user saw both on video before any metric flagged them
unpriced-foot-attitude-is-a-free-variable.List every posture quantity the hardware cares about (foot attitude, stance width, yaw, knee flexion) and make sure each is priced by a term with real gradient at the observed error; pricing one exposes the next, so re-inspect the feet-level video after every change.
Symptom
Watching the v2_5 video the user said the ankle roll after standing looked very strange - the feet were not flat. The feet view showed the right foot standing on its outer edge; median ankle roll at t = 8 s was -/+26 deg (limit +/-35), mirrored. This was the physical form of the ankle_roll saturation criterion that had failed since R0.5.
Context
No term priced foot attitude: feet_contact_upright counts an edge contact as contact, and stand_pose's wide exp kernel (sigma^2 = 9) gives almost no gradient at 0.45 rad. HoST carries an "ankle parallel" term (+20); this table never had one. Edge standing was made a blocking precondition for hardware (continuous ankle load, unstable contact, wear).
Change
V2.6 (single variable): flat_feet = (|q_l_ankle_roll| + |q_r_ankle_roll|) x upright gate x height gate, a linear hinge with a 5 deg margin, weight -2, continued from V2.5b.
Outcome
Ankle roll -/+26 -> 5.3/5.2 deg, ankle_roll saturation 6.6% (< 10%), feet flat on the feet-view video; success 99.6%, re-falls 0-1%. Then the user watched v2_6: hip yaw constantly tense and the legs very close together. The numbers: hip roll -/+4.9/4.8 deg against a nominal 25 - feet flat and hips open 25 deg cannot coexist without ankle compensation, the new term taxed that compensation, and nothing priced stance width. "Foot attitude as a free variable" was fixed and "stance width became the new free variable" - which the real robot then exposed as splits.
Mechanism
An optimizer spends every posture degree of freedom no term prices; closing one reallocates the slack to the next unpriced one.
Applies when
- a standing or landing posture looks wrong on video while gates pass
- a saturation criterion keeps failing on one joint
- a new posture term was just added
“用户看 v2_5 视频:"起身之后 ankle_roll 非常奇怪,脚根本不是平着站立"。 … 机理:奖励表**无任何脚掌姿态项** —— feet_contact_upright 边缘接触也算触地, stand_pose 的 exp 核(σ²=9)对 26°=0.45 rad 梯度≈0。脚掌姿态是自由变量。”