Determine the precise onset of falls in Poppy Humanoid locomotion

Determine the precise fall time after the first footstep at which a fall occurs in Poppy Humanoid locomotion, so that fall trajectories can be labeled with finer temporal resolution for learning the quadratic LQR cost function.

Background

The learned cost-function formulation distinguishes successful and failed walking runs by assigning higher accumulated cost to fall trajectories. However, the available experimental labels identify only the footstep during which a fall occurs, rather than the exact time at which the fall becomes unavoidable. This limits the temporal precision of the training constraints and requires the method to assume that the point of no return occurs no later than the first waypoint of the failed footstep.

After a fall begins, joint-angle observations no longer determine the complete robot state because the stance foot may lose firm contact with the ground, leaving the base-frame position and orientation unobserved. Establishing the precise fall time would therefore require finer-grained sensing or labeling before state observability is lost, potentially improving cost-function learning and credit assignment.

References

Our hand is forced in this regard, because manual fall labeling does not allow finer temporal resolution than the footstep level. Consequently, the precise fall time after $n_F$ is unknown -- and as soon as a fall begins, the observed joints will no longer determine the full system state (once the stance foot is not firmly planted, the base frame position and orientation is unknown).

— Closing the Loop on the Poppy Humanoid: Bipedal Locomotion with Linear-Quadratic Control and Learned Cost Functions  (2608.26505 - Chen et al., 27 Aug 2026) in Section 3, subsection “Learned Cost Function”