When the Gait Slips: Retrieval-Conditioned Reflex Plasticity for Humanoid Fall Recovery

Animals reliably recover from stumbles through rapid, experience-shaped mechanical responses. Inspired by this, we propose Retrieval-Conditioned Reflex Plasticity, a terrain-adaptive CoM trigger that detects falls from a 20 cm CoM height drop and low vertical velocity. A window of recent body states are encoded with an LSTM-VAE \cite{sailor} to retrieve recovery motions from MoCap libraries \cite{lee2025phuma, mahmood2019amass}, which conditions Feature-Wise Linear Modulation (FiLM) modulation \cite{perez2018film} of the actor trunk as $\(h'=(1+c\gamma_t)\odot h+c\beta_t\)$, where $\((\gamma_t,\beta_t)=P_\phi(\tau)[t]\)$ and $\(c\in\{0,1\}\)$. The reflex remains active until recovery, retrieving new segments if instability persists. We fine-tune the plasticity hypernetwork on 3.2G-step AMP-SAC policy \citep{lessa2025unlocking} for 0.44G steps on 2-5 cm rough terrain in Isaac Lab. Initial results and future work: Figure 1 shows a stumble-to-recovery sequence at 2 fps on random 2-10 cm rough terrain, along with CoM height and its adaptive threshold. The blue interval marks FiLM action modulation after retrieval is triggered. Here, a retrieved jumping motion produces a corrective step and brief jump, visible in both the frames and the CoM trajectory. Compared with the non-plastic baseline, retrieval-conditioned plasticity reduced falls by 8\% on random 2-5 cm rough terrain and recovered from 25\% of 120 triggered destabilizations on random 2-10 cm rough terrain. On random 2-15 cm rough terrain, it increased policy survival time by 18.4\%. Under a stronger matched push at $\(t{=}5\)\,s$, it sustained locomotion $\(\sim 16{\times}\)$ longer after disturbance. Nominal gait remains unchanged when $\(c{=}0\)$. Future work will expand retrieval across terrains and non-walking skills.