Online distillation of vision‑language‑action (VLA) policies usually relies on teacher action scores, which are often unavailable when the teacher only offers an inference‑only interface. Trajectory‑only Adversarial Distillation (\gad{}) addresses this by converting a fixed bank of teacher joint‑position features into adaptive feedback: a discriminator compares feature marginals from teacher and student rollouts and supplies rewards for on‑policy learning. The student interacts with the environment without querying the teacher online or receiving task rewards. Experiments on the ManiSkill \texttt{PutOnPlateInScene25Main} task show that \gad{} raises OpenVLA‑OFT (7B) success from $26.2\%$ to $76.6\%$ (+$50.4$ points) and another metric from $20.3\%$ to $58.6\%$. Compared with white‑box on‑policy distillation (OPD) at $81.6\%$ / $59.4\%$, \gad{} lags by only $5.0$ / $0.8$ points (evaluated on $256$ trajectories per checkpoint). A four‑epoch continuation achieves the highest \gad{} result without extra teacher data. This best run consumes $5\times$ the student interaction budget of OPD and $12.5\times$ its full‑rollout batch passes. The comparison demonstrates that VLA policies can be improved online without teacher scoring while quantifying unequal interaction and optimization budgets.
Review