Hierarchical robotic systems that perform long‑horizon manipulation must make high‑level semantic decisions to orchestrate stochastic low‑level skills. In this regime failed rollouts are ambiguous: a poor downstream state may stem from an invalid high‑level decision, partial observation, or a correct decision whose physical execution failed. Standard supervised learning lacks data for such recovery states, while reinforcement learning suffers from sparse rewards and non‑local credit assignment.
We introduce MAGMA‑GEN, an on‑policy data‑generation pipeline that converts ambiguous failed rollouts into validated recovery supervision. MAGMA‑GEN first employs a privileged coach to hypothesize an early decision‑level error and propose localized correction or recovery actions. Because this diagnosis can be fallible, candidates are retained only if re‑execution from the same state under matched conditions improves downstream progress. This yields supervised examples from the agent’s own failure distribution without step‑by‑step human demonstrations.
Evaluations on interactive long‑horizon manipulation tasks show that MAGMA‑GEN improves both task success rate and recovery capability compared to distillation and trajectory‑repair baselines, and it adapts to evolving task constraints in both simulation and real‑robot execution.
Review