π0-EqM boosts robotic task success by 10% with equilibrium matching decoder

A new paper on arXiv introduces π0-EqM, an upgrade to the π0 Vision-Language-Action (VLA) robotic control model. The key innovation replaces the flow-matching action decoder with an Equilibrium Matching (EqM) decoder, a change that leaves the upstream VLA stack untouched but dramatically improves closed-loop task performance. Under a fixed 300-step inference budget, π0-EqM raises the average success rate on the RoboTwin benchmark from 40.4% to 50.2% across 19 diverse manipulation tasks. It also achieves competitive results on LIBERO, with a standout 87.0% on the LIBERO-10 suite.

The authors identify what they call the “stationarity–executability gap”—a non-monotonic relationship between residual threshold (how long the decoder runs) and task success. This means that pushing inference depth too far can actually hurt performance, depending on the task. The work positions π0-EqM as an energy-based VLA framework, suggesting that future systems could dynamically adjust compute per control cycle rather than using a fixed sampling horizon. This opens the door to more efficient, task-aware robot controllers that adapt their reasoning in real time.

The practical implications are significant for roboticists deploying VLA models in real-world settings. By swapping in the EqM decoder, teams can achieve meaningful gains without retraining the entire model. The paper also hints at composable action generation across different tasks and robot embodiments, potentially accelerating the development of general-purpose manipulation systems. For now, π0-EqM offers a clear, drop-in improvement for one of the leading VLA architectures.