EVE: A Generator-Verifier System for Generative Policies

TL;DR

EVE introduces a modular generator-verifier system that augments frozen generative embodied policies with zero-shot vision-language verifiers to improve task success under distribution shifts, without additional training. The framework orchestrates a diverse verifier ensemble through a guided-diffusion action incorporator and an MMD-based intervention detector, enabling test-time action refinement across diffusion and flow-based policies. Across ManiSkill-HAB and SimplerEnv benchmarks, EVE yields consistent improvements, with verifier ensembles often delivering the largest gains and scaling with verifier capacity. The work demonstrates practical guidance for building scalable, modular verification-augmented policies in embodied robotics, reducing recoveries without retraining while highlighting tradeoffs in latency, complexity, and task-specific benefits.

Abstract

Visuomotor policies based on generative architectures such as diffusion and flow-based matching have shown strong performance but degrade under distribution shifts, demonstrating limited recovery capabilities without costly finetuning. In the language modeling domain, test-time compute scaling has revolutionized reasoning capabilities of modern LLMs by leveraging additional inference-time compute for candidate solution refinement. These methods typically leverage foundation models as verification modules in a zero-shot manner to synthesize improved candidate solutions. In this work, we hypothesize that generative policies can similarly benefit from additional inference-time compute that employs zero-shot VLM-based verifiers. A systematic analysis of improving policy performance through the generation-verification framework remains relatively underexplored in the current literature. To this end, we introduce EVE - a modular, generator-verifier interaction framework - that boosts the performance of pretrained generative policies at test time, with no additional training. EVE wraps a frozen base policy with multiple zero-shot, VLM-based verifier agents. Each verifier proposes action refinements to the base policy candidate actions, while an action incorporator fuses the aggregated verifier output into the base policy action prediction to produce the final executed action. We study design choices for generator-verifier information interfacing across a system of verifiers with distinct capabilities. Across a diverse suite of manipulation tasks, EVE consistently improves task success rates without any additional policy training. Through extensive ablations, we isolate the contribution of verifier capabilities and action incorporator strategies, offering practical guidelines to build scalable, modular generator-verifier systems for embodied control.

Figures (14)

• Figure 1: EVE: A Generator-Verifier Interaction Framework for Generative Embodied Policies
• Figure 2: EVE task performance on $6$ mobile manipulation tasks from the MS-HAB shukla2024maniskill benchmark. Across tasks, we consistently find that verifier-based steering improves base policy task performance. See App \ref{['tab:task_performance_fixed']} for detailed results.
• Figure 3: EVE Individual Component Ablations.
• Figure 4: EVE System Ablations.
• Figure 5: Violin Plots of EVE task performance with $\pi_0$ VLA policy on Simpler-Env simplerenvpizero. We find that verifier-based steering improves performance when using a large VLA policy. See App. Table \ref{['tab:pi_zero_results']} for detailed results.