EVE: A Generator-Verifier System for Generative Policies

Rather than finetuning or retraining a base policy with expensive in-domain recovery data, EVE (Embodied Verifier Ensembles) leverages frozen, frontier vision-language models (VLMs) as zero-shot verifiers in a unified generator-verifier architecture. At each intervention point, EVE queries multiple verifiers with distinct action-feedback strategies and fuses their corrections into the base VLA through a classifier-guidance-based action incorporator. The framework is organized around four components:

• Base Policy Candidate Generation. Given an instruction, observation, and proprioceptive state, the base VLA samples K diverse candidate action trajectories by denoising from independent noise samples.
• Verifier Agents. A collection of verifier modules each provide structured and dsitinct action feedback. Generator-Agnostic verifiers operate only on sensor observations and select from predefined recovery action primitives (the Primitive steerer). Generator-Conditioned verifiers additionally consume a representation of the candidate trajectories and select among them (the Pivot steerer, built on PIVOT-style prompting). Verifier outputs are aggregated into a common trajectory representation using a weighted average.
• Action Incorporator. Instead of overriding or naively averaging actions, EVE uses a guided diffusion process that steers denoising toward verifier-consistent behavior while preserving the base VLA's learned action prior. A guidance coefficient controls how much verifier feedback is injected at each reverse-diffusion step.
• Intervention Detection. To avoid costly VLM calls at every step, EVE invokes verifiers only at automatically detected intervention points — whenever the cumulative Maximum Mean Discrepancy (MMD) of the sampled action distribution exceeds a threshold, flagging erratic or out-of-distribution behavior.

The result is an inference-time action refinement algorithm that produces semantically grounded recovery feedback and interpolates it with the base VLA actiondistribution — requiring no finetuning of policy weights.

On the SimplerEnv real-to-sim benchmark, EVE with training-free, zero-shot verifiers outperforms state-of-the-art finetuned verifiers. EVE-Ensemble attains the best total average success rate of 72.2, improving over the base policy π₀ (67.1) and over trained-verifier baselines RoboMonkey (68.1, trained on 20M synthetic preferences) and V-GPS (27.3, trained on 175K demos) — while using zero training data. The largest gains concentrate on harder, lower-success tasks such as Put Apple in Drawer (29.9 → 40.3) and Stack Blocks (56.2 → 63.2).

EVE generalizes to new embodiments and diverse tasks where trained-verifier baselines (RoboMonkey, V-GPS) are not applicable, since they require extensive in-domain training. On the ManiSkill-HAB long-horizon mobile manipulation suite (mean over 1008 rollouts), EVE consistently improves a base diffusion policy. On RoboTwin 2.0 with an Aloha AgileX dual-arm embodiment, EVE improves the π_0.5 VLA.

We validate EVE on a real Franka Emika Panda arm. We finetune the π_0.5 VLA on 50 demonstrations of placing two blocks into a bin, and evaluate on 5 tasks (10 trials each): 1 in-distribution task, 2 OOD-Prompt tasks (pick blocks in a specific order), and 1 OOD-Object task (pick coffee pod). EVE matches RoboMonkey in-distribution and achieves the best performance across all OOD settings via semantically-informed verifier-guided corrections, while RoboMonkey degrades sharply in the OOD-Prompt setting.

We ablate the core components of EVE on the SetTable-Place task from ManiSkill-HAB, reporting the delta in success rate (%) between steered and unsteered runs. (a) Verifier model scaling: larger VLM verifiers steer more effectively, with Qwen-2.5-VL-72B consistently beating 7B and 32B. (b) Action incorporator design: guided diffusion outperforms both verifier override and naive averaging, integrating just enough verifier feedback to prevent failure while staying near the base policy's action distribution. (c) Guidance ratio: performance peaks at a guidance coefficient of 10, with sharp drops at neighboring values.

On a failed Place Apple in Drawer rollout, the base policy π₀ only partially opens the drawer, causing the apple to fall out. EVE detects an MMD spike and intervenes: the Primitive steerer proposes a "Nudge Left" action toward the handle, while the Pivot steerer selects trajectories that maximize outward pulling of the drawer. The averaged feedback guides denoising so the drawer opens fully and the apple is placed successfully — showing how complementary verifier guidance recovers a failed trajectory.

On a failed Move Near rollout, the base policy executes a faulty grip and the can tips over. EVE detects an MMD spike during the approach phase. Here the Primitive steerer proposes a corrective "Nudge Left", while the Pivot steerer incorrectly selects an unhelpful trajectory. Because EVE aggregates feedback across verifiers, the averaged signal still yields a small leftward correction that achieves a stable grasp — illustrating EVE's robustness to imperfect verifier feedback.