PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing

cs.RO cs.CV Yuheng Ji, Yuyang Liu, Huajie Tan, Xuchuan Huang, Fanding Huang, Yijie Xu, Cheng Chi, Yuting Zhao, Huaihai Lyu, Peterson Co, Mingyu Cao, Qiongyu Zhang, Zhe Li, Enshen Zhou, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang, Xiaolong Zheng · Mar 23, 2026
Local to this browser
What it does
PRM-as-a-Judge addresses the fundamental limitation of binary success metrics in robotic manipulation by repurposing Process Reward Models (PRMs) as dense evaluators. The paper introduces the OPD (Outcome–Process–Diagnosis) metric system,...
Why it matters
The paper introduces the OPD (Outcome–Process–Diagnosis) metric system, which decomposes execution quality via a task-aligned progress potential $\Phi(x_t) \in [0,1]$ induced from trajectory videos. Validated on the RoboPulse benchmark and...
Main concern
This is a methodologically solid contribution to robotic evaluation that moves beyond coarse success rates. The axiomatic foundation—macro-consistency via potential-based additivity and micro-resolution via fine-grained...
Community signal
0
0 up · 0 down
Sign in to vote with arrows
AI Review AI reviewed
Plain-language introduction

PRM-as-a-Judge addresses the fundamental limitation of binary success metrics in robotic manipulation by repurposing Process Reward Models (PRMs) as dense evaluators. The paper introduces the OPD (Outcome–Process–Diagnosis) metric system, which decomposes execution quality via a task-aligned progress potential $\Phi(x_t) \in [0,1]$ induced from trajectory videos. Validated on the RoboPulse benchmark and RoboTwin policy auditing, the work shows that trajectory-supervised PRMs achieve superior micro-resolution compared to foundation models, revealing behavioral signatures invisible to outcome-only evaluation.

Critical review
Verdict
Bottom line

This is a methodologically solid contribution to robotic evaluation that moves beyond coarse success rates. The axiomatic foundation—macro-consistency via potential-based additivity and micro-resolution via fine-grained sensitivity—provides a principled framework for dense evaluation. The OPD metrics (Milestone Coverage, Path-weighted Progress Length, Cumulative Regret Area, and Stagnation Ratio) are well-motivated and mathematically grounded. The empirical validation on RoboPulse demonstrates that PRM judges, particularly Robo-Dopamine, achieve substantially higher accuracy (0.80) than general foundation models (0.47–0.54) on fine-grained progress discrimination, supporting the core claim that domain-specific process reward models serve as effective execution judges.

“Macro-consistency requires an additive, path-independent potential so that local assessments aggregate into a coherent episode-level picture. Micro-resolution requires sensitivity to subtle, task-relevant physical evolution.”
Ji et al., Sec. 3.1 · Section 3.1
“trained on a vast 3,400+ hour dataset”
Tan et al., 2025 · Abstract
What holds up

The theoretical formulation is rigorous and insightful. The potential-difference construction $S(x_i, x_j) = \Phi(x_j) - \Phi(x_i)$ naturally satisfies the macro-consistency axiom (Equation 2), ensuring path-independent aggregation across trajectory segments. The OPD metric system effectively separates outcome reachability (MC, MP) from process efficiency (PPL) and diagnostic failure modes (CRA, STR), enabling granular policy auditing. The RoboPulse benchmark—comprising 1,800 pairwise judgments across 9 embodiment–setting categories with Small/Medium/Large hop stratification—is a valuable contribution for testing progress discrimination at controlled granularities. The failure fingerprint analysis (Figure 5) successfully distinguishes stagnation-dominant failures (high STR) from regret-dominant failures (high CRA), offering actionable diagnostic signals.

“Under the proposed potential-based formulation, a PRM judge assigns each observed state a scalar progress score under a fixed task context, thereby inducing a globally comparable progress ordering within that task.”
Ji et al., Sec. 3.3 · Section 3.3
“PPL (Path-weighted Progress Length): Efficiency of the execution path relative to net progress and path variation. CRA (Cumulative Regret Area): Measures the severity and duration of regression from the best-so-far progress level.”
Ji et al., Table 1 · Table 1
Main concerns

The comparison between PRM judges and foundation models is methodologically problematic. While the paper frames this as "zero-shot" evaluation, the PRMs (Robo-Dopamine, VLAC, GVL) are trained on thousands of hours of robotic trajectory data (3,400+ hours for Robo-Dopamine, 4,000+ hours for VLAC), whereas Gemini and GPT-5.2 are general-purpose models without domain-specific pretraining on manipulation videos. This creates an apples-to-oranges comparison that unfairly disadvantages the foundation models. The claim that PRM-as-a-Judge operates "directly from trajectory videos" elides the fact that these models required extensive domain-specific supervision to learn the progress manifold. Additionally, the theoretical guarantee of macro-consistency assumes a perfect potential function, but real PRMs implemented as neural networks may violate path independence due to stochastic inference or non-convex optimization, a limitation not quantified in the experiments.

“trained on a vast 3,400+ hour dataset spanning real-world, simulation, and human-centric videos”
Tan et al., 2025 · Abstract
“trained on more than 4,000 hours of language annotated manipulation data where temporal ordering yields the progress labels”
Zhai et al., 2025 · Introduction
“All evaluators are tested without task-specific fine-tuning and are queried through interfaces compatible with their native input formats.”
Ji et al., Sec. 5.1 · Section 5.1
Evidence and comparison

The RoboPulse results (Table 2) provide strong evidence for micro-resolution: Robo-Dopamine achieves 0.80 accuracy on Small hops versus 0.54 for Gemini 3 Pro and 0.47 for GPT-5.2. However, without controlling for the domain-specific training data advantage, these comparisons are difficult to interpret as evidence that PRMs are inherently better judges rather than simply better trained for this specific domain. The comparison to CLIP-based methods is more equitable (both trained on broad internet data), though CLIP is not optimized for temporal progress modeling. The RoboTwin policy auditing experiments (Table 3) effectively demonstrate OPD's utility: on Blocks Ranking RGB, $\pi_0$ reaches MC@75 of 40% while OpenVLA-OFT reaches only 6%, despite both having near-zero MC@100—revealing qualitatively different failure regimes. The success-conditioned analysis (Figure 4) validly separates high-quality successes (DP: PPL=94.9) from unstable ones.

“Robo-Dopamine achieves the highest overall accuracy of 0.83. Under the Small hop range, Robo-Dopamine reaches 0.80 average accuracy. Gemini drops to 0.54 and GPT-5.2 to 0.47.”
Ji et al., Table 2 · Table 2
“On Blocks Ranking RGB, $\pi_0$ reaches MC@75 of 40, while OpenVLA-OFT reaches MC@75 of 6, although both have near-zero MC@100.”
Ji et al., Sec. 5.3 · Section 5.3
Reproducibility

The mathematical formalization is thorough, with Appendix A providing rigorous definitions, range proofs, and robustness analyses for all OPD metrics (MC, MP, PPL, CRA, STR). The RoboPulse construction protocol is detailed in Appendices E and F, including hop-based normalization and sampling procedures. However, the paper does not explicitly state code availability or release checkpoints for the specific PRM-as-a-Judge implementation in the provided text. Reproduction would require obtaining the specific Robo-Dopamine or VLAC checkpoints (which are from prior work) and implementing the OPD metric calculations. The reliance on existing PRM checkpoints means hyperparameters for the judges are inherited from prior publications, though the OPD aggregation metrics themselves are deterministic given the potential estimates. Full reproducibility would benefit from explicit release of the RoboPulse dataset splits and evaluation scripts.

“This appendix formalizes the proposed metrics as functionals of a progress-potential trajectory $(\Phi_t)_{t=0}^T$, and records elementary properties used in the main text (range, monotonicity, and stability with respect to bounded judge perturbations).”
Ji et al., Appendix A · Appendix A
“We adopt hop-based normalization within semantically coherent phases. We annotate key frames to segment each episode into semantically coherent phases, retain only phases in which task progress is monotonic.”
Ji et al., Appendix F · Appendix F
Abstract

Current robotic evaluation is still largely dominated by binary success rates, which collapse rich execution processes into a single outcome and obscure critical qualities such as progress, efficiency, and stability. To address this limitation, we propose PRM-as-a-Judge, a dense evaluation paradigm that leverages Process Reward Models (PRMs) to audit policy execution directly from trajectory videos by estimating task progress from observation sequences. Central to this paradigm is the OPD (Outcome-Process-Diagnosis) metric system, which explicitly formalizes execution quality via a task-aligned progress potential. We characterize dense robotic evaluation through two axiomatic properties: macro-consistency, which requires additive and path-consistent aggregation, and micro-resolution, which requires sensitivity to fine-grained physical evolution. Under this formulation, potential-based PRM judges provide a natural instantiation of dense evaluation, with macro-consistency following directly from the induced scalar potential. We empirically validate the micro-resolution property using RoboPulse, a diagnostic benchmark specifically designed for probing micro-scale progress discrimination, where several trajectory-trained PRM judges outperform discriminative similarity-based methods and general-purpose foundation-model judges. Finally, leveraging PRM-as-a-Judge and the OPD metric system, we conduct a structured audit of mainstream policy paradigms across long-horizon tasks, revealing behavioral signatures and failure modes that are invisible to outcome-only metrics.

Challenge the Review

Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.

No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.