Technical Summary: Pre-VLA

Problem Statement

Large Vision-Language-Action (VLA) models and generative World Models (WM) have advanced embodied intelligence but face significant challenges regarding learning-based generative uncertainty during practical deployment. At each planning cycle, candidate actions generated by VLA policies may suffer from distributional shifts, temporal error accumulation, or insufficient confidence calibration.

These low-quality actions introduce risks in two critical scenarios:

1. Physical Execution: Erroneous actions can cause collisions, object drops, or irreversible kinematic violations, leading to cascading failures.
2. World Model Imagination: Biased or unsafe action inputs can cause the WM to produce blurry frames, target drift, distorted physical relationships, or "false-success" rollouts. This not only accumulates visual and physical errors but also incurs redundant GPU rendering costs.

Existing runtime verification frameworks often struggle with high computational overhead, difficulty in assessing candidate actions before execution, and severe class imbalance in offline data (where safe trajectories dominate, drowning out rare catastrophic failure samples).

Methodology: Pre-VLA Architecture

The authors propose Pre-VLA, a unified runtime verification architecture designed to perform preemptive action validity assessment before actions are dispatched to physical actuators or world model imaginations.

1. Safety-Aware Supervision Signals

To overcome the lack of human annotations for safety boundaries, Pre-VLA derives supervision signals from PPO (Proximal Policy Optimization) rollouts:

• Chunk Advantage Estimation: A pre-trained PPO Critic computes a $K$-step guided chunk advantage ($A^K_t$) based on multimodal context (language, vision, proprioception, and action chunks).
• Failure-Aware Refinement: A safety mask ($H_{t,K}$) truncates value estimation upon irreversible failures (e.g., collisions). A retrospective penalty term ($F_{t,K}$) is introduced to assign exponentially decayed penalties to actions that cause delayed failures.
• Label Generation: The refined advantage is standardized and converted into binary safety labels ($y_{binary}$) and continuous advantage targets ($y_{cont}$) using a task-specific threshold.

2. Lightweight Dual-Branch Monitor

Pre-VLA utilizes an efficient multimodal backbone (based on WorldVLA and Chameleon) with frozen parameters to preserve generative capabilities.

• Modality-Aware Pooling: A dedicated layer separates contributions from text, image, state, and action tokens using predefined masks, producing a unified global feature vector.
• Dual-Branch Head:
  • Classification Branch: Outputs a safety confidence score ($p_t$) using a lightweight Feed-Forward Network (FFN).
  • Regression Branch: Predicts the continuous chunk advantage score ($\hat{A}^K_t$) derived from the Critic.

3. Multi-Task Training Objective

To address class imbalance and boundary instability, the model is trained with a joint loss function:

• Focal Loss ($L_{cls}$): Focuses on hard negative examples (rare failure cases) to mitigate class imbalance.
• Regression Loss ($L_{reg}$): Fits the continuous advantage scores to regularize hidden representations.
• Soft-Threshold Loss ($L_{soft}$): Smooths the decision boundary around the safety threshold to prevent oscillation and improve calibration for ambiguous actions.

4. Dual-Mode Preemptive Resampling Scheduler

During deployment, Pre-VLA operates in two modes:

• Physical Execution: If an action chunk fails the safety threshold, it is intercepted. The system triggers resampling (injecting new noise) up to a limit $N$. If all samples fail, the regression branch selects the candidate with the highest advantage score as a "safe fallback" to maintain control continuity.
• World Model Imagination: Invalid actions are filtered before entering the WM, preventing expensive rendering of misleading futures. If resampling limits are reached, the system may truncate the imagination rollout to save resources.

Key Contributions

1. Unified Runtime Verification: Introduction of Pre-VLA, an architecture that integrates safety constraints directly into the inference pipeline of embodied agents, securing both physical actuators and world model simulators via ultra-low-latency filtering.
2. Multimodal Verification Module: A novel design featuring modality-aware pooling and a dual-branch head that jointly encodes heterogeneous inputs and simultaneously predicts safety confidence and action advantage scores.
3. Robust Training Strategy: A multi-task objective combining Focal classification, advantage regression, and soft-threshold calibration to handle severe class imbalance and stabilize decisions near safety boundaries.
4. Comprehensive Evaluation: Rigorous testing on the LIBERO benchmark demonstrating improvements in closed-loop success rates, execution efficiency, and world model stability.

Experimental Results

Evaluated on the LIBERO manipulation benchmark (LIBERO-10, LIBERO-Goal, LIBERO-Object, LIBERO-Spatial) using RynnVLA-002 as the baseline:

• Action Discrimination: On an independent test set, Pre-VLA achieved an F1 score of 0.8303 and accuracy of 0.9542. It reduced the false pass rate of invalid actions to 0.0200 while maintaining a low false reject rate (0.0491).
• Closed-Loop Success: Pre-VLA improved the average success rate across four suites from 30.79% to 37.62% (an absolute gain of 6.83 percentage points).
• Efficiency: The method reduced the average number of task execution steps, indicating more efficient entry into valid trajectories.
• Runtime Overhead: The average forward verification time per action chunk is 183.9 ms, with total per-step inference times generally remaining under 1000 ms (except in complex spatial tasks), satisfying real-time requirements.
• World Model Stability: Qualitative analysis showed that Pre-VLA effectively mitigates error accumulation, target drift, and visual artifacts in WM rollouts by filtering low-quality actions before prediction.

Significance and Claims

The paper claims that Pre-VLA provides a fundamental prerequisite for trustworthy embodied intelligence by establishing a runtime verification mechanism capable of preemptive safety assessment.

• Safety & Efficiency: It demonstrates that preemptive action verification can serve as an effective safety layer for foundation-model-based robotic control, significantly reducing physical failures and computational waste in world model simulations without requiring extensive retraining of the base policy.
• Generalizability: The approach is shown to be effective even when applied to stronger baseline policies, suggesting that runtime verification is a scalable solution for improving the reliability of generative embodied agents.
• Limitations: The authors acknowledge that supervision signals rely on PPO Critic estimates (potentially biased under distribution shifts) and that repeated resampling in highly adverse states may accumulate inference costs. Future work is proposed to explore online self-calibrating mechanisms.