Technical Summary: Onto-EV-WM for Failure Diagnosis and Closed-Loop Repair

1. Problem Statement

Physical AI systems utilizing world models (e.g., EV-WM) predict candidate futures and score them based on feature or event vectors. However, these scalar scores or event vectors often lack explicit semantic information regarding why a candidate fails. Specifically, they do not record:

• Which specific task predicate remains unsatisfied (e.g., a spatial relation or joint constraint).
• The typed arguments associated with the failure.
• A route label indicating which correction mechanism should be applied.
• A post-correction acceptance result based on native task predicates.

Consequently, while a system might identify a low-scoring candidate, it lacks a structured interface to diagnose the specific unmet condition, assign a compatible correction strategy, and verify the outcome using a closed-loop protocol. The paper addresses the gap between high-dimensional prediction and symbolic task verification, aiming to provide an explicit interface for diagnosis and repair without replacing the underlying world model or visuomotor controller.

2. Methodology: Onto-EV-WM

The authors propose Onto-EV-WM, an ontology-grounded diagnosis and verification-gated correction interface layered above the existing EV-WM architecture. It is not a replacement world model but a symbolic layer that manages the flow from prediction to correction.

2.1 Architecture and Components

The system operates through a structured pipeline involving three main components:

1. Ontological Record Layer (TBox and ABox):
   • TBox (Task Box): Defines a task-local schema including entity types (e.g., Object, Region, Joint), predicate signatures (e.g., in_region, contact), and type constraints. It serves as a reusable vocabulary for the specific task.
   • ABox (Assertion Box): Instantiates the current state. The system constructs three provenance-specific ABoxes:
     • $A^\star_g$: The required assertions compiled from the structured task specification ($g$).
     • $\hat{A}^{pred}_{g,t+H}$: Assertions grounded from the world model's predicted outputs.
     • $A^{sim}_{g,t}$: Assertions grounded directly from the simulator state.
2. Deterministic Diagnosis and Routing:
   • Diagnosis ($\Pi_D$): Compares the required ABox ($A^\star_g$) against the observed/predicted ABox. It identifies unsatisfied assertions ($\Delta F_{src}$), preserving the specific predicate, its typed arguments, and any continuous margin violations. This generates a typed failure record (e.g., relation_missing(plate, stove_front)).
   • Routing ($\Pi_R$): A deterministic rule base maps the typed failure to a specific "route label." A route is a dispatch label for a compatible correction mechanism (e.g., "source joint pose," "object relation predicate"), not a direct robot action.
3. Verification-Gated Correction Loop:
   • Proposal Generation: A proposer (learned, heuristic, or planning-based) generates a correction $q$ based on the failure record and the selected route.
   • Application: The correction is applied either via direct simulator state mutation (e.g., modifying $qpos$) or controller-mediated execution.
   • Verification: A native task verifier checks the resulting state against the task predicate.
   • Bounded Retry: If verification fails, the system increments an attempt budget ($b$), regrounds the state, and may retry the same route or reroute. The loop terminates on success or budget exhaustion.

2.2 Operational Flow

The interface separates prediction, grounding, diagnosis, application, and verification.

• Input: Task specification, source tag (prediction or simulation), and state.
• Process: Ground the state into an ABox $\rightarrow$ Diagnose missing predicates $\rightarrow$ Assign a route $\rightarrow$ Generate correction $\rightarrow$ Apply $\rightarrow$ Verify.
• Output: An accepted state, a retained planning candidate, or an unresolved typed failure trace.

3. Key Contributions

The paper makes three primary technical contributions:

1. Operational Robot Task Ontology: A design for an explicit grounding interface that represents predicted and simulator-observed states as distinct, task-specific ABoxes under a shared vocabulary and type constraints.
2. Deterministic Diagnosis and Routing: A rule-based system that preserves violated predicates and their typed arguments, mapping them to task-specific correction routes without discarding the original context.
3. Verification-Gated Correction Contract: A protocol that records the full lifecycle of a correction attempt (diagnosis, route selection, proposal, application mode, and native-predicate acceptance) and enforces a bounded retry mechanism.

4. Experimental Results

The authors evaluate Onto-EV-WM across three distinct benchmark settings using simulation protocols.

4.1 PointMaze (Aligned Comparison)

• Setup: 50-trial random-state planning comparison between EV-WM and Onto-EV-WM.
• Results: Both methods achieved a 94% success rate. However, Onto-EV-WM achieved a significantly lower mean final-state distance (0.61177) compared to EV-WM (0.90573), indicating higher precision in satisfying task conditions.
• Note: A separate configuration with a larger search budget and success-first selection reached 100% success, but this is not directly comparable to the aligned setting due to budget differences.

4.2 LIBERO-Goal (Sampled-Window Correction)

• Setup: Evaluation on 10 manipulation tasks using 4 evaluation-sampling seeds. The protocol samples 25-step demonstration windows. Corrections are applied via a fixed learned source/joint $qpos$-delta mechanism directly to the simulator state.
• Results:
  • Seed 0: 93.8% corrected-window success (469/500).
  • Across 4 Seeds: 94.05 ± 0.30% corrected success.
  • Recovery: Out of 261 replay failures, 142 were "rescued" (54.4%) by the correction mechanism.
• Context: The ontology provides the diagnostic record linking the failure to the fixed correction head; the reported metric reflects the complete ontology-grounded configuration.

4.3 LIBERO-Plus (Fixed Registry)

• Setup: Evaluation on a fixed registry of 10,030 perturbation tasks across four suites (LIBERO-10, Goal, Object, Spatial).
• Results:
  • Overall Success: 85.00% (8,526/10,030 tasks).
  • Suite Breakdown:
    • LIBERO-Goal: 91.39%
    • LIBERO-Object: 91.38%
    • LIBERO-Spatial: 91.38%
    • LIBERO-10: 65.98% (identified as the suite with the highest residual failure count).
• Comparison: The Onto-EV-WM configuration outperforms several baselines, including DINO-WM + CEM (61.57%) and various VLA models, though the comparison is system-level and does not isolate the causal contribution of the ontology alone.

5. Significance and Claims

The paper positions Onto-EV-WM as a system-level interface that enhances physical AI systems by adding explicit, typed reasoning to the failure diagnosis and repair process.

• Modest Scope: The authors explicitly state that the results do not constitute:
  • Real-robot recovery or sim-to-real validation.
  • A general open-world knowledge graph or universal robot ontology.
  • An ontology-only causal share (the performance gains are attributed to the complete configuration).
  • A new learned policy class replacing existing controllers.
• Core Value: The significance lies in the ability to record why a task failed (typed predicates and arguments) and to structure the repair process (routes and verification) in a way that is decoupled from the underlying prediction model. This allows for "closed-loop" repair within simulation protocols where verification gates the acceptance of corrections.
• Limitations: The evaluation relies on simulation-based protocols. The "closed loop" refers to a bounded verify-retry cycle within the simulator, not real-time feedback control on physical hardware. The quantitative results reflect the performance of the integrated system under specific protocols rather than the isolated efficacy of the ontology.