HarnessWAM: Bridging Prediction and Deliberation in World Action Models
1. Paper overview
In one sentence: HarnessWAM improves long-task execution by combining evidence memory, constrained skill compilation, and event-triggered verification, while remaining limited by the executor’s validated skills and the reliability of semantic judgments. e-probleme-architecturee-statee-projectione-protocole-main
| At a glance | What to know |
|---|---|
| Research problem | Source description A finite-horizon controller can execute a local instruction yet lack the cross-stage knowledge needed to finish a partially observed task. In the drawer example, exploration must reveal the target before placement can be planned. The authors call this mismatch the prediction–deliberation gap. e-problem |
| Core mechanism | Source description An external runtime combines a VLM Task Manager, evidence-linked scene belief, task graph, deterministic executable-space projection, progress monitoring, and recovery without changing the WAM architecture. e-architecturee-statee-projection |
| A key reported result | RoboMemArena overall manipulation success: 59.6 / 69.9 Full-task / subtask success (%). 26 tasks; HarnessWAM and same-WAM diagnostics use 20 rollouts per task with matched initial states, seeds, interfaces, and task budgets; macro-averaged across tasks and rollouts. Whole Task: 44.4 / 52.3; Static Plan: 47.9 / 62.0; published PrediMem: 38.5 / 55.2. The gain over Static Plan is 11.7 / 7.9 percentage points. Full-task success requires all stage predicates; subtask success averages the completed fraction. Published baselines provide context with weaker experimental control. e-benchmarkse-metricse-protocole-main |
| Reading caution | Source description The conclusion leaves real-world manipulation, broader validated skill repertoires, and calibrated uncertainty-aware deliberation to future work. Benchmark evidence does not establish physical deployment. e-limits |
Core contributions
- Source description
An external runtime combines a VLM Task Manager, evidence-linked scene belief, task graph, deterministic executable-space projection, progress monitoring, and recovery without changing the WAM architecture. e-architecturee-statee-projection
- Author claim
The authors attribute improved task reliability to persistent orchestration beyond initial decomposition. Same-WAM comparisons and component removals support this system-level claim within the evaluated benchmarks. e-maine-ablatione-protocol
Figure 2. Semantic planning reaches the robot through a compiler and an execution feedback loop. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the left with the instruction, camera views, and robot state. The upper blocks exchange a task graph and evidence memory with the VLM Task Manager. Follow the downward path into executable-space projection: semantic subgoals pass through capability, dependency, and robot-state checks before becoming WAM-supported primitives. At the bottom, the WAM interacts with the execution runtime and environment. New observations feed subsequent state updates and decisions. The runtime’s milestone, timeout, and stage-change labels identify occasions for deliberation. Section 3.4 and Algorithm 1 clarify that a progress event requests semantic judgment; it does not itself authorize task advancement. e-architecturee-projectione-progresse-decisionse-implementation
What it supports. The architecture separates choosing a valid local goal from generating its continuous actions. This makes the compiler and verifier part of inference-time behavior. Qwen3-VL-32B-Instruct supplies the Task Manager, while benchmark-fine-tuned LingBot-VA remains the local executor; the paper changes their orchestration without redesigning the WAM.
Where the evidence stops. Figure 2 ends its action strip at \(a_{t+H}\), whereas Equation (1) ends an H-action chunk at \(a_{t+H-1}\). This edition follows the equation and preserves the original graphic; the indexing discrepancy is unresolved.
2. Motivation
2.1 The problem and the proposed response
A finite-horizon controller can execute a local instruction yet lack the cross-stage knowledge needed to finish a partially observed task. In the drawer example, exploration must reveal the target before placement can be planned. The authors call this mismatch the prediction–deliberation gap. e-problem
2.2 What this reading follows
A robot may know how to open a drawer and still fail to place a cookie in the correct one. The target can become identifiable only after exploration, then disappear from view before placement. HarnessWAM makes this cross-stage reasoning explicit: a VLM maintains evidence and a task graph, a compiler translates the graph into supported skills, and a runtime checks execution before advancing. The underlying LingBot-VA controller keeps its architecture. Read the results through that division of responsibilities: the clearest controlled evidence concerns orchestration of a shared executor, while the published comparisons span different systems. e-probleme-architecturee-statee-projectione-protocole-main
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The separate Qwen Task Manager and LingBot-VA executor support Dual-system; persistent state supports Memory & long-horizon modeling. Other mechanisms and Outside quadrants fit the contribution’s external compilation and verification loop: no new joint future/action predictor or inverse-dynamics decoder is introduced. Policy post-training is supported only by benchmark-specific fine-tuning; no WM-RL objective or reinforcement-learning procedure is established. e-architecturee-statee-projectione-implementation
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Keep the target unresolved until observation earns a binding
Consider the cookie task before any drawer is opened. A useful plan can specify that a target drawer must eventually be selected without pretending its identity is already known. HarnessWAM represents that target as an unresolved variable in a graph containing both motor and cognitive nodes. Opening a drawer changes what can be observed; a subsequent cognitive update can record its contents and bind a variable when the evidence is sufficient. The scene-fact representation distinguishes an unknown fact from an observed negative, and links claims to RGB evidence. Closing the drawer therefore need not erase the acquired information. In Figure 3, the later return to the top drawer is the source’s concrete example. Reader interpretation: the graph provides a place for information acquisition and delayed decisions to influence control explicitly. e-statee-rollout
5.2 Treat the plan as a constrained interface to learned skills
The VLM can describe operations more freely than the WAM can execute them. Section 3.3 handles this mismatch by restricting plans to an empirically validated primitive set and checking feasibility in the current runtime state. PLACE, for example, requires acquisition of the same object; a single arm’s holding state limits what can happen next. A supported semantic operation may expand into several primitives, while an infeasible graph returns violated constraints for replanning. Revisions preserve executed nodes and change the future suffix. Equation (12) summarizes these checks as a projection into an intersection of supported plan language and state-dependent feasibility. Table 4 tests a key distinction: normalization already improves executability, but full projection improves it further. Reader interpretation: this is evidence for constrained composition beyond prompt vocabulary cleanup, not proof that every compiler check is individually necessary. e-projectione-plan
5.3 Use progress as evidence, then verify the physical outcome
The progress estimator is trained to recognize local temporal advancement from five timesteps of dual-view RGB and a skill instruction. Its six supervision terms shape progress and completion predictions, while the inference runtime converts their history into candidate events. Algorithm 1 then gives the Task Manager responsibility for the actual transition decision. An uncertain result can request another observation; a failed strategy can trigger replanning; a local deviation can trigger recovery toward the saved arm and gripper state. Recovery clears local WAM state but retains scene knowledge, so subsequent attempts can use earlier evidence. Success still requires final visual verification of the global goal. Reader interpretation: Table 3’s progress-only sequence scores illustrate why partial completion and correct task termination must be evaluated separately, even when the local progress signal looks strong. e-progresse-implementatione-decisionse-recoverye-ablatione-metrics
5.4 Training and inference
During training
LingBot-VA is fine-tuned separately on each benchmark’s training data with its architecture unchanged; controlled variants share that checkpoint. Qwen3-VL-32B-Instruct serves as Task Manager without task-specific fine-tuning. RoboMemArena WAM RGB inputs are 256×256. Task Manager visual inputs exclude depth, segmentation, and privileged simulator state. e-implementation
The progress model uses five timesteps of dual-view RGB and skill text. Frozen SigLIP2-base-patch16-256 features feed a four-layer causal Transformer. Supervision combines regression, interval classification, ranking, endpoint anchoring, completion prediction, and local monotonicity. Episodes separate training/validation; checkpoint selection minimizes validation progress error. e-progresse-implementation
During inference
Algorithm 1 initializes belief and memory, plans and projects a graph, then selects ready nodes. Cognitive nodes acquire evidence and reproject future work. Motor nodes execute chunks until an event decision changes execution. Success requires all required nodes plus final visual goal verification; infeasibility, termination, or budget exhaustion can yield failure. e-decisionse-recovery
5.5 Implementation flow
- Maintain evidence and delayed bindings
Scene facts store subject, predicate, object, value, epistemic status, confidence, and supporting RGB evidence. Observed, inferred, and unknown facts remain distinct. Memory retains attempts, variable bindings, and revisions. A directed graph combines motor nodes with cognitive operations such as observing and verifying; unresolved targets remain symbolic until evidence supports binding. e-state
- Compile within the executor’s capabilities
The deterministic compiler checks types, dependencies, preconditions and effects, single-arm holding state, and acyclicity. It expands supported compositions and canonicalizes nodes into local WAM instructions. PLACE requires prior acquisition of the same object. Infeasible plans return a failure symbol and violated constraints for replanning. Only the unexecuted graph suffix may change. e-projection
- Execute and detect candidate events
LingBot-VA generates action chunks from observations, robot state, and the active skill prompt. A separate estimator predicts continuous progress, completion likelihood, and progress-bin probabilities. Milestones, exhausted skill budgets, or changed conditions trigger deliberation; estimator output alone cannot complete a motor node. e-architecturee-progress
- Verify, adapt, and recover
The Task Manager combines RGB, progress history, belief, graph, evidence, and execution state to continue, advance, observe, replan, recover, or terminate. Recovery drives joints and gripper toward their saved subtask-initial state, clears local WAM state, retains scene knowledge, and reprojects the remaining plan. It does not restore the environment. e-decisionse-recovery
6. Experiments & results
HarnessWAM adds persistent scene evidence, constrained skill planning, event-triggered verification, and local recovery around LingBot-VA. Its matched-checkpoint comparisons support improved composition of manipulation skills; the strongest measured ablation concerns the interface that compiles semantic plans into executable operations.
6.1 Read the original evidence
Table 1. The average gain is strongest evidence when the underlying WAM checkpoint is shared. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each task-family heading as a pair: SRtask requires every stage predicate to hold, while SRsub averages the fraction completed within each task. The far-right columns summarize the benchmark. Keep the two row groups distinct: the upper block contains published baseline scores, while Whole Task, Static Plan, and HarnessWAM share the LingBot-VA checkpoint in the controlled comparison. Section 4.1 specifies 20 rollouts per task with matched initial states, seeds, observation interfaces, and execution budgets for these diagnostics. Compare the bottom three rows first, then examine individual families to see where aggregate gains conceal weaker outcomes. e-maine-metricse-protocol
What it supports. HarnessWAM reaches 59.6% full-task and 69.9% subtask success, exceeding Static Plan’s 47.9% and 62.0% by 11.7 and 7.9 percentage points. This supports a benefit beyond an initial decomposition. The family columns show that the improvement is uneven across task types and success definitions.
Where the evidence stops. Counting subtask success is lower than Static Plan’s 90.1%, at 88.2%. PrediMem also leads the transfer columns. Published methods are contextual comparisons, and the table supplies no uncertainty estimates for any score.
Table 2. RoboCerebra Ideal tests extended composition under a different success definition. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the upward arrow as higher-is-better and the entries as percentages. The compact heading says Ideal SR, whose meaning is specified by Equation (20) on page 9: average completion of key object-state transitions within tasks, then across tasks. It therefore records partial progress. The paper describes the Ideal subset as static and fully observable, so these scores concern compositional execution when the initial scene is available. The comparator rows name different controller/planner systems, rather than the matched LingBot-VA variants of Table 1. That distinction matters when attributing the small numerical lead to a specific component. e-rce-metricse-protocol
What it supports. HarnessWAM reports 23.70%, compared with 21.92% for GPT-4o Planner + OpenVLA and 21.10% for HPE Framework. The lead is 1.78 percentage points over the strongest listed comparator. The metric supports a modest reported improvement on Ideal key-transition completion, with substantial room remaining.
Where the evidence stops. Do not interpret 23.70% as the fraction of tasks completing every stage, or compare it directly with RoboMemArena’s 59.6% full-task score. Different underlying systems and absent uncertainty estimates limit causal and statistical conclusions.
Figure 3. The target discovered during exploration determines a later manipulation sequence. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read left to right across the top row, then continue at the lower left. The robot opens and closes the top drawer, then the middle drawer, then the bottom drawer. In the last three frames it reopens the top drawer, picks the cookie, and places it inside. The caption identifies the top drawer as the non-empty target in this Task 4 rollout. Connect the earlier observation to the later reopening: Section 3.2’s scene belief and delayed variable binding explain how an earlier visible fact can remain useful after closure removes it from the current RGB view. e-rolloute-statee-problem
What it supports. This selected sequence illustrates the intended division between information-gathering actions and goal-directed manipulation. A drawer is inspected before the final target-dependent placement is executed. The accompanying source text attributes that continuity to retained evidence and target binding, making the example concrete for the abstract task-state representation.
Where the evidence stops. The frames are selected from one representative rollout. They do not expose internal memory records, verify each belief update, or quantify failure recovery. Figure 1 uses a different schematic target label; this rollout specifically returns to the top drawer.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboMemArena overall manipulation success 26 tasks; HarnessWAM and same-WAM diagnostics use 20 rollouts per task with matched initial states, seeds, interfaces, and task budgets; macro-averaged across tasks and rollouts. | 59.6 / 69.9 Full-task / subtask success (%) | Whole Task: 44.4 / 52.3; Static Plan: 47.9 / 62.0; published PrediMem: 38.5 / 55.2. The gain over Static Plan is 11.7 / 7.9 percentage points. Full-task success requires all stage predicates; subtask success averages the completed fraction. Published baselines provide context with weaker experimental control. e-benchmarkse-metricse-protocole-main |
| RoboMemArena family-specific tradeoffs Table 1 task-family breakdown under the same-WAM protocol. | Occlusion: 55.0 / 63.9; counting: 73.6 / 88.2; transfer: 21.3 / 31.5. Full-task / subtask success (%) | Static Plan: occlusion 38.2 / 49.8 and counting 67.2 / 90.1; published PrediMem transfer 22.5 / 45.2. Improvement is uneven: counting subtask success falls 1.9 points versus Static Plan, and transfer remains weak. e-maine-protocol |
| RoboCerebra Ideal key-transition completion Static, fully observable Ideal subset; official key object-state transition metric; published comparator scores. | 23.70 Ideal SR (%) | GPT-4o Planner + OpenVLA: 21.92; HPE Framework: 21.10. The 1.78-point lead over the strongest listed comparator concerns mean transition completion. It is not an all-stages full-task success rate and cannot be directly equated with RoboMemArena’s 59.6%. e-metricse-protocole-rc |
| RoboMemArena component removal Matched LingBot-VA weights, initial states, and evaluation seeds; Table 3. | Full system 59.6 / 69.9; without projection 18.5 / 38.3; without progress events 38.3 / 50.7; without task state 47.7 / 61.1; progress-only switching 55.4 / 68.4; without recovery 54.2 / 67.7. Average full-task / subtask success (%) | Projection removal costs 41.1 / 31.6 percentage points, the largest component-removal loss. This intervention jointly removes multiple checks. Fixed-interval deliberation changes event selection; equal VLM-call cost is not established. e-ablatione-protocol |
| Offline plan-quality diagnosis Raw, lexically normalized, and fully projected graphs; reference decompositions used only in offline analysis. | Fully projected: 72.9 / 92.9 Executability / dependency satisfaction (%) | Raw: 13.8 / 58.1; normalized: 42.3 / 67.5. Projection adds 30.6 executability points beyond normalization. These are intermediate plan diagnostics; sample size and uncertainty are unspecified. e-plan |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Projection, event selection, state, and recovery contribute differently to task completion. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the full HarnessWAM row as the reference, then scan the average task/subtask pair on the right. The intervention definitions span pages 10–11: removing task state drops scene facts, history, and bindings; removing projection bypasses multiple validity checks; removing progress events substitutes fixed-interval VLM invocation. Progress-only switching instead retains the estimator but lets it decide completion without semantic verification. Removing recovery ends execution after detected failure or budget exhaustion. These controls ask different questions. Once the averages are clear, inspect the sequence columns to see how local completion can diverge from full-task reliability. e-ablatione-protocole-metrics
What it supports. Without projection, average full-task success falls from 59.6% to 18.5%. Progress-only switching illustrates another failure mode: sequence subtask success rises from 93.0% to 97.9%, while sequence full-task success drops from 86.3% to 72.5%. High partial completion therefore does not establish reliable end-to-end composition.
Where the evidence stops. Projection removal bundles capability, argument, dependency, precondition, and holding-state checks, so it does not isolate their individual effects. Fixed-interval invocation is not reported at matched VLM-call cost; the table alone cannot establish an efficiency advantage.
Table 4. Canonical wording helps, but the full compiler adds a larger layer of validity checking. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the rows as progressively stronger processing of a VLM-generated graph. Raw plans are evaluated first; normalization maps operator expressions and safe entity aliases into canonical WAM vocabulary; full projection then applies capability, dependency, binding, and embodiment constraints. Each column measures a distinct plan property, with percentages increasing upward in quality. Compare the middle and bottom rows to separate lexical alignment from the full constrained interface. Section 4.3 says reference decompositions are used for this offline node/dependency analysis only. They are not additional instructions supplied to HarnessWAM while it executes benchmark tasks. e-plane-projection
What it supports. Executability increases from 13.8% for raw plans to 42.3% after normalization, then to 72.9% with projection. The last step also raises dependency satisfaction from 67.5% to 92.9%. These diagnostics support the mechanism that canonical language alone leaves substantial structural and embodiment-related plan errors unresolved.
Where the evidence stops. Executability here is an intermediate diagnostic, not observed task success. The paper does not report the analysis sample size or enough scoring detail to reconstruct every judgment, and the projected plans still fall short of universal validity.
7. Analysis & limitations
7.1 What the evidence leaves open
The conclusion leaves real-world manipulation, broader validated skill repertoires, and calibrated uncertainty-aware deliberation to future work. Benchmark evidence does not establish physical deployment. e-limits
Tables 1–4 give no confidence intervals or dispersion. Hardware, latency, and VLM-call counts are unreported, limiting significance and efficiency conclusions. Recovery restores the robot toward a saved configuration but cannot undo environmental changes. e-maine-rce-ablatione-plane-implementatione-recovery
7.2 Questions for discussion
- Does executable projection retain its advantage when prompt vocabulary, VLM calls, and execution budgets are matched?
- Can event-triggered verification improve full-task success at the same deliberation cost, and how often does recovery encounter irreversible scene changes?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires benchmark-specific LingBot-VA checkpoints, the named VLM and progress encoder, episode-separated progress training, a validated primitive inventory, graph compiler, event logic, and joint/gripper recovery control. Preserve matched initial states and budgets. e-implementatione-projectione-progresse-recoverye-protocol
Optimizer schedules, dataset sizes/split ratios, loss weights, exact prompts, complete validated skill inventory, numerical event thresholds, budget settings, hardware, and software versions are unspecified. These remain implementation requirements; the illustrated checks are proposed experiments. e-projectione-progresse-recoverye-implementatione-protocol
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Separate lexical normalization from constraint enforcement
Reader-proposed, not run: use the same benchmark-trained LingBot-VA checkpoint, Task Manager outputs, initial states, and execution budgets for two arms: canonical normalization/aliasing alone versus full executable-space projection. Preserve the same proposed graph before compilation, and prevent offline reference decompositions from entering inference. Use the paper’s 20-rollout-per-task protocol for online evaluation. Measure offline dependency satisfaction and executability, then full-task and subtask success; log which compiler checks reject or expand nodes. The prediction is that full projection improves both plan validity and full-task success beyond lexical normalization. Equal task success despite better offline validity would weaken the claim that the measured plan diagnostic explains the execution gain. e-projectione-protocole-maine-ablatione-plan
Check 2: Compare event timing at a matched deliberation budget
Reader-proposed, not run: hold the executor, progress estimator, task memory, compiler, semantic verifier, and recovery logic fixed. Compare progress-triggered Task Manager invocation with a fixed-interval schedule, giving both conditions the same per-episode VLM-call cap and environment-step budget. Select the interval on held-out validation episodes and report actual calls, latency, full-task/subtask success, premature advances, and missed failure detections on matched evaluation starts. Sweep the shared call cap to compare success at similar realized cost. The event-selection mechanism predicts fewer mistimed transitions and higher full-task success at comparable cost. If the advantage disappears after matching calls, Table 3’s fixed-interval loss would not by itself establish better timing efficiency. e-progresse-decisionse-protocole-ablation
8.3 Reading coverage
Visual audit: Visually inspected the title/authors/version on page 1; Figure 1 on page 2; Figure 2 and local-control equations on page 4; state, compiler, progress, Algorithm 1, and recovery definitions on pages 5–8; training, evaluation, and metric definitions on page 9; Tables 1–4 and their surrounding protocol/diagnostic text on pages 10–11; and Figure 3 plus the conclusion on page 12. All six final original crops were viewed and retain complete graphic/table content and labels. Figure 2’s action endpoint was cross-checked against Equation (1), and event flow against Algorithm 1; the endpoint discrepancy is disclosed. All five text chunks, including references on pages 12–15, were read. Page 3 and reference-only pages 13–15 were not visually inspected; no retained technical claim depends on their layout. No appendix is present, and separate supplements remain unverified.
PDF pages inspected for this edition: 1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 12. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; 1 Introduction (pp. 1–3)
- 2 Related Work, including 2.1–2.2 (p. 3)
- 3 Method, including 3.1–3.6, Equations (1)–(17), and Algorithm 1 (pp. 4–8)
- 4 Experiments, including 4.1–4.4 and Tables 1–4 (pp. 8–12)
- 5 Conclusion (p. 12)
- References (pp. 12–15)
Outside the original text pass
- Identity: the inspected title page identifies arXiv:2608.09516v1 [cs.RO], 10 August 2026. Its title and all eleven authors match the catalog; no revision or edition substitution is involved.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction limitation was addressed by visual inspection of Figures 1–3, Tables 1–4, and all pages supporting retained technical claims.
- Separate supplemental material availability has not been fully verified.
- No appendix is present in the supplied 15-page PDF. All five text chunks were read; references were read as supplied, without opening the cited works.
- Code and external resources were not inspected; no training, inference, or reproduction experiment was run.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author/affiliation block, and arXiv margin identifier
HarnessWAM: Bridging Prediction and Deliberation in World Action Models; eleven listed authors; arXiv:2608.09516v1 [cs.RO], 10 August 2026.
Go to primary source ↓e-problemPDF pp. 1–2, Abstract, Section 1, and Figure 1/caption
The prediction–deliberation gap concerns local WAM control versus persistent planning, evidence retention, verification, and recovery; Figure 1 illustrates sequential drawer exploration.
Go to primary source ↓e-architecturePDF p. 4, Figure 2/caption and Section 3.1, Equations (1)–(3)
The architecture connects an external Task Manager, task state, deterministic compiler, WAM, runtime, and environment. Equation (1) ends an H-action chunk at a(t+H−1); Figure 2 labels its endpoint a(t+H). Runtime state contains belief, graph, memory, and active-skill execution state.
Go to primary source ↓e-statePDF p. 5, Section 3.2, Equations (4)–(8)
Evidence-linked facts distinguish observed, inferred, and unknown state; memory stores attempts and revisions; motor/cognitive graph nodes carry preconditions, effects, termination, recovery, and delayed entity bindings.
Go to primary source ↓e-projectionPDF pp. 5–6, Section 3.3, Equations (9)–(13)
The extensible ontology is restricted to empirically validated WAM primitives. Deterministic compilation enforces types, dependencies, holding-state and precondition/effect consistency, and acyclicity; infeasible graphs return a failure symbol; executed history remains invariant.
Go to primary source ↓e-progressPDF p. 6, Section 3.4, Equations (14)–(15)
Progress, completion probability, and progress intervals are estimated from recent dual-view RGB and skill text. Six supervision terms train the estimator; milestones, budgets, and changed conditions trigger deliberation without directly advancing nodes.
Go to primary source ↓e-decisionsPDF p. 7, Algorithm 1, lines 1–38, and Equation (16)
Inference initializes and projects a graph; handles cognitive and motor nodes; executes environment actions; event decisions include continue, advance, observe, replan, recover, and terminate. Required nodes plus final goal verification govern success.
Go to primary source ↓e-recoveryPDF p. 8, Sections 3.5–3.6, Equation (17)
Recovery targets saved arm/gripper state with multi-step control, clears local WAM state, retains environment and memory, and reprojects subsequent plans. Task execution and recovery budgets bound termination.
Go to primary source ↓e-benchmarksPDF p. 8, Section 4.1, Benchmarks
RoboMemArena has 26 tasks across transfer, occlusion, counting, and sequence; RoboCerebra Ideal evaluates long-horizon compositional manipulation.
Go to primary source ↓e-metricsPDF p. 9, Section 4.1, Metrics, Equations (18)–(20)
RoboMemArena full-task success requires all stage predicates; subtask success averages the within-task completed fraction. RoboCerebra SR averages key object-state transition completion.
Go to primary source ↓e-protocolPDF p. 9, Section 4.1, Evaluation protocol and Baselines
HarnessWAM and same-WAM diagnostics use 20 rollouts per task with matched initial states, seeds, observation interfaces, and budgets. Published scores come from benchmark evaluations. Whole Task lacks decomposition; Static Plan fixes its initial linear sequence.
Go to primary source ↓e-implementationPDF p. 9, Section 4.1, Models and implementation details
Benchmark-specific LingBot-VA fine-tuning preserves architecture. Qwen3-VL-32B-Instruct is not task-fine-tuned. The progress estimator uses five timesteps, frozen SigLIP2-base-patch16-256, and a four-layer causal Transformer; episodes separate training and validation.
Go to primary source ↓e-mainPDF p. 10, Table 1, all rows and task-family/average SR columns; Section 4.2
HarnessWAM average task/subtask scores are 59.6/69.9, versus Static Plan 47.9/62.0 and Whole Task 44.4/52.3. PrediMem averages 38.5/55.2 and leads transfer at 22.5/45.2. HarnessWAM counting subtask score 88.2 trails Static Plan 90.1.
Go to primary source ↓e-rcPDF p. 10, Table 2, Ideal SR column, and Section 4.2, RoboCerebra paragraph
Ideal SR is 23.70 for HarnessWAM, 21.92 for GPT-4o Planner + OpenVLA, 21.10 for HPE, 7.84 for OpenVLA, and 1.88 for π0.5. Ideal is described as static and fully observable.
Go to primary source ↓e-ablationPDF pp. 10–11, Section 4.3 intervention definitions; p. 11, Table 3, all rows and columns
Average task/subtask scores: full 59.6/69.9; no task state 47.7/61.1; no projection 18.5/38.3; no progress events 38.3/50.7; progress-only 55.4/68.4; no recovery 54.2/67.7. Sequential progress-only switching yields 72.5/97.9 versus full 86.3/93.0. Fixed-interval invocation replaces progress events; no recovery terminates on failure or budget exhaustion.
Go to primary source ↓e-planPDF p. 11, Table 4, all rows/columns, and Section 4.3, Plan-level diagnosis
Raw/normalized/projected syntax: 60.8/84.6/95.2; dependencies: 58.1/67.5/92.9; binding: 21.3/63.8/88.3; executability: 13.8/42.3/72.9. Reference decompositions support offline analysis only, never inference. Sample counts and uncertainty are not reported here.
Go to primary source ↓e-rolloutPDF p. 12, Figure 3/caption and Section 4.4
Ten selected Task 4 keyframes show exploration of top, middle, and bottom drawers, followed by reopening the non-empty top drawer and picking/placing the cookie. The caption attributes target retention to stored evidence.
Go to primary source ↓e-limitsPDF p. 12, Section 5, final sentence
Real-world manipulation, broader WAM skill repertoires, and calibrated uncertainty-aware deliberation are identified as future directions.
Go to primary source ↓8.5 Primary sources
HarnessWAM: Bridging Prediction and Deliberation in World Action Models ↗
PDF · 7,585 extracted words
Source fingerprint
e123b7056eb7011730c4ce5fc289f3df7a74b187228cdb5ba029daf46dddcad9