The Door Opened. The Task Isn't Done.

TL;DR for operators

A robot that opens a door has made a route accessible, but it has not necessarily found the object it was sent to retrieve. In a 3,000-episode simulated benchmark, an agent that maintains interaction-state memory and verifies physical changes achieves 41.63% terminal success, compared with 10.53% for an adapted baseline. Replacing execution-informed state updates with perception-only updates produces a particularly large performance decline in a smaller component test.

For indoor-service robot developers, the engineering implication is to record what an action actually changed, revise navigation assumptions accordingly, and measure final task completion separately from intermediate actions. The results concern simulated navigation with restricted-ground-truth perception and a shared interaction executor, not demonstrated production-ready robotic manipulation.

Opening a Door Is Not the Same as Finding the Object

Consider an indoor-service robot asked to locate an apple behind a closed door and inside a refrigerator. Opening the door establishes a possible route into the next room. Opening the refrigerator can expose previously hidden objects. The robot still needs to observe the apple and reach a position from which the task’s success conditions are satisfied.

This distinction is central to Interactive Navigation Beyond Reachability: Benchmark and Agentic Framework, which introduces MultINav-Bench and MultINav-Agent.¹

The paper’s Channel experiments reveal a substantial difference between performing required interactions and completing navigation. MultINav-Agent achieves a 60.09% Interaction Success Rate (ISR), compared with 54.28% for an adapted MoMa-LLM baseline. Yet their terminal success rates are 55.80% and 19.60%, respectively (Section 5.2, Table 3).

The interaction-effect gap is only 5.81 percentage points; the terminal-success gap is 36.20 points.

Here, ObjectGoal navigation requires reaching and visually identifying a requested object, with success defined by both distance and head-camera visibility. ISR instead measures the proportion of required physical effects successfully verified. It can award partial credit when some interactions succeed but the overall task remains unfinished.

The distinction becomes particularly consequential when an agent must perform several dependent actions. On Mixed tasks, MultINav-Agent records 45.07% ISR but only 24.10% terminal success. Just 247 of 1,000 Mixed episodes complete all required interaction effects, and 223 of those subsequently reach the goal.

A system can therefore make measurable progress through its required actions without reliably completing the requested job.

A Benchmark That Tests Whether Interaction Changes What Is Possible

MultINav-Bench addresses an evaluation problem that ordinary navigation tests leave largely untouched: the robot may need to change its environment before the destination becomes reachable or observable.

Its frozen ObjectGoal-light benchmark contains 3,000 episodes across 639 simulated houses, divided equally among three domains:

• Channel: Opening a door changes which rooms are connected by a traversable route.
• Container: Opening a cabinet, drawer, or refrigerator can reveal previously concealed objects.
• Mixed: Sequential interactions combine access changes with target exposure.

The benchmark also distinguishes necessary from unnecessary interactions. Of its 3,000 episodes, 2,654 require interaction and 346 do not. The unnecessary cases involve distractor doors in Channel tasks.

These labels come from benchmark construction using counterfactual reachability, visibility checks, articulation-state validation, and feasible reference plans (Section 3; Appendix B).

This design matters because attempting every available interaction can produce physical effects without improving goal completion. A robot that unnecessarily opens doors should not receive the same assessment as one that correctly determines no interaction is needed.

The evaluation consequently separates terminal success, path efficiency, verified required effects, target-class interaction precision, and normalized cost. These metrics answer different questions. In particular, interaction precision credits qualifying effects at the annotated interaction-class level; it does not necessarily demonstrate manipulation of the exact required object instance.

The Critical Decision Comes After the Physical Action

MultINav-Agent’s architecture addresses the problem of retaining and revising knowledge about a changing environment.

It maintains a persistent, state-conditioned interaction graph representing rooms, objects, containers, connectivity, containment, visibility, and estimated interaction states. Entries also carry information about confidence and observation recency.

A door that leaves the camera view remains in memory. Its recorded state determines whether the planner considers the connection between rooms traversable. Likewise, opening a container changes the opportunity to observe its contents, without establishing that a particular hidden object has been found.

The agent generates navigation, exploration, and interaction candidates. A vision-language model selects among admissible actions according to their anticipated relevance to the goal.

Crucially, execution does not automatically establish success.

After an attempted interaction, the system checks new observations and execution feedback. A requested state is recorded only when supported by verification. Conflicting or incomplete evidence can produce an uncertain state; a failed execution without observed change preserves the previous estimate.

The revised graph then informs the next decision (Section 4; Appendix D).

This mechanism prevents the planner from treating an intended physical change as an established fact. An unverified door opening should not automatically create a traversable connection in the robot’s internal representation.

What the component tests actually show

The full interactive benchmark favors MultINav-Agent over adapted MoMa-LLM: terminal success rises from 10.53% to 41.63%, ISR from 23.69% to 49.38%, and target-class interaction precision from 27.59% to 46.79%. Both systems use matched episodes, a shared high-level decision model, restricted-ground-truth observations, and the same physical interaction executor.

That comparison establishes an end-to-end difference between complete systems. The separate ablation study provides more focused evidence about architectural components.

Source: Section 5.3, Table 4. All variants use the same 200 Mixed episodes across 165 houses.

Replacing execution-informed graph updates with perception-only updates produces the largest observed success decline: 17.5 percentage points.

The other interventions tell a more qualified story. Flat Object Memory has lower interaction precision than Full but slightly higher ISR. Greedy Selection has slightly lower terminal success while achieving higher success-weighted path efficiency (0.134 versus 0.107) and lower normalized Cost (0.933 versus 0.961).

Neither relational memory nor semantic action selection dominates every metric. Differences of one or two percentage points in success correspond to only two or four episodes in this subset, and the paper reports no statistical uncertainty intervals.

The strongest component-level finding concerns execution-informed state updating. The experiments provide less decisive grounds for claiming that every additional planning component independently improves performance.

What Robotics Teams Should Measure Before Deployment

For developers of indoor-service and object-retrieval robots, the paper suggests a concrete architecture and evaluation decision.

A navigation system should distinguish whether an interaction was attempted, whether its intended physical effect was verified, whether that effect changed the available route or target visibility, and whether the final requested object was reached and observed.

Fleet dashboards can preserve these as separate operational states rather than reporting a successful door opening as completion of a retrieval task.

Reliability teams can also test whether uncertain interaction outcomes are handled conservatively. If a door remains closed or its state cannot be confirmed, subsequent route planning should not rely on an assumed opening.

These are engineering implications drawn from the demonstrated mechanism, not separately measured improvements in commercial robot fleets.

Cost evaluation needs similar care. MultINav-Agent’s overall normalized Cost is 0.7743, versus 0.9309 for the baseline. However, the benchmark assigns the maximum cost of one to failed episodes. Consequently, lower aggregate Cost partly reflects differences in success rates and does not independently establish lower resource consumption among successful tasks.

Simulation Gains Are Not Yet Manipulation Reliability

Three boundaries materially affect deployment interpretation.

First, the primary comparison uses restricted-ground-truth observations that provide identities, categories, and spatial information for currently visible entities. Hidden objects and oracle plans remain withheld, but the result does not establish equivalent performance using raw camera perception alone.

Second, both systems share a physical interaction executor that includes force-driven articulation and a kinematic completion fallback. Reliable autonomous low-level manipulation is therefore not demonstrated by their comparative scores.

Third, the physical Go2 quadruped demonstration illustrates navigation and graph updates across a door-and-refrigerator sequence, but both opening operations are human-assisted (Section 5.5).

Sequential reliability also remains limited in simulation: even the better-performing agent succeeds on only 24.10% of Mixed episodes. The benchmark’s target-category imbalance and repeated houses further constrain generalization.

The paper establishes a stronger testing framework and a promising state-management mechanism. It does not establish robust operation across unfamiliar physical environments.

Verify the Change Before Planning the Next Step

MultINav shifts interactive navigation evaluation toward a more complete account of task execution. The robot must determine which interactions are necessary, verify what changed after acting, preserve that information, and use it to complete the requested objective.

The most consequential result is not simply the large advantage over an adapted baseline. It is the combination of that advantage with a substantial ablation penalty when execution-informed updates are removed.

For robotics teams, that makes verified environmental state a defensible design priority. The broader analogy to software agents—updating workflow state only after checking execution outcomes—is plausible, but remains outside this paper’s empirical evaluation.

Ultimately, a completed action should change an agent’s plan only to the extent that its observed consequences justify doing so.

Cognaptus: Automate the Present, Incubate the Future.