Source-linked AI summary
Preemptive Instruction Following with Resumption
Zerun Liu, Lechao Cheng, Yimin Liu, Shengeng Tang, Yaxiong Wang, Zhun Zhong
TL;DR
Long-horizon embodied navigation lacks systematic evaluation of how agents handle an inserted task and recover an unfinished sequence from the resulting environment state. PIFR evaluates interruption and resumption across several interruption structures, and PIFR-Rank provides a candidate-ranking baseline; results vary across settings, and aggregate success can obscure recovery differences.
Problem
Existing embodied-navigation benchmarks do not systematically evaluate executing an inserted task and resuming an unfinished sequence from the resulting state.
Source: §1 · p. 2, §2.1 · p. 3Method
PIFR evaluates single, repeated, and nested interruptions with stage-specific measures, alongside PIFR-Rank, a candidate-ranking baseline for re-grounding after interruption.
Source: §1 · p. 2Results
Takeaways & Limitations
PIFR provides a testbed for studying task recovery in long-horizon embodied navigation.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Long-horizon embodied navigation may be interrupted by a higher-priority instruction before the current task sequence is complete. The agent must finish its current atomic task, execute the inserted task, and resume the remaining sequence from the updated environment state, where its viewpoint and available navigation candidates may have changed. To evaluate this capability, we introduce PIFR benchmark, a novel task for runtime task preemption and resumption in embodied navigation. PIFR includes single, repeated, and nested interruption settings,with stage-specific metrics for inserted-task completion, resumption of the original task, and completion of the full sequence. Evaluations show that recovery performance varies across interruption conditions, while aggregate success can obscure differences in interruption handling and resumption. We further propose a lightweight candidate-ranking baseline to improve target re-grounding after interruption. Together, PIFR and the baseline provide a systematic framework for studying interruption handling and task recovery in long-horizon embodied navigation.
1 INTRODUCTION
PIFR addresses the lack of systematic evaluation of runtime task preemption and resumption in embodied navigation. It evaluates recovery across interruption settings and introduces PIFR-Rank as a candidate-ranking baseline for re-grounding after interruption.
1 INTRODUCTION
- Existing embodied-navigation benchmarks do not systematically evaluate completing an atomic task, executing an inserted task, and resuming the unfinished sequence from the resulting state.Aggregate navigation success alone cannot distinguish inserted-task handling, resumption, and full-sequence completion.Source: §1 · p. 1
- After an inserted task, the agent resumes the unfinished base sequence at its next atomic task and re-grounds the objective from the updated environment state.The agent need not return to its pre-interruption position or reuse its original route.Source: §1 · p. 2
- Evaluations show recovery performance varies with interruption conditions and goal specifications, while aggregate success can mask differences in handling inserted tasks and resuming suspended ones.PIFR-Rank is proposed as a lightweight candidate-ranking baseline for target re-grounding after interruption.Source: §1 · p. 2
- PIFR evaluates single, repeated, and nested interruptions across diverse goal specifications.Its curated datasets support these different interruption structures.Source: §1 · p. 2
- The benchmark distinguishes inserted-task completion, base-task resumption, and full-sequence completion through stage-specific evaluation measures.The continuous-execution protocol evaluates performance under different interruption conditions.Source: §1 · p. 2
- PIFR-Rank uses candidate-level supervision to improve target re-grounding after interruption while preserving the original navigation pipeline.It is presented as a lightweight baseline.Source: §1 · p. 2
2 RELATED WORK
Prior embodied-navigation work broadens task types, goal specifications, and long-horizon execution, while interaction research studies instruction-driven behavior. The related work identifies unfinished-task recovery after an intervening task as a distinct evaluation setting.
2.1 EMBODIED NAVIGATION AND LONG-HORIZON TASK EXECUTION
- Embodied-navigation research expands supported task types and goal modalities through unified policies, language interfaces, and visual grounding with exploration and spatial memory.These approaches broaden the navigation tasks agents can handle.Source: §2.1 · pp. 2–3
- Multi-goal and long-horizon benchmarks evaluate sequential targets or predefined task sequences, but not systematic resumption after an inserted task changes the environmental state.The section contrasts this gap with PIFR’s interruption-and-recovery setting.Source: §2.1 · p. 3
2.2 INTERRUPTIONS IN EMBODIED TASK EXECUTION
- Instruction-driven interaction and dialogue studies do not evaluate recovery of an unfinished navigation task after an intervening task is executed.InterruptBench studies web-agent request changes, whereas PIFR considers embodied navigation and resumption from the resulting environmental state.Source: §2.2 · p. 3
3 PIFR BENCHMARK
PIFR formalizes interruption and resumption in persistent embodied navigation, where an unfinished task sequence is suspended for an inserted task and then resumed. Its benchmark construction covers three interruption structures and multiple goal specifications.
3.1 TASK DEFINITION
- PIFR preserves logical task progress across interruptions, resuming the next unfinished task from the updated environment rather than restoring the prior position or route.The next task instruction is reissued and must be grounded in the current physical and perceptual context.Source: §3.1 · p. 4
3.2 CANONICAL INSTANTIATIONS
- Single interruption inserts one sequence at one boundary; Repeated inserts sequences at distinct base-sequence boundaries; Nested interrupts an inserted sequence and resumes in last-in, first-out order.The three structures differ in how inserted episodes are embedded in the base sequence.Source: §3.2 · p. 4
3.3 DATASET CURATION
- PIFR composes tasks from SG3D, GOAT-Bench, and HM3D-OVON, retaining their respective description-based, category/language/image, and open-vocabulary goal specifications.Construction samples valid interruption boundaries, checks geometric navigability across the full execution chain, and balances events across interruption structures and scenes.Source: §3.3 · pp. 4–5
- Unlike benchmarks that generally follow predefined objective order, PIFR supports recursive interruption and resumption, including recovery when an inserted sequence is itself interrupted.The Nested setting evaluates this recursively embedded interruption case.Source: §3.3 · pp. 5–6
4 PIFR-RANK
PIFR-Rank trains a grounding scorer to re-ground unfinished tasks after an interruption, using candidate-level supervision and a ranking loss that gives greater weight to resumption decisions.
4 PIFR-RANK
- PIFR-Rank targets re-grounding the next unfinished task from the changed environment after an inserted sequence completes.It resumes the most recently suspended sequence rather than requiring a return to the prior position or route.Source: §4 · p. 6, §4.1 · p. 6
4.1 CANDIDATE CONSTRUCTION AND TRAINING OBJECTIVE
- Candidates combine observed-region objects and unexplored-region frontiers, scored against the active instruction.The scorer retains the navigation system’s candidate-generation and online procedures.Source: §4.1 · pp. 6–7, Figure 3 · p. 6
- Training uses candidate-validity labels and combines padding-aware binary cross-entropy with a ranking loss to favor target-consistent candidates over distractors.Multiple matching object candidates may be positive; training-only target and scene information can label a frontier when the target is unobserved.Source: §4.1 · p. 6, Figure 3 · p. 6
- The ranking loss compares the highest-scoring positive and negative with a margin, and is zero when no valid negative exists.Decisions without a valid positive are excluded from both candidate-level losses.Source: §4.1 · p. 6
- Resumption decisions receive greater training weight, including resumptions in repeated and nested events.Execution-role labels affect training weights but are unavailable to the online policy.Source: §4.1 · pp. 6–7
5 EXPERIMENTS
The experiments evaluate navigation baselines with overall and stage-specific metrics across interruption structures, goal splits, and benchmarks. Results show that performance depends on these conditions, and PIFR-Rank’s ablations trade broad gains against stronger stage-specific metrics.
5 EXPERIMENTS
- SR and SPL measure success and path efficiency over completed atomic segments; Interrupt SR and Resume SR measure success on interruption-entry and resumption-entry segments.An atomic segment is successful when the agent terminates within 0.25 m geodesic distance of a valid target viewpoint.Source: §5 · p. 7
5.1 BENCHMARKING RESULTS WITH NAVIGATION BASELINES
- Under nested interruption on PIFR-GOAT (Seen), PIFR-Rank improves SR from 44.3% to 49.6%; its SR also exceeds MTU3D across all three structures on PIFR-SG3D and PIFR-OVON.On PIFR-GOAT, MTU3D performs better under single interruption, while PIFR-Rank has higher SR under repeated and nested interruption.Source: §5.1 · p. 7, Table 3 · p. 8
- On PIFR-OVON under single interruption, PIFR-Rank has higher SR than MTU3D (51.7% vs. 49.6%) but lower Resume SR (49.0% vs. 56.1%).This illustrates that overall success and recovery performance do not always change together.Source: §5.1 · p. 8
- On PIFR-OVON’s Unseen split, PIFR-Rank has higher SR than MTU3D under all three interruption structures; results vary across splits and structures on PIFR-GOAT.The passage cautions that comparisons should be made within each benchmark and split protocol because Synonym and Unseen splits are independently defined.Source: §5.1 · p. 8, Figure 4 · p. 8
- Repeated and nested recovery profiles vary by benchmark and resumption position, so later or nested recovery does not produce a uniform decline.For repeated interruption, PIFR-Rank’s Resume SR changes from 32.5% to 17.5% on PIFR-SG3D, 53.8% to 38.7% on PIFR-OVON, and 44.0% to 66.0% on PIFR-GOAT; Figure 5 separates first/second and inner/outer resumptions.Source: Figure 5 · p. 9, §5.1 · p. 9
5.2 ABLATION STUDY OF PIFR-RANK
- Candidate-level BCE supervision improves all four baseline metrics; adding candidate ranking or resumption weighting shifts the profile toward stronger stage-specific performance, with resumption weighting yielding the highest Interrupt SR and Resume SR.Table 4 reports BCE-only SR 39.82%, SPL 23.32%, Interrupt SR 33.46%, and Resume SR 44.91%; resumption weighting reaches Interrupt SR 35.25% and Resume SR 45.04%.Source: Table 4 · p. 9, §5.2 · p. 9
6 CONCLUSION
The conclusion presents PIFR as a benchmark for interruption and resumption in embodied navigation. It reports that aggregate success can obscure recovery differences and introduces PIFR-Rank, whose gains vary across settings.
6 CONCLUSION
- PIFR evaluates interruption and resumption of unfinished task sequences in embodied navigation.Its single, repeated, and nested interruptions assess navigation success and recovery from the post-interruption state.Source: §6 · p. 9
- Aggregate success alone can obscure differences in interruption handling and resumption, which vary across execution structures..Source: §6 · p. 9
- PIFR-Rank is a candidate-ranking baseline that improves resumption-related metrics, though its gains vary across settings.PIFR is presented as a testbed for studying task recovery in long-horizon embodied navigation.Source: §6 · p. 9