acceptodds
Under review as a conference paper at ICLR 2027

AeroSOP: Bridging Task Intent and AAV Execution through Executable Procedures

Abstract

Large language models (LLMs) enable flexible mission planning for autonomous aerial vehicles (AAVs), but reliance on remote reasoning makes mission progress vulnerable to inference delays and communication disruptions. Action failures and changing observations can invalidate planned steps, making reliable continuation depend on execution feedback and updates consistent with the vehicle's current state. We introduce AeroSOP, a framework that decouples base-station reasoning from onboard execution through verifiable, incrementally updatable standard operating procedures (SOPs). The base station composes reusable skills into validated procedures, while an onboard runtime checks step contracts, verifies completion evidence, and maintains explicit execution state without task-level LLM inference. State-compatible patches introduce subsequent tasks at verified task-completion boundaries, preserving completed results and activating atomically after local compatibility and energy checks. Bounded waiting and return-feasibility monitoring regulate continuation under delayed updates and disturbances. Evaluation against eleven baselines across simulated urban missions shows that AeroSOP achieves 79.67% overall success and 97.00% inspection success, with no false completions or invalid proposals observed on inspection and search–confirmation tasks. Its patches achieve 91.0% useful continuation versus 90.0% for state-preserving full refresh, with 95th-percentile switching latency reduced from 6.8 to 4.7 s. Relative to remote skill-by-skill dispatch, AeroSOP reduces control traffic per mission by 75.5% and response latency at 24 AAVs by 79.7%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Large language models, Autonomous aerial vehicles, Standard operating procedures, State-aware execution, Incremental procedure updates.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_S7FFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AeroSOP, a framework that bridges high-level LLM-based mission reasoning and reliable onboard execution for autonomous aerial vehicles. AeroSOP represents missions as executable SOPs composed of pre-tested skills with explicit execution contracts, completion criteria, bounded retries, and failure handling. Task-level reasoning is performed at the base station, while the onboard runtime executes and verifies procedures locally without requiring LLM inference at every step. For continuous missions, AeroSOP introduces state-compatible SOP patches that preserve completed results and are activated at task boundaries after version, state, capability, and energy checks. Experiments in simulated urban environments compare AeroSOP with eleven baselines and show strong overall task completion, high inspection success, low execution errors, reduced switching latency, and lower communication and coordination overhead. Ablation studies further validate the contributions of SOP pre-testing, contract checking, mission verification, and local recovery.

Critical strengths
  1. Find the critical problem of LLM-AAV that the gap between LLM reasoning and execution physically.
  2. Full design of the whole system withcomplete ablation and experiments.
Critical weaknesses
  1. The proposed framework is comprehensive and well engineered, but many of its core ingredients—pre-tested procedural templates, execution contracts, runtime verification, bounded retries, state-machine/behavior-tree-style execution, version checks, atomic updates, and idempotency—are individually well established. The paper would benefit from a sharper characterization of what is fundamentally new beyond integrating these mechanisms into an AAV execution architecture, ideally supported by stronger formal guarantees or a clearer distinction from existing executable-program and behavior-tree frameworks.
  2. TypeFly—a particularly close prior system combining remote LLM reasoning with an edge runtime—is discussed in related work but not included as a baseline.
Questions for authors

see weakness

Reviewer_LkECOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AeroSOP, a framework for bridging natural-language mission intent and autonomous aerial vehicle execution through executable standard operating procedures. The main design is to separate task-level language reasoning at the base station from onboard execution. The base station constructs and validates SOPs using pre-tested templates and deployed skills. Each SOP contains explicit skill bindings, typed parameters, execution dependencies, preconditions, outcome checks, retry limits, failure responses, and mission-level completion criteria. The onboard runtime executes accepted SOPs, verifies intermediate and final outcomes, performs bounded recovery, and maintains local safety monitoring without requiring task-level LLM inference at every step. The paper also introduces a state-compatible next-task patch mechanism for continuous missions. A new task can be installed at a verified task-completion boundary while retaining completed results and execution state. Patches are checked against the current revision, task boundary, available capabilities, current vehicle state, and remaining energy before activation. The authors evaluate the system in simulated urban AAV missions and compare against eleven baselines. AeroSOP achieves competitive overall success, strong inspection performance, low observed execution errors, lower communication overhead, and favorable scaling behavior.

Critical strengths
  1. The paper addresses an important systems problem: how to keep physical-agent execution reliable without requiring task-level LLM inference at every step. The separation between high-level reasoning and local execution is sensible for communication-constrained robotic systems. Once a procedure is accepted, the onboard runtime can check preconditions, dispatch skills, verify results, handle bounded failures, and confirm completion without repeatedly consulting the LLM. This is a practically meaningful design choice.
  2. The state-compatible patch mechanism is carefully designed. The paper does more than simply regenerate an entire plan or append a new task. A patch is tied to a particular vehicle, procedure revision, task-completion boundary, and expiry condition. It is validated against the current local state and then activated atomically. The ablations suggest that state preservation, version checks, activation-time validation, and idempotency are important for avoiding duplicate execution, invalid activation, and state loss.
Critical weaknesses
  1. The framework appears to obtain much of its reliability by restricting execution to a pre-specified procedural space, but the paper does not characterize the size or coverage of this space. AeroSOP assumes that the necessary skills, SOP templates, verification predicates, and bounded failure responses are already available. The base-station agent therefore does not solve an unrestricted open-world planning problem. It retrieves and composes procedures from a predefined vocabulary, binds parameters, and validates the resulting candidate. This restriction may be exactly what makes the system reliable, and I do not consider the restriction itself a flaw. However, it is central to interpreting the contribution. The paper does not report the number of skills, SOP templates, verification predicates, recovery rules, or legal compositions available to the planner. More importantly, it does not estimate what fraction of realistic AAV missions can be represented within this procedural space. Therefore, it is difficult to tell whether AeroSOP provides a broadly general interface for autonomous AAV operation or primarily performs well on tasks that have already been anticipated by the designers of the skills, templates, contracts, and verifiers. The composition-OOD experiment only partially addresses this issue. Unseen compositions may still be assembled entirely from the same predefined procedural vocabulary. Generalization to a new combination of known components is substantially different from handling a genuinely novel capability, success criterion, failure mode, or procedural structure.
  2. The strongest contribution appears to be runtime and protocol engineering rather than a new learning or planning principle, and the current experiments do not establish sufficiently broad generality. Many individual ingredients of AeroSOP are well-established systems ideas: planner–executor separation, executable programs, preconditions and postconditions, bounded retries, state machines or behavior trees, version checks, idempotency, atomic updates, and energy-aware return policies. The main contribution is the integration of these ideas into a coherent execution and update protocol for continuous AAV missions. I find this useful, but I am less convinced that the current paper establishes a sufficiently general new principle for ICLR. The evaluation is entirely simulated, considers fleets of limited size, and focuses on a relatively structured family of inspection, search-confirmation, reporting, and compound missions. The current design also emphasizes adaptation between tasks at verified completion boundaries rather than arbitrary structural replanning during a partially executed task. This distinction is important. Many real-world failures cannot be solved through bounded retries or pre-specified local recovery. They require changing the procedure itself while execution is ongoing. The current paper does not demonstrate how AeroSOP behaves in this regime.
Questions for authors
  1. What is the effective task and program space covered by AeroSOP? Could the authors quantify the current procedural vocabulary, including the number of skills, SOP templates, verification predicates, failure responses, and valid task compositions? More importantly, can the authors estimate or empirically measure what fraction of realistic AAV missions can be expressed using this vocabulary? This would help determine whether the reported reliability comes from a broadly reusable execution abstraction or from operating within a heavily curated task space.
  2. What happens when a task falls outside the predefined SOP space? For example, what happens if a mission requires:
  • a skill that is not currently deployed;
  • a completion criterion for which no verifier exists;
  • a failure mode for which no approved recovery rule exists;
  • or a structural replan while the current task is only partially completed? Does AeroSOP reject the task, return control to the base-station LLM, generate a new procedure, or use another mechanism? An explicit experiment on such out-of-vocabulary cases would materially affect my assessment.
  1. Can the authors separate the benefit of AeroSOP’s runtime semantics from the benefit of restricting execution to a pre-tested procedural space? A useful controlled experiment would keep the same skill library, SOP templates, and verification coverage fixed while changing only the runtime and update semantics. This would help quantify how much of the reliability improvement comes specifically from AeroSOP’s execution protocol rather than from the fact that the available solution space has already been extensively engineered and validated.
  2. How much mid-task adaptation is actually supported? The current design primarily supports inter-task adaptation at verified completion boundaries. Could the authors provide representative cases where bounded recovery is insufficient and the system must structurally modify an ongoing task? How frequent are such cases in the current workload, and how would AeroSOP handle them?
Reviewer_LbB7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a framework that decouples base-station reasoning from autonomous aerial vehicle onboard execution through verifiable, incrementally updatable standard operating procedures (SOPs). The proposed framework was evaluated with other baselines in a simulated environment.

Critical strengths
  • This paper presents AeroSOP, a framework connecting the base-station reasoning with autonomous aerial vehicle onboard execution. The authors conducted extensive simulation experiments with many baseline methods.
Critical weaknesses
  • The first claimed contribution is not novel. Many works have covered the verifiable flight mission planning.
  • The second claimed contribution is not novel. In most open-sourced and commercial drone settings, the "AutoMode" naturally decouples the ground control station mission planning and the UAV or drone sensing, estimation and control.
  • For simulation experiments, it's hard to image a ground control station can be equipped with 8 NVIDIA H100's. Even they are the ground control station on the cloud, the paper does not mention which onboard computing devices the drone is using. The paper does not mention if these onboard computing devices are sufficient for local validation of version, capability, state, and energy compatibility.
Questions for authors
  • The authors may want to set up the problem formulation and experiments with a strong real-world implementation flavor. For example, consider to use software-in-the-loop simulation or even hardware-in-the-loop simulation such as Ardupilot for flight software, running Python on Raspberry Pi, and MissionPlanner or QGroundControl as the ground control station software paired with your ground control station Python codes and GPU.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.