{"id":"f2ab06e9-0996-4fe0-b4ea-6391a0a2930f","arxiv_id":"2505.01953","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Tunnel couples F-16 nonlinear flight dynamics with customizable Gymnasium sensors and tasks, plus an informal comparison of RL, behavioral cloning, and PID baselines.","lead":"This preprint introduces Tunnel, a small Python environment that puts F-16 flight dynamics into the standard Gymnasium interface for training AI agents. It reports a week of informal experiments comparing reinforcement learning, imitation learning, and a hand-tuned autopilot, but provides no code link and no quantitative results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central deliverable is missing: no code, repository, or commit hash appears anywhere, so the core claim that Tunnel is an open-source Gymnasium environment cannot be verified and the trade study is not reproducible.","rationale":"The reader's weakest_assumption targets fidelity transfer from the 13-state F-16 model to operationally relevant flight. That is a reasonable concern, but it is not the most load-bearing one here, because the paper explicitly frames Tunnel as a low-cost training environment rather than a high-fidelity simulator and concedes in Section IX that the missionized extension is 'not high fidelity'. The more fundamental problem is that the central deliverable — the open-source code — is absent from the manuscript. The abstract and Section V.A both advertise open source, but no repository, URL, or commit hash is given, and no code is included beyond the PID equations in the Appendix. Without the environment, the physics integration, sensor primitives, and reward structure cannot be checked, and the empirical trade study in Section VI cannot be reproduced. The trade study's internal validity is also weak: Section VI.A describes a sequence of ad hoc observation and reward changes, with no seeds, hyperparameters, or numerical results, and Table 1 reports only categorical outcomes. The paper does include a concrete PID controller in the Appendix, which is a useful independent artifact, but it does not compensate for the missing environment. Because the central claim of an open-source tool cannot be verified, the REJECT verdict is correct, though the primary reason should be the missing artifact rather than the fidelity assumption. I therefore leave the reader's verdict unchanged.","tokens_in":11292,"tokens_out":3211,"duration_ms":33345,"concrete_test":"Attempt to retrieve the Tunnel code from any public source (e.g., GitHub, the author's affiliation, or the DAF AI Accelerator). If no public repository exists, the open-source claim fails and the REJECT verdict stands. If a repository exists, install it and run a fixed-seed PPO baseline using the final observation space from Section VI.A (last four timesteps of the 3x3 sensor returns, no internal state), then compare end-reach success and wall-hit rates against Table 1; if the claimed 'Marginal' or 'No' outcomes cannot be reproduced with the documented setup, the trade-study claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is a software artifact, not a theory. The abstract promises a 'simple, open source' environment, and Section V.A lists 'Open source' as a design feature, yet the manuscript contains no URL, repository name, commit hash, or code listing — only PID equations in the Appendix. Without the environment, no one can instantiate the F-16 13-state dynamics, sensor geometry, or reward structure. Consequently, the 'week long trade study' in Section VI is unrepeatable. The trade study itself is also anecdotal: Section VI.A reports post hoc changes to the observation and reward design with no seeds, hyperparameters, or quantitative success rates, and Table 1 uses only categorical labels 'Yes/Marginal/No'. The author's own concession that the missionized extension is 'not high fidelity' (Section IX) further limits any operational-transfer reading, though this is secondary because the paper positions Tunnel as a training environment rather than a validated high-fidelity simulator.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces Tunnel, an OpenAI Gymnasium environment for reinforcement learning with F-16 nonlinear flight dynamics. The manuscript describes the environment's design features (Section V), reports a week-long trade study comparing PPO-trained RNN and MLP agents and behavioral cloning from a PID expert (Section VI), and presents a \"missionized\" extension with threat zones and sensor modifications. The abstract and Section V.A claim the environment is open source and written in fewer than 300 lines of code. The paper does not include a repository URL, code listing, or quantitative experimental results; it reports mostly categorical outcomes and anecdotal findings, and it acknowledges in Section IX that the missionized extension is \"not high fidelity.\"","tokens_in":11475,"tokens_out":7913,"duration_ms":79647,"significance":"If the environment were publicly available and the demonstrations reproducible, Tunnel could be a useful community resource: it addresses a real gap between inaccessible proprietary simulators and simpler drone environments, and it is explicitly positioned as a training environment rather than a validated high-fidelity simulator. The author correctly distinguishes training environments from simulation (Section IV.E) and cites related open-source efforts. However, as submitted, the contribution cannot be assessed: there is no code to inspect, no quantitative evaluation of agent performance, and the operational-relevance claims are not supported by the evidence in the trade study.","major_comments":[{"comment":"The paper repeatedly claims that Tunnel is open source, but no repository URL, commit hash, or code listing is provided anywhere in the manuscript. Since the paper's contribution is a software artifact, this absence prevents the reader from instantiating the F-16 13-state dynamics, sensor geometry, or reward structure, and it makes the Section VI trade study unrepeatable. Even the dynamics are not described beyond a citation to [23]. This is a load-bearing omission, not a presentation issue.","section":"Abstract; Section V.A; Section VIII.A"},{"comment":"The trade study is unreplicated and unquantified. No random seeds, episode counts, hyperparameters, training curves, or success-rate statistics are reported. The statement that \"there was no significant difference seen between RNN and MLP agents\" is made without a statistical test or error bars, and Table 1 uses only the categorical labels \"Yes/Marginal/No.\" As a result, the paper's central demonstration—that Tunnel enables rapid, meaningful comparison of training methods—is not supported.","section":"Section VI.A; Table 1"},{"comment":"The environment configuration and reward function were altered after observing agent failures: the reward was changed \"based on the results of the Trackmania effort,\" and the final observation used the last four timesteps of sensor data with no internal state. Because the final configuration is selected post hoc and no held-out validation or separate test set is described, the reported results are at risk of overfitting to the observed failures and do not establish generalizable performance.","section":"Section VI.A"},{"comment":"The PID \"expert\" controller is not fully specified. The equations contain only proportional and derivative terms, despite the name PID, and omit the integral term. The manuscript does not define the sampling interval, the units or body-axis conventions for Nz and Ps, or how these commands map to the action space. Since the behavioral-cloning comparison in Section VI.B depends on this expert, the result cannot be reproduced or checked.","section":"Appendix A; Section VI.B"}],"minor_comments":[{"comment":"The nomenclature list omits Nz, Ps, and NX, all of which are used in the equations and observation descriptions; please define these symbols.","section":"Section I; Appendix A"},{"comment":"Several references have incomplete bibliographic data (e.g., [7], [9], [10], [21], [34], and [38] contain placeholders such as \"????\" or missing years/pages); these should be completed.","section":"References"},{"comment":"The paper states that Tunnel is \"written using three files and less than 300 lines of code,\" but no code is available to verify this claim; please provide the repository or remove the specific count.","section":"Section III"},{"comment":"The text first says the missionized navigation task \"was reliably accomplished by use of an autopilot\" and then reports instances of trespass; please clarify the success criterion and the observed failure rate.","section":"Section VI.C"},{"comment":"The conclusion that the missionized extension is \"not high fidelity\" should be reflected in the abstract's claim of \"operationally relevant aircraft physics\" to avoid overstating operational transfer.","section":"Section IX"}],"recommendation":"reject","confidential_remarks":"I agree with the stress-test concern: the absence of any code repository or quantitative evaluation is decisive for the current submission. The manuscript reads as a project report rather than a refereed research paper. If the author can provide a public repository with the full environment and a properly repeated experimental study with seeds, episode counts, and error bars, a resubmission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a tool paper whose tool is missing. The abstract promises a 'simple, open source' Gymnasium environment for F-16 dynamics, but the manuscript contains no code, no repo URL, no commit hash, and no runnable example—only a few PID equations in the appendix. You cannot instantiate the environment, verify the dynamics integration, or reproduce the trade study. For a software contribution, that is the whole ballgame.\n\nWhat is genuinely new is the pitch: a deliberately small, Python-only wrapper around Heidlauf et al.'s 13-state F-16 model, with sensor arrays, boundary primitives, and a 'missionized' mode for non-air-air tasks. The writing is clear, the author openly cites JSBSim and Harfang3D as existing Gymnasium-capable simulators, and the paper is honest about the missionized extension being 'not high fidelity.' The trade study design—comparing PPO, behavioral cloning, and a simple PID within a week—is a sensible demonstration of the intended workflow, and the finding that a PID beats RL in this narrow corridor task is unsurprising but worth a footnote.\n\nThe soft spots are large and mostly load-bearing. No seeds, no episode counts, no training curves, no error bars; Table 1 uses 'Yes/Marginal/No.' The observation and reward schedule were changed after failures, so the final configuration is post hoc. The PID 'expert' is two proportional-derivative lines, so the behavioral-cloning comparison is more a smoke test than a study. The operational-relevance language in the abstract—'national military advantage'—is not supported by anything in the paper, and the author's own fidelity caveat undercuts any transfer claims. None of this would be fatal if the environment itself were downloadable and the benchmark reproducible; as is, the core claim cannot be checked.\n\nWho is this for? A reader interested in a lightweight F-16 environment for RL prototyping might find the design rationale useful, but they should wait for the code. I would not cite it in its current form, and I would not send it to a full referee. A desk reject with an invitation to resubmit alongside a public code release and a proper empirical appendix is the right call.","headline":"A useful idea undermined by a missing artifact: the paper advertises an open-source Gymnasium environment but ships no code, and the trade study is unrepeatable.","tokens_in":11989,"tokens_out":1821,"would_cite":false,"duration_ms":18024,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tunnel embeds nonlinear F-16 flight dynamics in an open-source Gymnasium environment, letting air-combat researchers and mission planners change tasks, sensors, and training methods in days rather than months.","keywords":["Tunnel","reinforcement learning","F-16 flight dynamics","Gymnasium","air combat autonomy","imitation learning","mission planning","training environment"],"falsifier":"Run a policy trained in Tunnel on the same tunnel task inside a higher-fidelity F-16 simulation that includes sensor noise, actuator limits, and atmospheric effects; if the policy's success rate collapses, the claim that Tunnel's dynamics are operationally relevant would be falsified.","tokens_in":11085,"feed_emoji":"✈️","tokens_out":5836,"duration_ms":53327,"temperature":0.7,"pith_summary":"This paper introduces Tunnel, an open-source reinforcement-learning training environment that wraps F-16 nonlinear flight dynamics in the standard Gymnasium Python interface. The author's central claim is that because the environment is deliberately simple, researchers and mission planners can rapidly change observations, actions, sensors, and tasks, compressing what normally takes months into days. A week-long trade study is presented as evidence: the author varied training algorithms, observation spaces, and threat presentations, and built a missionized tunnel with missile engagement zones and GPS-denied sensing. If Tunnel works as claimed, it would give the air-combat autonomy community a shared, low-cost testbed that connects machine-learning research to operationally relevant aircraft physics.","feed_headline":"Open-source tool trains air-combat agents on F-16 physics in days","feed_subtitle":"A 300-line environment lets researchers and mission planners swap sensors, tasks, and algorithms in under a week.","key_machinery":"The central machinery is the Tunnel environment itself: a Gymnasium-compatible wrapper that combines a 13-state nonlinear F-16 flight model (forces, moments, kinematics, position, plus thrust lag) with configurable sensor nodes arranged in body-axis elevation and azimuth. The sensors return distances to the nearest wall, giving the agent a LiDAR-like image; observations can be swapped among sensor returns and a 16-element aircraft-state vector, and rewards can be reshaped, as when the study switched from a centerline-distance penalty to arithmetic target rewards. This plug-and-play design is what carries the argument: it makes observation, action, and task changes cheap enough to iterate in days, and it standardizes the interface with the machine-learning community.","core_discovery":"The paper claims that high-performance air-combat autonomy can be productively explored in a training environment rather than a full simulation, and that Tunnel is such an environment. Tunnel couples a 13-state, six-degree-of-freedom F-16 model with Gymnasium's reset-and-step interface, a configurable sensor array that returns range-to-wall distances, and a default action space of stick, throttle, and rudder. The environment's walls stand in for the tactical constraints of a mission, and the author argues this abstraction is enough to test which observation spaces, algorithms, and training schemes deserve further investment. The demonstration trade study found that a simple behavioral-cloning agent could navigate the tunnel while reinforcement-learning agents could not reliably do so, and that a missionized variant with moving threat zones could be built and evaluated in under a week.","pith_inferences":["Editorial inference: the paper's strongest implied claim is about transfer from Tunnel to real flight, yet the reported experiments stop at the abstract tunnel task; Tunnel's real value is likely as a cheap first filter that identifies which algorithm-sensor combinations deserve expensive high-fidelity evaluation, not as a substitute for it.","Editorial inference: if the simplicity claim holds, Tunnel could serve as a community benchmark for partial observability and adversarial air combat, letting different research groups compare policies on identical dynamics for the first time.","Editorial inference: the imitation-learning result suggests a testable extension, collecting expert trajectories from the waypoint autopilot and varying observation noise to see how much sensor realism is needed before behavioral cloning breaks down.","Editorial inference: because the environment supports on-off threat control, self-play between two Tunnel agents could be used to study how engagement zones and sensor visibility shape adversarial policies."],"forward_implications":["Researchers can benchmark machine-learning algorithms against a common high-performance-aircraft task without months of simulator-specific training.","Mission planners can prototype sensor configurations, threat zones, and navigation tasks as primitives, then pass the resulting environment to a reinforcement-learning or imitation-learning pipeline.","A week-long trade study becomes feasible: the paper shows a single developer can compare PPO with recurrent and MLP policies, imitation learning, and classical autopilot control in one week.","The missionized tunnel demonstrates that GPS-denied, partially observed, adversarial scenarios can be represented well enough to begin exploring them before moving to higher-fidelity simulators."],"supporting_citations":[{"why":"Supplies the 13-state nonlinear F-16 dynamics that Tunnel wraps; its fidelity is the load-bearing physics of the whole environment.","marker":"[23]"},{"why":"The Trackmania effort inspires the reward function of rewarding passage of regularly spaced targets inside the tunnel.","marker":"[40]"},{"why":"The A* algorithm is used in the missionized environment to plot the agent's course to the goal while avoiding threat zones.","marker":"[41]"},{"why":"An open-source flight dynamics model with Gymnasium support that Tunnel positions itself against as simpler and easier to modify.","marker":"[11]"},{"why":"An air-to-air reinforcement-learning sandbox for fighter aircraft, used as the comparison point for customizable air-combat training environments.","marker":"[12]"},{"why":"A drone-focused training environment whose speed and maneuverability, the paper argues, make it invalid for high-performance aircraft conclusions.","marker":"[7]"}],"fun_headline_variants":["Train air-combat AI on F-16 physics in days","Open-source F-16 env trains air-combat AI in a week","F-16 physics env for air-combat AI: days not months","Air-combat AI dev faster with open-source F-16 env"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the simplified 13-state F-16 flight model used in Tunnel is realistic enough that conclusions about agent behavior transfer to real high-performance aircraft; the paper itself concedes the extended mission version is not high fidelity.","fun_headline_variants_meta":{"raw":{"variants":["Train air-combat AI on F-16 physics in days","Open-source F-16 env trains air-combat AI in a week","F-16 physics env for air-combat AI: days not months","Air-combat AI dev faster with open-source F-16 env"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001055,"raw_usage":{"total_tokens":4406,"prompt_tokens":900,"completion_tokens":3506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":3430}},"tokens_in":516,"tokens_out":3506,"duration_ms":23001,"temperature":1.0,"reasoning_tokens":3430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:05:32.645604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a policy trained in Tunnel on the same tunnel task inside a higher-fidelity F-16 simulation that includes sensor noise, actuator limits, and atmospheric effects; if the policy's success rate collapses, the claim that Tunnel's dynamics are operationally relevant would be falsified.","supporting_citations":[{"cited_title":"Verification Challenges in F-16 Ground Collision Avoidance and Other Automated Maneuvers,","cited_arxiv_id":null,"evidence_quote":"Supplies the 13-state nonlinear F-16 dynamics that Tunnel wraps; its fidelity is the load-bearing physics of the whole environment."},{"cited_title":"Automated Learning: An Implementation of The A* Search Algorithm over The Random Base Functions","cited_arxiv_id":"2211.05085","evidence_quote":"The A* algorithm is used in the missionized environment to plot the agent's course to the goal while avoiding threat zones."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An open-source flight dynamics model with Gymnasium support that Tunnel positions itself against as simpler and easier to modify."},{"cited_title":"Harfang3D Dog-Fight Sandbox: A Reinforcement Learning Research Platform for the Customized Control Tasks of Fighter Aircrafts","cited_arxiv_id":"2210.07282","evidence_quote":"An air-to-air reinforcement-learning sandbox for fighter aircraft, used as the comparison point for customizable air-combat training environments."},{"cited_title":"Airsim: High-fidelity visual and physical simulation for autonomous vehicles","cited_arxiv_id":null,"evidence_quote":"A drone-focused training environment whose speed and maneuverability, the paper argues, make it invalid for high-performance aircraft conclusions."}],"review_version":1}