REVIEW 4 major objections 6 minor 34 references
Human-like Bots for Tactical Shooters Using Compute-Efficient Sensors
T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read This paper claims that imitation learning with a pixel-free ray-cast sensor grid can produce human-like tactical-shooter bots that make decisions in single-digit milliseconds on a CPU, with a video Turing test showing 30% of bot clips…
desk verdict Real compute-efficiency contribution, but the human-likeness evidence is undermined by training-set evaluation and a mathematically impossible negative JS divergence. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the coupled sensor-action ray grid: 225 range finders arranged non-uniformly, dense near the crosshair and sparse at the periphery, plus one extra ray per important tracked object inside a 90-degree field of view. Each sensor direction has a matching discrete aiming action, so the bot has a direct action to aim at whatever it senses, and the network learns the sensor-to-action mapping without processing pixels. Audio sensors, distance-and-direction vectors, and game-state features are concatenated into the same model, and a long short-term memory (LSTM) layer provides the temporal memory needed for imitation learning from human trajectories via backpropagation through time.
What would settle it
Compare the bots' behavior against a held-out corpus of human gameplay that was never used for training, using the same Jensen-Shannon divergence and heatmap-distance metrics; if similarity to held-out humans is substantially worse than similarity to the training humans, the human-likeness claim is inflated. A preregistered live-interaction Turing test, where players face the bots in real time with a larger participant pool, would test whether the 30% human-judgment rate survives outside curated clips.
Extended reading notes
Core claim
The central claim is that imitation learning with pixel-free sensors is sufficient to produce believable tactical-shooter bots that meet production latency budgets. Concretely, a 14.9-million-parameter network with convolutional and LSTM layers, fed by a 15-by-15 ray-cast grid over ten object types together with audio and game-state features, reproduces human behavioral distributions for round duration, shots fired, kills, plant and defuse attempts, and spatial heatmaps of map coverage. In the paper's video Turing test, 30% of bot clips were judged human and 16% drew uncertainty, while human clips were correctly identified 60% of the time. The authors conclude that this model best balances computational efficiency, quantitative similarity to human play, and perceived believability.
Load-bearing premise
The believability conclusion depends on treating the training set's own statistics as ground truth for 'human-like': the distributional comparisons use the same 48.6 hours of gameplay the models were trained on, with no held-out human data, and the human study relies on 20 participants judging curated 15-second clips rather than interactive play.
Editorial extensions
If this is right
- A single CPU thread can run the best model's decisions in 9.59 ms on average, keeping the GPU free for rendering, which matches the resource budget of commercial shooters.
- Behavior cloning from human trajectories alone—without reinforcement learning—reproduces human distributions on rounds, shots, kills, and bomb events, suggesting imitation can yield believable agents.
- The best model sits near the middle of the size range tested; larger networks (roughly 25M to 36M parameters) do not consistently improve human-likeness and run 2–3 times slower.
- The sensor-action coupling idea may transfer to other 3D games where spatial perception, rather than pixel detail, drives decisions.
- Bot clips were judged human 30% of the time with 16% uncertainty, a similar ambiguity range to human clips (60% correct, 25% mistakenly called bots).
Reading between the lines
- The paper's evaluation likely overstates human-likeness, since the distributional baselines come from the same data the bots were trained on; a held-out human dataset would give a fairer, probably lower, similarity score.
- The 15-second curated clips remove the interactive loop; in a live match, players could exploit telltale patterns (such as the reported higher average movement speed) that static clips may hide.
- The ray-grid perception scheme is a general design pattern: any 3D game with a spatial state could replace pixel encoders with structured range sensors, trading visual detail for a large cut in inference cost, at the price of losing non-spatial visual information like UI cues and subtle animations.
- The method's failure modes, including weak handling of unseen situations, could be addressed with the paper's own suggested extensions such as GAIL or Decision Transformers, and the sensor-action coupling would remain usable in those frameworks.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces a sensor-based perception architecture for training human-like bots in a VALORANT-like 2v2 tactical shooter (Lyra:Ascent). Instead of pixel input, the agents use ray-cast visual sensors, directional audio sensors, distance/direction features, and game-state information. Six LSTM-based behavior-cloning models (A-F) are trained on 48.6 hours of human gameplay and evaluated on inference time, distributional similarity to human data (Jensen-Shannon divergence), spatial similarity via heatmaps (EMD/ASD), and a human perception study. The authors report that model D (14.9M parameters) runs at 9.59 ms per decision on a consumer CPU, significantly faster than a comparable CS:GO pixel-based model (24.10 ms), and is the most human-like according to their metrics.
Significance. If the human-likeness claims survive scrutiny, the paper makes a valuable contribution to practical game AI: it demonstrates that compute-efficient, non-pixel sensors can support behavior-cloned agents in a commercial-standard shooter, with concrete and favorable inference-time comparisons. The sensor design is novel and the engineering effort is substantial. However, the believability evidence currently rests on comparisons against the training data itself and on a small human study without a baseline or significance tests; the quantitative model-selection table also contains an impossible negative JS divergence. The compute-efficiency result is the strongest part of the paper, while the human-likeness conclusion is not yet supported.
major comments (4)
- [Section VIII-B, Table IV] The distributional similarity analysis compares bot-generated data to the same 48.6 hours of human gameplay used to train the models, with no held-out split. Consequently, the reported JS divergences partly measure how well the models fit their training distribution rather than how human-like they are in novel situations. The conclusion in Section VIII-B that model D is the best-performing model is therefore not supported as a generalization claim; a held-out human dataset or cross-validation is needed before this metric can support the human-likeness claim.
- [Table IV, Defence/Speed row] The Defence/Speed row reports a JS divergence of -0.249 for model A, which is impossible because Jensen-Shannon divergence is non-negative by construction (Eq. 1). This indicates an error in the computation or in the table itself. Since this table is used to select model D as the best model, the error undermines the quantitative model comparison unless corrected.
- [Section IX, Human perception study] The human perception study uses only 20 participants classifying 32 curated 15-second clips, with no non-human baseline (e.g., a scripted bot or random guessing) and no significance tests or confidence intervals. The 54% correct bot identification rate is close to chance and is not interpretable without such statistical context, so the conclusion that the bots are 'convincing' or 'sophisticated' is not statistically supported.
- [Section VIII-C, Table V] The spatial similarity evaluation (EMD and ASD) also uses heatmaps derived from the human training data as the reference distribution. As with the JS divergence analysis in Section VIII-B, this measures fit to the training set rather than generalization to unseen human behavior. The claim in Section VIII-C that model D is 'the most robust in replicating human-like moving patterns' is therefore not supported for novel situations.
minor comments (6)
- [Section I, Contributions] There is a typo in the contributions list: 'beliavability' should be 'believability'.
- [Equation 1] The text spells 'Kullback-Liebler'; the correct spelling is 'Kullback-Leibler'.
- [Section V] The phrase 'The data was then further cleaned' should be 'The data were then further cleaned' for subject-verb agreement.
- [Section VIII-A, Table III] The inference-time comparison would be easier to interpret if the authors stated how many runs were averaged for each model and whether the measurements were on a single CPU thread; the reported error bars are standard deviations but the number of trials is not given.
- [Figure 8] The heatmap figure caption lists only models A, C, D, and E in addition to the human condition; it is unclear why models B and F are omitted from the visualization.
- [Section IX-A] The paper does not state whether the 20 participants in the human study overlapped with the 28 players whose gameplay was used for training; if there was any overlap, it should be disclosed as a potential familiarity bias.
Circularity Check
Quantitative believability measures compare against the training data itself, making model D's selection partly a fit to its own inputs; the human Turing-test component is independent but small.
-
fitted input called prediction
[Section VIII-B (Distributional Similarity) and Section VIII-C (Spatial Similarity); Table IV and Table V; Section X conclusion]
"For these two scenarios we perform quantitative comparisons between the generated bot data and the human dataset used for training. ... From these similarity comparisons we can conclude that model D is the best (quantitatively) performing model from those we trained."
The reference distribution for every JS divergence entry in Table IV and every EMD/ASD value in Table V is the same 48.6-hour human dataset on which the behavior-cloning models were trained (Section V). Low divergence/EMD therefore partly reports how well the policy reproduces its own training distribution, not an independent measure of human-likeness; the model-selection conclusion ('model D is the best') is drawn from the same training-set comparisons and then carried into the claim that D is 'the most robust in replicating human-like behavior'.
full rationale
No load-bearing self-citation or imported uniqueness theorem was found; citations to the authors' prior work are background or future-work. The compute-efficiency claim (Table III) is externally anchored by a comparison to the independently published CS:GO model [3] and is not circular. The main circularity concern is that the quantitative believability evidence compares bots to the training data with no held-out split, so the JS/EMD results partly measure training fit; the human video Turing test is an independent signal, though small (20 participants, 32 curated clips) and without a scripted-bot baseline. In addition, Table IV reports a mathematically impossible negative JS divergence (-0.249 for Model A, Defence/Speed), a correctness red flag for the exact table used in model selection. Overall, the central human-like claim rests partly on a training-set fit, giving a moderate circularity score.
Assumptions & free parameters
free parameters (4)
- Ray-cast angle sets (yaw and pitch) =
yaw: 15 values from -70 to +70 degrees; pitch: 15 values from -45 to +45 degrees; action pitch set: 11 values from -20…
- Number of ray directions =
225 (15x15 grid) plus targeted extra range finders
- Sensor max range =
100 m
- Training hyperparameters =
learning rate 0.0003, decay 0.001, batch size 96, epochs 600, timesteps 64, dropout 0.5
assumptions (4)
- domain assumption The collected 48.6 hours of gameplay from 28 LAN players is representative of human tactical-shooter behavior in Lyra:Ascent.
- domain assumption The 15x15 ray-cast grid with extra targeted range finders is a sufficient perception model for competent play in Lyra:Ascent.
- domain assumption Turing-style classification of curated 15-second video clips by 20 experienced gamers measures gameplay human-likeness.
- standard math Jensen-Shannon divergence and Earth Mover Distance on the chosen features are valid similarity metrics for behavior.
Cite this review
Pith. "Pith review of Human-like Bots for Tactical Shooters Using Compute-Efficient Sensors." pith.science (2026). https://pith.science/paper/WDUPP2SE
@misc{pith2026250100078,
author = {Pith},
title = {Pith review of: Human-like Bots for Tactical Shooters Using Compute-Efficient Sensors},
year = {2026},
howpublished = {\url{https://pith.science/paper/WDUPP2SE}},
note = {Machine review of arXiv:2501.00078}
}
read the original abstract
Artificial intelligence (AI) has enabled agents to master complex video games, from first-person shooters like Counter-Strike to real-time strategy games such as StarCraft II and racing games like Gran Turismo. While these achievements are notable, applying these AI methods in commercial video game production remains challenging due to computational constraints. In commercial scenarios, the majority of computational resources are allocated to 3D rendering, leaving limited capacity for AI methods, which often demand high computational power, particularly those relying on pixel-based sensors. Moreover, the gaming industry prioritizes creating human-like behavior in AI agents to enhance player experience, unlike academic models that focus on maximizing game performance. This paper introduces a novel methodology for training neural networks via imitation learning to play a complex, commercial-standard, VALORANT-like 2v2 tactical shooter game, requiring only modest CPU hardware during inference. Our approach leverages an innovative, pixel-free perception architecture using a small set of ray-cast sensors, which capture essential spatial information efficiently. These sensors allow AI to perform competently without the computational overhead of traditional methods. Models are trained to mimic human behavior using supervised learning on human trajectory data, resulting in realistic and engaging AI agents. Human evaluation tests confirm that our AI agents provide human-like gameplay experiences while operating efficiently under computational constraints. This offers a significant advancement in AI model development for tactical shooter games and possibly other genres.
Figures
Figures from the paper (12 more)
Reference graph
Works this paper leans on
-
[1]
A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,
David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al., “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,”Science, vol. 362, no. 6419, pp. 1140–1144, 2018
work page 2018
-
[2]
Georgios N Yannakakis and Julian Togelius,Artificial intelligence and games, vol. 2, Springer, 2018
work page 2018
-
[3]
Counter-strike deathmatch with large-scale behavioural cloning,
Tim Pearce and Jun Zhu, “Counter-strike deathmatch with large-scale behavioural cloning,” in2022 IEEE Conference on Games (CoG). IEEE, 2022, pp. 104–111
work page 2022
-
[4]
Mastering the game of go with deep neural networks and tree search,
David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al., “Mastering the game of go with deep neural networks and tree search,”nature, vol. 529, no. 7587, pp. 484–489, 2016
work page 2016
-
[5]
Starcraft ii: A new challenge for reinforcement learning,
Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexan- der Sasha Vezhnevets, Michelle Yeo, Alireza Makhzani, Heinrich Küt- tler, John Agapiou, Julian Schrittwieser, et al., “Starcraft ii: A new challenge for reinforcement learning,”arXiv preprint arXiv:1708.04782, 2017
arXiv 2017
-
[6]
Imitation learning: A survey of learning methods,
Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne, “Imitation learning: A survey of learning methods,” ACM Computing Surveys (CSUR), vol. 50, no. 2, pp. 1–35, 2017
work page 2017
-
[7]
Imitation learning: Progress, taxonomies and challenges,
Boyuan Zheng, Sunny Verma, Jianlong Zhou, Ivor W Tsang, and Fang Chen, “Imitation learning: Progress, taxonomies and challenges,”IEEE Transactions on Neural Networks and Learning Systems, , no. 99, pp. 1–16, 2022
work page 2022
-
[8]
Explo- ration in deep reinforcement learning: A survey,
Pawel Ladosz, Lilian Weng, Minwoo Kim, and Hyondong Oh, “Explo- ration in deep reinforcement learning: A survey,”Information Fusion, vol. 85, pp. 1–22, 2022
work page 2022
Show all 34 references
-
[9]
Tail: Task-specific adapters for im- itation learning with large pretrained models,
Zuxin Liu, Jesse Zhang, Kavosh Asadi, Yao Liu, Ding Zhao, Shoham Sabach, and Rasool Fakoor, “Tail: Task-specific adapters for im- itation learning with large pretrained models,” arXiv preprint arXiv:2310.05905, 2023
2023 arXiv
-
[10]
Long short-term memory,
Sepp Hochreiter and Jürgen Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997
1997
-
[11]
Behavior transformers: Cloning𝑘 modes with one stone,
Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto, “Behavior transformers: Cloning𝑘 modes with one stone,” Advances in neural information processing systems, vol. 35, pp. 22955– 22968, 2022
2022
-
[12]
Transformers for one-shot visual imitation,
Sudeep Dasari and Abhinav Gupta, “Transformers for one-shot visual imitation,” in Conference on Robot Learning. PMLR, 2021, pp. 2071– 2084
2021
-
[13]
A generalist agent,
ScottReed,KonradZolna,EmilioParisotto,SergioGomezColmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al., “A generalist agent,”arXiv preprint arXiv:2205.06175, 2022
2022 arXiv
-
[14]
Generalized behavior learning from diverse demonstrations,
Varshith Sreeramdass, Rohan RPaleja, Letian Chen, Sanne van Waveren, and Matthew Gombolay, “Generalized behavior learning from diverse demonstrations,” in First Workshop on Out-of-Distribution Generaliza- tion in Robotics at CoRL 2023, 2023
2023
-
[15]
Uncertainty- driven exploration for generalization in reinforcement learning,
Yiding Jiang, J Zico Kolter, and Roberta Raileanu, “Uncertainty- driven exploration for generalization in reinforcement learning,” inDeep Reinforcement Learning Workshop NeurIPS 2022, 2022
2022
-
[16]
Playing atari with deep reinforcement learning,
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller, “Playing atari with deep reinforcement learning,”arXiv preprint arXiv:1312.5602, 2013
2013 arXiv
-
[17]
Dota 2 with large scale deep reinforcement learning,
Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemysław Dębiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, et al., “Dota 2 with large scale deep reinforcement learning,”arXiv preprint arXiv:1912.06680, 2019
1912 arXiv
-
[18]
Outracing champion gran turismo drivers with deep reinforcement learning,
Peter R Wurman, Samuel Barrett, Kenta Kawamoto, James MacGlashan, Kaushik Subramanian, Thomas J Walsh, Roberto Capobianco, Alisa Devlic, Franziska Eckert, Florian Fuchs, et al., “Outracing champion gran turismo drivers with deep reinforcement learning,” Nature, vol. 602, no. 7...
2022
-
[19]
Conditional imitation learning for multi-agent games,
Andy Shih, Stefano Ermon, and Dorsa Sadigh, “Conditional imitation learning for multi-agent games,” in2022 17th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 2022, pp. 166– 175
2022
-
[20]
Imitation learning with concurrent actions in 3d games,
Jack Harmer, Linus Gisslén, Jorge del Val, Henrik Holst, Joakim Bergdahl, Tom Olsson, Kristoffer Sjöö, and Magnus Nordin, “Imitation learning with concurrent actions in 3d games,” in2018 IEEE Conference on Computational Intelligence and Games (CIG). IEEE, 2018, pp. 1–8
2018
-
[21]
Human-level performance in 3d multiplayer games with population-based reinforcement learning,
Max Jaderberg, Wojciech M Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castaneda, Charles Beattie, Neil C Rabinowitz, Ari S Morcos, Avraham Ruderman, et al., “Human-level performance in 3d multiplayer games with population-based reinforcement learning,” Scie...
2019
-
[22]
Navigates like me: Understanding how people evaluate human-like ai in video games,
Stephanie Milani, Arthur Juliani, Ida Momennejad, Raluca Georgescu, Jaroslaw Rzepecki, Alison Shaw, Gavin Costello, Fei Fang, Sam Devlin, and Katja Hofmann, “Navigates like me: Understanding how people evaluate human-like ai in video games,” in Proceedings of the 2023 CHI Conf...
2023
-
[23]
Assessing believability,
Julian Togelius, Georgios N Yannakakis, Sergey Karakovskiy, and Noor Shaker, “Assessing believability,”Believable Bots: Can Computers Play Like People?, pp. 215–230, 2012
2012
-
[24]
The 2k botprize,
Philip Hingston, “The 2k botprize,” in 2009 IEEE Symposium on Computational Intelligence and Games, 2009, pp. 1–1
2009
-
[25]
The turing test track of the 2012 mario ai championship: Entries and evaluation,
Noor Shaker, Julian Togelius, Georgios N. Yannakakis, Likith Poovanna, Vinay S. Ethiraj, Stefan J. Johansson, Robert G. Reynolds, Leonard K. Heether, Tom Schumann, and Marcus Gallagher, “The turing test track of the 2012 mario ai championship: Entries and evaluation,” in2013 I...
2012
-
[26]
Imitating human playing styles in super mario bros,
Juan Ortega, Noor Shaker, Julian Togelius, and Georgios N Yannakakis, “Imitating human playing styles in super mario bros,”Entertainment Computing, vol. 4, no. 2, pp. 93–104, 2013
2013
-
[27]
Plat- former level design for player believability,
Elizabeth Camilleri, Georgios N Yannakakis, and Alexiei Dingli, “Plat- former level design for player believability,” in2016 IEEE Conference on Computational Intelligence and Games (CIG). IEEE, 2016, pp. 1–8
2016
-
[28]
How humans perceive human-like behavior in video game navigation,
Evelyn Zuniga, Stephanie Milani, Guy Leroy, Jaroslaw Rzepecki, Raluca Georgescu, Ida Momennejad, Dave Bignell, Mingfei Sun, Alison Shaw, Gavin Costello, et al., “How humans perceive human-like behavior in video game navigation,” in CHI Conference on Human Factors in Computing ...
2022
-
[29]
Backpropagation through time: what it does and how to do it,
Paul J Werbos, “Backpropagation through time: what it does and how to do it,”Proceedings of the IEEE, vol. 78, no. 10, pp. 1550–1560, 1990
1990
-
[30]
Generative adversarial imitation learning,
Jonathan Ho and Stefano Ermon, “Generative adversarial imitation learning,” Advances in neural information processing systems, vol. 29, 2016
2016
-
[31]
Decision transformer: Reinforcement learning via sequence modeling,
Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch, “Decision transformer: Reinforcement learning via sequence modeling,” Advances in neural information processing systems, vol. 34, pp. 15084– 15097, 2021
2021
-
[32]
Offline reinforcement learning as one big sequence modeling problem,
Michael Janner, Qiyang Li, and Sergey Levine, “Offline reinforcement learning as one big sequence modeling problem,”Advances in neural information processing systems, vol. 34, pp. 1273–1286, 2021
2021
-
[33]
Generalized decision transformer for offline hindsight information matching,
Hiroki Furuta, Yutaka Matsuo, and Shixiang Shane Gu, “Generalized decision transformer for offline hindsight information matching,”arXiv preprint arXiv:2111.10364, 2021
2021 arXiv
-
[34]
Skill decision transformer,
Shyam Sudhakaran and Sebastian Risi, “Skill decision transformer,” arXiv preprint arXiv:2301.13573, 2023. Appendix Feature Range Player Is on attacking team [0;1] Is jumping [0;1] Is falling [0;1] Is shooting [0;1] Is being shot [0;1] Is crouching [0;1] Has main ability Zero [...
2023 arXiv
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.