MicroRTS: Minimalist RTS Research Platform

• MicroRTS is a lightweight real-time strategy game designed as an affordable, reproducible research platform that exposes combinatorial state and action spaces.
• It provides a standardized tensorized observation interface and a low-level action vocabulary, enabling controlled evaluations across diverse algorithmic approaches.
• Its integration in competitions and studies highlights its impact on advancing adversarial planning, deep reinforcement learning, program synthesis, and procedural content generation.

MicroRTS, also written as $\mu$RTS in parts of the literature, is a lightweight, minimalist, open-source, two-player, zero-sum real-time strategy game testbed designed for research. It preserves core RTS elements—resource gathering, building construction, unit production, simultaneous control of multiple units, and combat with asymmetric unit properties—while removing much of the engineering overhead of commercial RTS games. That combination has made it a widely used platform for work on adversarial planning, program synthesis, deep reinforcement learning, uncertainty handling, opponent modeling, and procedural content generation, and it underlies a long-running competition ecosystem (Goodfriend, 2024, Huang et al., 2021, Moraes et al., 2024).

1. Research platform and competition setting

MicroRTS is repeatedly described as a research-oriented RTS environment that is computationally affordable while still exposing the combinatorial state and action spaces, sparse rewards, delayed consequences, and simultaneous-action structure that make full-game RTS difficult. In contrast to environments such as StarCraft II, it simplifies unit types, terrain, and production, but still requires tactical micro and strategic macro to be coordinated within a real-time control loop (Goodfriend, 2024, Huang et al., 2021).

The platform has also become a standardized benchmark. The IEEE microRTS competitions were hosted at the Conference on Computational Intelligence and Games starting in 2017 and at the IEEE Conference on Games nearly every year since 2019. Competition fields include hand-written programmatic bots, synthesized programmatic agents, search-based agents, and more recently deep reinforcement learning agents, with evaluation commonly conducted through round-robin play on fixed suites of maps (Goodfriend, 2024).

Its suitability for algorithmic research is reinforced by two technical properties emphasized across the literature. First, the environment has a fast forward model and discrete frame-based time, which enables adversarial look-ahead and controlled computation budgets for search-based agents. Second, although MicroRTS supports fog of war and non-determinism, many benchmark settings disable both, yielding full observability and deterministic transitions that simplify controlled comparisons without eliminating the strategic and combinatorial structure of the game (Barriga et al., 2017, Richoux, 2020).

2. Rules, entities, and formal environment structure

MicroRTS is played on a grid map. Players control buildings and units that gather resources, construct infrastructure, produce new units, and attack the opponent until one side has no remaining units or buildings, or until a time limit yields a draw. Core entities repeatedly referenced in the literature are resources, bases, barracks, workers, and the combat-unit types light, heavy, and ranged. Workers harvest resources, return them to bases, build structures, and can attack; bases store resources and train workers; barracks train combat units (Moraes et al., 2023, Huang et al., 2021).

The low-level action vocabulary is also standardized. Across the Gym-$\mu$RTS and LLM-oriented papers, atomic or component actions include NOOP, move, harvest, return, produce, and attack, typically with directional parameters or relative attack-position parameters. Actions are durative and simultaneous: units may continue executing previously issued actions while new commands are assigned elsewhere, and the environment advances in discrete time steps rather than turn-taking phases (Cui et al., 16 Sep 2025, Huang et al., 2019).

A canonical tensorized observation interface appears in Gym-$\mu$RTS. For a map of size $h \times w$, the observation is a tensor of shape $(h,w,27)$ with one-hot planes for hit points, carried resources, owner, unit type, and current action. The corresponding low-level unit action is represented as an 8-component discrete vector over source unit, action type, move/harvest/return parameters, produce direction, produce type, and relative attack position, with $a_r=7$ giving 49 relative attack offsets (Huang et al., 2021).

Research papers also exploit the fact that MicroRTS can be instantiated in different informational regimes. Some work uses deterministic, fully observable settings to analyze semantics of policies from single roll-outs; other work studies the partially observable track with fog of war and hidden enemy composition. The latter line frames production planning as a decision problem under uncertainty, rather than as pure perfect-information search (Moraes et al., 2024, Richoux, 2020).

3. Programmatic representations and search spaces

A major branch of MicroRTS research represents policies as programs in a domain-specific language. In the program-synthesis literature, strategies are encoded as abstract syntax trees over high-level constructs such as loops, conditionals, build and train commands, movement, attack priorities, and resource-harvesting routines. The Microlanguage used for semantic-space search provides primitives such as harvest, build, train, move, attack, and state predicates, together with loop-based prioritization: once an entity receives an action earlier in the loop, later instructions cannot overwrite it in the same tick. Formally, a program $p$ induces a deterministic policy $\pi_p$, with semantics $\mathrm{Sem}(p): S \to A^U$ over controllable entities $U$ (Moraes et al., 2024).

This representation supports explicit optimization over program space. One formulation casts synthesis as a robust max–min search over the DSL-induced policy set,

$\mu$0

with approximate best responses computed by local search inside Programmatic PSRO. The Local Learner (2L) line of work argues that the quality of the opponent meta-strategy determines the quality of the search signal, and it constructs a support of “helpful” reference opponents through a greedy set-cover approximation, retaining multiple informative adversaries while pruning redundant ones (Moraes et al., 2023).

MicroRTS has also been a testbed for studying the geometry of program spaces. “Searching for Programmatic Policies in Semantic Spaces” replaces purely syntax-guided neighborhoods with a library-induced semantic neighborhood that swaps AST subparts for executable subprograms with distinct action signatures. On MicroRTS, this reduces the probability that a random program has a semantically identical neighbor from $\mu$1 in syntax space to $\mu$2 in the learned semantic space, and the resulting LISS method consistently outperforms syntax-guided baselines on five of six test maps while changing only the neighborhood function (Moraes et al., 2024).

Other synthesis work uses demonstrations or foundation models to bias program search. Sketch-learning showed that behavioral cloning can supply effective partial structures even when demonstrations come from weak players, and Sketch-SA(O) was the only approach in that study to achieve strong performance on both $\mu$3 and $\mu$4 maps against COAC. InnateCoder extends the idea by extracting subprograms from foundation-model-generated MicroLanguage code and treating them as options in a mixed syntax/semantic search; with combined GPT-4o and Llama 3.1 option libraries, the resulting system achieved average win rate $\mu$5 against COAC, RAISocketAI, and Mayari across four maps (Medeiros et al., 2022, Moraes et al., 18 May 2025).

4. Deep reinforcement learning and hybrid control

MicroRTS is equally prominent as an affordable full-game RL environment. Gym-$\mu$6RTS introduced an OpenAI Gym interface together with a practical PPO-based recipe built around action composition, full invalid-action masking, opponent diversity, and CNN backbones such as IMPALA-CNN and encoder–decoder models. In the single-map setting studied there, the best agent defeated every tested past competition bot on the $\mu$7 basesWorkers map, achieving a cumulative win rate of $\mu$8 across the opponent pool after roughly $\mu$9–$\mu$0 hours of training on a single machine with one GPU, three vCPU, and 16 GB RAM (Huang et al., 2021).

A more focused DRL representation study compared global and local control for a harvesting task. The global representation observes the whole map and chooses both the unit and its action; the local representation observes a window centered on a scheduled unit and predicts only that unit’s action. On $\mu$1, $\mu$2, and $\mu$3 harvesting maps, local per-unit control substantially outperformed the global controller, with the best $\mu$4 local setting reaching $\mu$5 total gathered resources versus $\mu$6 for the global representation under the reported protocol (Huang et al., 2019).

The first DRL competition winner in MicroRTS was RAISocketAI. It used PPO with invalid action masking, GridNet-style simultaneous issuance, multiple value heads, iterative fine-tuning against strong opponents, and map-specific transfer learning. In IEEE-CoG 2023 it became the first deep reinforcement learning agent to win the microRTS competition, and in the reported competition table it achieved $\mu$7 overall win rate across submissions, defeating every opponent head-to-head, including $\mu$8 against 2L and $\mu$9 against Mayari. The same paper also reports that transfer learning to specific maps was critical to its performance (Goodfriend, 2024).

MicroRTS has additionally served as a setting for hybrid architectures that separate macro strategy from local tactics. A notable example combines a fully convolutional policy that imitates Puppet Search’s script selection with a tactical NaïveMCTS restricted to units near enemies. On large $h \times w$0 $h \times w$1RTS maps ported from StarCraft: Brood War AIIDE competition maps, PolicyNet+NaïveMCTS achieved average win rate $h \times w$2, compared with $h \times w$3 for PuppetSearch+NaïveMCTS and $h \times w$4 for PuppetSearch alone, illustrating a recurrent MicroRTS theme: strategic abstraction is effective, but tactical refinement remains decisive (Barriga et al., 2017).

5. Recent extensions: uncertainty, LLM planning, style analysis, and procedural generation

MicroRTS has become a substrate for research beyond conventional search and RL. Under partial observability, microPhantom models the Unit Production Problem with constraint programming and rank-dependent utility, using online beliefs over hidden enemy unit types and a 100 ms anytime solver. Against a matched baseline, the reported average normalized scores rose from $h \times w$5 to $h \times w$6 on basic maps and from $h \times w$7 to $h \times w$8 on open maps, while “chaotic” variants with changing game attributes incurred only a small loss of efficiency overall (Richoux, 2020).

LLM-based planning work uses MicroRTS precisely because it is adversarial, long-horizon, and unfamiliar to most LLMs. PLAP introduces a library of five parameterized skills—Deploy Unit, Harvest Mineral, Build Building, Produce Unit, and Attack Enemy—together with an LLM planner and a deterministic executor that replans every $h \times w$9 steps. On the zero-shot LLM-vs-LLM leaderboard for basesWorkers8x8, GPT-4o reached Scores $(h,w,27)$0, Win Rate $(h,w,27)$1, Resource Harvesting Ratio $(h,w,27)$2, Resource Utilization Rate $(h,w,27)$3, Unit Production Rate $(h,w,27)$4, and Combat Efficiency Ratio $(h,w,27)$5, while Qwen2-72B under fs-tip-PLAP surpassed CoacAI in baseline matches (Cui et al., 16 Sep 2025).

A related line, Strategy-Augmented Planning, constructs an explicit strategy space and trains a Strategy Evaluation Network offline, then uses an LLM online to recognize opponent strategy from summarized trajectories and greedily select a best response. On basesWorkers8x8, SAP achieved $(h,w,27)$6 average win rate across the compared LLM-based opponents, corresponding to an $(h,w,27)$7 relative improvement over the strongest plan-level baseline, and it generalized strongly to unseen strategies, with reported $(h,w,27)$8 win rate against the unseen set (Xu et al., 13 May 2025).

MicroRTS has also been used to study behavior as data. “Play Style Identification Using Low-Level Representations of Play Traces in MicroRTS” trains CNN–BiLSTM autoencoders directly on low-level state/action trace tensors from 13 agents on the 12x12basesWorkers family. On the held-out map $(h,w,27)$9, the joint state+action model with $a_r=7$0 clusters achieved Completeness $a_r=7$1, Homogeneity $a_r=7$2, ARI $a_r=7$3, and AMI $a_r=7$4, showing that agent-specific structure can be separated without handcrafted style features (Xia et al., 14 Jul 2025).

The environment has even been extended for search-based procedural content generation. One study modifies the standard $a_r=7$5 competition map to add a fourth barracks unit and a new ability system, then uses greedy hill climbing with an MCTS-based fitness to generate ten units such as Revenger, Phoenix, Lawman, and Barrage. The reported units were evaluated through two-round simulated play to measure both utility and balance (Sorochan et al., 2022).

6. Benchmarking ecosystem, typical baselines, and recurring limitations

MicroRTS research is unusually cumulative because many papers reuse common maps, bots, and evaluation protocols. Frequent baselines include scripted rushes such as LightRush and WorkerRush, search agents such as NaiveMCTS, programmatic competition winners such as COAC and Mayari, synthesized programmatic agents such as 2L, and increasingly DRL systems such as RAISocketAI or PPO GridNet. Round-robin tournaments, swapped starting positions, and map suites ranging from small rush maps to large macro maps such as BloodBath.scmB recur across the literature (Moraes et al., 2023, Goodfriend, 2024).

This ecosystem has made it possible to compare distinct methodological traditions on the same game. In a simulated tournament over seven maps, 2L achieved the highest overall Total of $a_r=7$6 and was the only method with average winning rate $a_r=7$7 against all opponents. LISS later beat recent competition winners on average, with reported overall average $a_r=7$8 across COAC, Mayari, and RAISocketAI in its external-opponent evaluation. RAISocketAI then marked the first DRL competition victory, while PLAP and SAP reframed the environment as a benchmark for skill-grounded and strategy-grounded LLM planning (Moraes et al., 2023, Moraes et al., 2024, Goodfriend, 2024, Cui et al., 16 Sep 2025, Xu et al., 13 May 2025).

At the same time, the literature identifies recurring failure modes. Larger maps remain difficult for many methods; RAISocketAI, for example, was weak on BloodBath.scmB in its reported benchmarks, and both local and global DRL controllers deteriorated sharply as map size increased in the harvesting study. Programmatic DSLs often emphasize macro-strategy over fine-grained micromanagement, which can leave micro-heavy maps favorable to specialized DRL or tactical-search agents. Many studies also adopt deterministic, fully observable settings even though MicroRTS supports fog of war and non-determinism, so results do not automatically transfer to the partially observable or stochastic variants (Goodfriend, 2024, Huang et al., 2019, Moraes et al., 2024).

Taken together, these lines of work establish MicroRTS as a compact but technically rich RTS benchmark. It supports direct low-level control through tensorized observations and compositional action spaces, high-level control through interpretable DSLs and parameterized skills, and a competition culture that exposes methods to strong scripted, search-based, synthesized, and learned opponents. A plausible implication is that its continuing importance derives not from maximal realism, but from the unusual degree to which it makes full-game RTS research reproducible, comparable, and algorithmically analyzable across otherwise disparate paradigms (Huang et al., 2021, Xia et al., 14 Jul 2025).