Game or device state
Simulation today; visual device path as the deployment direction.
mixed maturityNPLUS / Kodo
Building agentsthat learn to play.
Kodo rebuilds Brawl Stars as a controllable game environment first, then uses that system as the foundation for reinforcement learning.
FIG. 001Kodo ENVIRONMENT / MEDIA SLOT
STATE / FRAME
OBS / SCHEMATIC
PATH / SAMPLED
OVERVIEW / CURRENT DIRECTION
One game system. Two observation paths.
The current architecture direction keeps a stable actor-facing contract while the observation source changes between simulation and a device perception path. Exact model and library choices remain under review.
SOURCE / SIMULATION
Kodo environment
structured state
ACTOR CONTRACT
Observation→Representation→Memory→Policy→Action
SOURCE / DEVICE
Screen → perception
planned path
TRAINING ONLY
Learning may use information unavailable to the deployed actor. Any such branch must stay visibly separate from actor input.
STACK / WORKING ARCHITECTURE
From source to action.
This is the public contract, not a frozen implementation spec. Components show their role and authority state; exact algorithm, model-family, dimension, and library choices are deliberately withheld until locked.
Actor input contract
A representation the policy can consume across execution regimes.
working designEncoder + memory
Temporal and entity representation are still under architectural review.
under reviewMulti-agent policy
The policy family is a current direction rather than a locked public claim.
under reviewControl contract
Actions return to the simulated game or, later, the device-control path.
working designTRAINING PATHoptimisation / privileged signals may branch here
DEVICE PATHscreen perception must supply only what the actor can observe
PLAN / WORKING ROADMAP
Human video → imitation → multi-agent → device.
The latest simplified planning revision is used here as a roadmap, not as proof that every stage or stack choice is already implemented.
Prerequisites first
Training results do not count until the current environment and observation blockers are closed.
details held until confirmedHuman play video
Collect representative gameplay from the real game.
working routeCV reverse-engineering
Infer observation and action labels from recorded play.
working routeBehaviour cloning
Use human data to initialise a policy before reinforcement learning.
plannedMulti-agent self-play
Continue learning against a changing population of agents.
plannedDevice test
Evaluate perception, delay, aim control, and the sim-to-device gap.
test directionPUBLIC SIMPLIFICATION Detailed blocker IDs, exact model names, action dimensions, observation dimensions, and hyperparameters are intentionally not reproduced here.
RETRAINING RISK Device-side changes may invalidate earlier checkpoints; that dependency stays visible in the roadmap rather than being hidden as implementation detail.
STATUS/ TRUTH LAYER
Confirmed build vs working plan.
Kodo now separates what has been reported implemented from architecture that is likely but not yet locked.
CONFIRMED BUILD
Core game logicImplemented
User interfaceImplemented
Gas systemIn development
WORKING DIRECTION
Architecture shown aboveNot locked
Model / library choicesUnder review
Roadmap stagesPlanning sequence
PUBLICATION POLICY
Exact dimensions /hyperparametersWithheld
Detailed blocker listWithheld
Public team rosterOmitted
AUTHORITY BASIS / NPLUS teammate reply + revised planning diagram µ 2026.08.22 µ deleted chat content is not used as evidence.