Training Smarter Pilots

NullG combines native racing simulation with reinforcement-learning tools to develop learned driving and combat policies. Training produces compact models that run inside the game through a native C++ evaluator.

A native-first gameplay path

Training uses Gymnasium, Stable-Baselines3 PPO and PyTorch in a development Python environment. Isolated native workers run the engine’s race simulation, including physics, pickups, weapons and race rules.

Gameplay uses frozen .ngpolicy packages containing the actor’s weights, observation contract, physical scaling and model identity. The native loader validates each package before a race, keeping inference within the C++ game loop.

Observations and controls

The pilot observes its motion, health, connected road geometry, nearby racers, available pickups, visible projectiles, inventory and previous controls. Explicit road guidance provides heading, curvature and clearance information for steering.

Actions map to player-equivalent throttle, analog steering, airbrakes, pitch and weapon commands. Policy decisions run at 30 Hz within the 60 Hz session schedule, with the engine retaining control of simulation timing and source-specific physics.

Driving practice and combat training

The curriculum explores navigation, traffic, pickups, naturally earned Turbo and complete races. The current XL development rotation covers Talon’s Reach, Sagarmatha, Gare d’Europa and Odessa Keys.

Driving studies compare observations, rewards, exploration and steering quality. Native snapshots also support corrective practice: testing small steering interventions from states reached by the learner, then fitting accepted examples and checking full-race and Turbo behaviour.

Recent combat experiments freeze the learned driving parameters while training weapon outputs and the critic, which estimates future return. A placement-first objective prioritises race outcome, with bounded credit for effective combat and useful pickups. Ordinary races alternate with short practice windows beginning at naturally reached weapon opportunities. Original AI, a retained learned challenger and mixed fields provide varied opposition.

Reviewing progress

Evaluation compares complete races under matched conditions, tracking wins, completion, pace, damage, steering behaviour and effective attacks. Checkpoint reviews use multiple race seeds and additional confirmation panels to examine the trade-offs between candidates.

Run specifications, source and model hashes, optimiser state, random seeds, checkpoints and recorded races keep experiments traceable. Native export checks and exact checks on frozen parameters connect training changes to the policy evaluated in the engine.

Research references include Gran Turismo Sophy’s training regimen, TMRL’s real-time racing framework, CAPS smooth-control research and NeuralKart’s search-and-learn approach. Their ideas inform focused experiments within NullG’s simulation and control system.

Watch the progression

This horizontal overview follows five checkpoints from driving foundation to combat racing, ending with a complete Odessa Keys race.

Watch the training playlist for earlier checkpoints and the latest full races.