{"authorId":"","cover":"/content/images/2026/10/9e08d4941150-social-background.png","heroes":{"39b4d054f7e0bce8c6537f8d80a0a2e3":"/content/images/2017/04/IMG_0528.JPG","59ab1e54b3a5386e6900ca8678d4e839":"/content/images/2026/10/b1e12503f058-nullg-rl-full-preview.png","b00b035b3794d4eb5846a7122bf026cf":"/content/images/2026/10/ab60de3338bd-nullg-weekly-fresh-xl-shield.png"},"kind":"tag","logo":"/content/images/2016/08/projectable-me-4.png","nav":[{"id":"62e854d471683dea63488bcbca5c6188","label":"Home","url":"/","sortOrder":0},{"id":"1bc84fb15620acd635df7735a0b8ddd8","label":"Projects","url":"/projects/","sortOrder":1},{"id":"3207e2fec51d66d2a5b3d8d04feb9dd8","label":"Sim Racing","url":"/tag/sim-racing/","sortOrder":2},{"id":"59115afd52e59a74faa4b386e00ae3fa","label":"About","url":"/about-me/","sortOrder":3}],"page":1,"perPage":10,"posts":[{"id":"b00b035b3794d4eb5846a7122bf026cf","siteId":"projectable","slug":"nullg-weekly-progress-xl-fleet-rl-pulse-pure","title":"NullG updates: the XL fleet, shields and a first Pulse race","markdown":"This week's [NullG](https://projectable.me/projects/nullg/) work covered the five rebuilt XL ships, weapon effects, split-screen and the learning pilots. There's also a first playable slice of WipEout Pulse. I've put the ships and all eleven XL weapon types into a new video:\n\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/2VYHD8xlsoM\" frameborder=\"0\" allowfullscreen\u003e\u003c/iframe\u003e\n\n## The XL ships\n\nFEISAR, AG Systems, Auricom, Qirex and Piranha now have rebuilt geometry and physically based materials, with more detail around the cockpits, engines and moving control surfaces. The [Blender article](https://projectable.me/nullg-recreating-wipeout-xl-models-in-blender/) covers how the models were reconstructed with the assistant and reviewed.\n\n![FEISAR and AG Systems in the Model Viewer](https://projectable.me/content/images/2026/10/0f9bf03cf410-nullg-weekly-fresh-fleet-a.png)\n\nA still render only tells part of the story with the moving parts. Qirex has a broad hull to clear when its surfaces move, while the thin surfaces on AG Systems and Piranha needed further refinement. This week's changes included those clearance and thickness checks, then reviews under the menu and track lighting.\n\n![Auricom, Qirex and Piranha with their control surfaces posed](https://projectable.me/content/images/2026/10/a300c3095728-nullg-weekly-fresh-fleet-b.png)\n\nFor the racing footage I used the newest completed RL checkpoint available when capture preparation started on October 25. It was an experimental pace-training checkpoint on simulation revision 48, still unaccepted for gameplay. I wanted the video to show where the training had reached that day.\n\nAll eight racers in each recorded field use the same saved policy, so continuing the training couldn’t change the driver halfway through recording. This week's training work also covered difficulty objectives, rewards, checkpoint retention and resuming earlier runs. The [RL page](https://projectable.me/projects/nullg/reinforcement-learning/) explains how the pilot produces steering, throttle, airbrake, pitch and weapon inputs.\n\n## Shields and weapons\n\nThe shield now has an animated surface around the hull, with the ship and engine flare visible through it. The close-up shows the surface moving while the ship stays visible underneath.\n\n![Blue shield surrounding Qirex on Sagarmatha](https://projectable.me/content/images/2026/10/ecc5894c0244-nullg-weekly-fresh-shield-closeup.png)\n\nFor the weapon shots I used controlled demonstrations with granted inventory. Rockets, missile and electrobolt show their launch, flight and impact; mines, quake, plasma and energy bomb show the larger track effects. Turbo, autopilot and the energy pack have their activation and status feedback too, making eleven types including the shield. The mine sequence also places a live mine to demonstrate the impact.\n\n![Rockets leaving Qirex on Sagarmatha](https://projectable.me/content/images/2026/10/e4a12b0e8f41-nullg-weekly-fresh-xl-rocket-0.png)\n\nTargeting trails, status placement and effect draw order also had fixes. The video uses chase, cockpit and wider views, and the XL camera handoffs and results view needed some stabilization work.\n\nThere was also work on preparing assets and keeping textures available on the GPU. Menus, inspection views and races use the shared lighting and effects; the [architecture page](https://projectable.me/projects/nullg/architecture/) explains how that fits together.\n\n## Split-screen\n\nI've changed the HUD layout to fit each player's viewport. The top readouts stay at the top, speed and thrust sit near the bottom, and status information stays around the edges. That leaves more of the road visible in both horizontal and vertical layouts.\n\n![Two-player split-screen with separate weapon and shield readouts](https://projectable.me/content/images/2026/10/169198aae575-nullg-weekly-fresh-split-hud.png)\n\n*Weapon, shield and lap states were staged here to check the two-player HUD layout.*\n\nHD2097 is used for the racing labels and large figures. Brighter numbers and colored labels distinguish the shield countdown, health, boost and thrust, while compact text keeps names and results readable in the smaller views. Targeting trails and weapon feedback stay with the relevant player's camera.\n\nThree-player mode now uses the spare quadrant for a circuit map and live standings, including lap and speed information. The numbered ship markers, legend and standings use the same colors for each local player.\n\n![Three player views with a map and standings in the fourth quadrant](https://projectable.me/content/images/2026/10/cf608e45de29-nullg-weekly-fresh-split-overview.png)\n\nFinished and retired racers have their own result states, with each driver's results kept inside their view. The [multiplayer page](https://projectable.me/projects/nullg/multiplayer/) covers the rest of the local and network setup.\n\n## A first Pulse race\n\nFor this update, the playable Pulse preview is **Talon's Junction White, FEISAR and Venom**. It goes through the normal import and menus into a one-lap Time Trial or a Single Race against seven native opponents, with difficulty selectable in Race Setup.\n\n![FEISAR with an active shield and HUD countdown on Talon's Junction White](https://projectable.me/content/images/2026/10/e14ca56845db-nullg-weekly-fresh-pulse-shield.png)\n\nGetting that combination running involved grid placement, the countdown and launch behavior, steering, airbrakes and the sideshift gesture. Banking and barrel-roll visuals are in, along with a chase camera that follows the track horizon and feedback for landings, wall contact and recovery.\n\nTime Trial gives you a Turbo at the start/lap line. It uses the normal weapon controls and common HUD, and can be absorbed for energy. Ship and track contact feed into the shared damage handling.\n\nPulse's Shield and three-shot Rocket bursts are connected to the models, rendering and HUD, though the weapon work is still incomplete. Tests covered shielded and unshielded encounters, and all eight racers completed a lap in the validation run. The weapon screenshots use inventory granted by the inspection tool.\n\n![Three Rocket darts ahead of FEISAR on Talon's Junction White](https://projectable.me/content/images/2026/10/9899d3185aba-nullg-weekly-fresh-pulse-rockets.png)\n\nPure is further back, at the import stage. The shared PSP work decodes geometry, materials, textures, collision and animation data while keeping each game's content separate. CPU loader checks cover **16 Pure track layouts and 24 Pulse layouts**, and the ship metadata covers **eleven Pure ships and eight Pulse ships**, including their dependent resources. Sound-bank voices and soundtrack resources have also been decoded and checked.\n\nThose numbers describe what the loaders can read, not a list of playable content. Pulse is the preview described above; Pure isn't at that point yet. The imports feed the existing model cache, staged GPU uploads and track streaming. Pulse imports are checked in a separate installation before being activated in the asset library.\n\n## Working with the inspection tools\n\nA lot of the model and weapon work is done with the engine paused. The MCP harnesses let an agent or script operate the native tools: load a model, position the camera, pose a control surface, inspect its state and save a frame.\n\nFor the fleet, that means going from Blender to the engine's Model Viewer, checking materials and movement, then checking the menus and a race. A control surface can look fine in Blender and still need a clearance adjustment when the engine applies steering or an airbrake.\n\n![Qirex in the Model Viewer with bounds and control-surface hierarchy visible](https://projectable.me/content/images/2026/10/e00793761eb9-nullg-weekly-fresh-mcp-qirex-surfaces.png)\n\nThe weapon harness starts a paused production race and grants a selected weapon. It can then advance the simulation a few frames at a time. A snapshot saved before firing allows the same setup to be restored, with ship state, timers and events available beside the pictures. The demonstration still goes through the game's firing and rendering paths.\n\nPulse now has a couple of useful feedback probes too. One starts the ship above the road with downward velocity and lets hover and collision produce the landing. Another moves it beyond the track so the recovery controller has to respond. The reports record the racer, time and magnitude of the resulting feedback, giving a repeatable case to inspect.\n\n![Track Validator showing geometry and texture inspection controls](https://projectable.me/content/images/2026/10/a43258e42685-nullg-weekly-fresh-mcp-track-inspection.png)\n\nThe replay tools let the camera work happen after the race is recorded. The saved inputs and reports make the captures repeatable, and the same race can be used for landscape and portrait edits. There's more on the harnesses on the [MCP tooling page](https://projectable.me/projects/nullg/mcp-tooling/).\n\nThe music work this week included playlist playback and track announcements, as well as **Zero G Chokehold**, **On Fire** and the **Warehouse Rage** track used in the video. NullG's original soundtrack is generated with Suno.\n\n*WipEout names, designs and imported game content belong to their respective owners. These images show that content and its reconstructed presentation in NullG.*","html":"\u003cp\u003eThis week\u0026#39;s \u003ca href=\"https://projectable.me/projects/nullg/\" rel=\"nofollow\"\u003eNullG\u003c/a\u003e work covered the five rebuilt XL ships, weapon effects, split-screen and the learning pilots. There\u0026#39;s also a first playable slice of WipEout Pulse. I\u0026#39;ve put the ships and all eleven XL weapon types into a new video:\u003c/p\u003e\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/2VYHD8xlsoM\" frameborder=\"0\" allowfullscreen=\"\"\u003e\u003c/iframe\u003e\n\u003ch2\u003eThe XL ships\u003c/h2\u003e\n\u003cp\u003eFEISAR, AG Systems, Auricom, Qirex and Piranha now have rebuilt geometry and physically based materials, with more detail around the cockpits, engines and moving control surfaces. The \u003ca href=\"https://projectable.me/nullg-recreating-wipeout-xl-models-in-blender/\" rel=\"nofollow\"\u003eBlender article\u003c/a\u003e covers how the models were reconstructed with the assistant and reviewed.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/0f9bf03cf410-nullg-weekly-fresh-fleet-a.png\" alt=\"FEISAR and AG Systems in the Model Viewer\"\u003e\n\u003cp\u003eA still render only tells part of the story with the moving parts. Qirex has a broad hull to clear when its surfaces move, while the thin surfaces on AG Systems and Piranha needed further refinement. This week\u0026#39;s changes included those clearance and thickness checks, then reviews under the menu and track lighting.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/a300c3095728-nullg-weekly-fresh-fleet-b.png\" alt=\"Auricom, Qirex and Piranha with their control surfaces posed\"\u003e\n\u003cp\u003eFor the racing footage I used the newest completed RL checkpoint available when capture preparation started on October 25. It was an experimental pace-training checkpoint on simulation revision 48, still unaccepted for gameplay. I wanted the video to show where the training had reached that day.\u003c/p\u003e\n\u003cp\u003eAll eight racers in each recorded field use the same saved policy, so continuing the training couldn’t change the driver halfway through recording. This week\u0026#39;s training work also covered difficulty objectives, rewards, checkpoint retention and resuming earlier runs. The \u003ca href=\"https://projectable.me/projects/nullg/reinforcement-learning/\" rel=\"nofollow\"\u003eRL page\u003c/a\u003e explains how the pilot produces steering, throttle, airbrake, pitch and weapon inputs.\u003c/p\u003e\n\u003ch2\u003eShields and weapons\u003c/h2\u003e\n\u003cp\u003eThe shield now has an animated surface around the hull, with the ship and engine flare visible through it. The close-up shows the surface moving while the ship stays visible underneath.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/ecc5894c0244-nullg-weekly-fresh-shield-closeup.png\" alt=\"Blue shield surrounding Qirex on Sagarmatha\"\u003e\n\u003cp\u003eFor the weapon shots I used controlled demonstrations with granted inventory. Rockets, missile and electrobolt show their launch, flight and impact; mines, quake, plasma and energy bomb show the larger track effects. Turbo, autopilot and the energy pack have their activation and status feedback too, making eleven types including the shield. The mine sequence also places a live mine to demonstrate the impact.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/e4a12b0e8f41-nullg-weekly-fresh-xl-rocket-0.png\" alt=\"Rockets leaving Qirex on Sagarmatha\"\u003e\n\u003cp\u003eTargeting trails, status placement and effect draw order also had fixes. The video uses chase, cockpit and wider views, and the XL camera handoffs and results view needed some stabilization work.\u003c/p\u003e\n\u003cp\u003eThere was also work on preparing assets and keeping textures available on the GPU. Menus, inspection views and races use the shared lighting and effects; the \u003ca href=\"https://projectable.me/projects/nullg/architecture/\" rel=\"nofollow\"\u003earchitecture page\u003c/a\u003e explains how that fits together.\u003c/p\u003e\n\u003ch2\u003eSplit-screen\u003c/h2\u003e\n\u003cp\u003eI\u0026#39;ve changed the HUD layout to fit each player\u0026#39;s viewport. The top readouts stay at the top, speed and thrust sit near the bottom, and status information stays around the edges. That leaves more of the road visible in both horizontal and vertical layouts.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/169198aae575-nullg-weekly-fresh-split-hud.png\" alt=\"Two-player split-screen with separate weapon and shield readouts\"\u003e\n\u003cp\u003e\u003cem\u003eWeapon, shield and lap states were staged here to check the two-player HUD layout.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eHD2097 is used for the racing labels and large figures. Brighter numbers and colored labels distinguish the shield countdown, health, boost and thrust, while compact text keeps names and results readable in the smaller views. Targeting trails and weapon feedback stay with the relevant player\u0026#39;s camera.\u003c/p\u003e\n\u003cp\u003eThree-player mode now uses the spare quadrant for a circuit map and live standings, including lap and speed information. The numbered ship markers, legend and standings use the same colors for each local player.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/cf608e45de29-nullg-weekly-fresh-split-overview.png\" alt=\"Three player views with a map and standings in the fourth quadrant\"\u003e\n\u003cp\u003eFinished and retired racers have their own result states, with each driver\u0026#39;s results kept inside their view. The \u003ca href=\"https://projectable.me/projects/nullg/multiplayer/\" rel=\"nofollow\"\u003emultiplayer page\u003c/a\u003e covers the rest of the local and network setup.\u003c/p\u003e\n\u003ch2\u003eA first Pulse race\u003c/h2\u003e\n\u003cp\u003eFor this update, the playable Pulse preview is \u003cstrong\u003eTalon\u0026#39;s Junction White, FEISAR and Venom\u003c/strong\u003e. It goes through the normal import and menus into a one-lap Time Trial or a Single Race against seven native opponents, with difficulty selectable in Race Setup.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/e14ca56845db-nullg-weekly-fresh-pulse-shield.png\" alt=\"FEISAR with an active shield and HUD countdown on Talon\u0026#39;s Junction White\"\u003e\n\u003cp\u003eGetting that combination running involved grid placement, the countdown and launch behavior, steering, airbrakes and the sideshift gesture. Banking and barrel-roll visuals are in, along with a chase camera that follows the track horizon and feedback for landings, wall contact and recovery.\u003c/p\u003e\n\u003cp\u003eTime Trial gives you a Turbo at the start/lap line. It uses the normal weapon controls and common HUD, and can be absorbed for energy. Ship and track contact feed into the shared damage handling.\u003c/p\u003e\n\u003cp\u003ePulse\u0026#39;s Shield and three-shot Rocket bursts are connected to the models, rendering and HUD, though the weapon work is still incomplete. Tests covered shielded and unshielded encounters, and all eight racers completed a lap in the validation run. The weapon screenshots use inventory granted by the inspection tool.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/9899d3185aba-nullg-weekly-fresh-pulse-rockets.png\" alt=\"Three Rocket darts ahead of FEISAR on Talon\u0026#39;s Junction White\"\u003e\n\u003cp\u003ePure is further back, at the import stage. The shared PSP work decodes geometry, materials, textures, collision and animation data while keeping each game\u0026#39;s content separate. CPU loader checks cover \u003cstrong\u003e16 Pure track layouts and 24 Pulse layouts\u003c/strong\u003e, and the ship metadata covers \u003cstrong\u003eeleven Pure ships and eight Pulse ships\u003c/strong\u003e, including their dependent resources. Sound-bank voices and soundtrack resources have also been decoded and checked.\u003c/p\u003e\n\u003cp\u003eThose numbers describe what the loaders can read, not a list of playable content. Pulse is the preview described above; Pure isn\u0026#39;t at that point yet. The imports feed the existing model cache, staged GPU uploads and track streaming. Pulse imports are checked in a separate installation before being activated in the asset library.\u003c/p\u003e\n\u003ch2\u003eWorking with the inspection tools\u003c/h2\u003e\n\u003cp\u003eA lot of the model and weapon work is done with the engine paused. The MCP harnesses let an agent or script operate the native tools: load a model, position the camera, pose a control surface, inspect its state and save a frame.\u003c/p\u003e\n\u003cp\u003eFor the fleet, that means going from Blender to the engine\u0026#39;s Model Viewer, checking materials and movement, then checking the menus and a race. A control surface can look fine in Blender and still need a clearance adjustment when the engine applies steering or an airbrake.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/e00793761eb9-nullg-weekly-fresh-mcp-qirex-surfaces.png\" alt=\"Qirex in the Model Viewer with bounds and control-surface hierarchy visible\"\u003e\n\u003cp\u003eThe weapon harness starts a paused production race and grants a selected weapon. It can then advance the simulation a few frames at a time. A snapshot saved before firing allows the same setup to be restored, with ship state, timers and events available beside the pictures. The demonstration still goes through the game\u0026#39;s firing and rendering paths.\u003c/p\u003e\n\u003cp\u003ePulse now has a couple of useful feedback probes too. One starts the ship above the road with downward velocity and lets hover and collision produce the landing. Another moves it beyond the track so the recovery controller has to respond. The reports record the racer, time and magnitude of the resulting feedback, giving a repeatable case to inspect.\u003c/p\u003e\n\u003cimg src=\"https://projectable.me/content/images/2026/10/a43258e42685-nullg-weekly-fresh-mcp-track-inspection.png\" alt=\"Track Validator showing geometry and texture inspection controls\"\u003e\n\u003cp\u003eThe replay tools let the camera work happen after the race is recorded. The saved inputs and reports make the captures repeatable, and the same race can be used for landscape and portrait edits. There\u0026#39;s more on the harnesses on the \u003ca href=\"https://projectable.me/projects/nullg/mcp-tooling/\" rel=\"nofollow\"\u003eMCP tooling page\u003c/a\u003e.\u003c/p\u003e\n\u003cp\u003eThe music work this week included playlist playback and track announcements, as well as \u003cstrong\u003eZero G Chokehold\u003c/strong\u003e, \u003cstrong\u003eOn Fire\u003c/strong\u003e and the \u003cstrong\u003eWarehouse Rage\u003c/strong\u003e track used in the video. NullG\u0026#39;s original soundtrack is generated with Suno.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eWipEout names, designs and imported game content belong to their respective owners. These images show that content and its reconstructed presentation in NullG.\u003c/em\u003e\u003c/p\u003e\n","excerpt":"I've been working on the rebuilt XL ships, shields, split-screen and learning pilots, plus the first playable Pulse preview and more PSP imports.","metaTitle":"NullG updates: the XL fleet, shields and a first Pulse race","metaDescription":"I've been working on the rebuilt XL ships, shields, split-screen and learning pilots, plus the first playable Pulse preview and more PSP imports.","status":"published","featured":false,"publishedAt":"2026-10-25T23:34:35Z","authorId":"190160de54f2192bd5b8734ec8808086","authorName":"Daniel Chote","authorSlug":"daniel","heroMediaId":"b6d17adfd043ab59aaf5a462247eea23","projectId":"ae4bdb66da6dc1729a263a475e1cdbe3","tags":[{"id":"10a974c0a281d4fcf3d44f5ef8dae079","name":"NullG","slug":"nullg"},{"id":"1c9e8a1339a8f8b797767e91a9673016","name":"Game","slug":"game"},{"id":"60103f47eb65b46c6138b9d6fc7b17d0","name":"Machine Learning","slug":"machine-learning"}],"createdAt":"2026-10-25T23:12:18Z","updatedAt":"2026-10-26T18:08:27Z"},{"id":"59ab1e54b3a5386e6900ca8678d4e839","siteId":"projectable","slug":"nullg-reinforcement-learning-journey","title":"Teaching NullG pilots to drive and use weapons","markdown":"I've been working on getting a learned pilot to race in [NullG](https://projectable.me/projects/nullg/). It trains against the native simulation, exports into a model the game can run, and uses the same controls as a player. The latest work is on weapons: keeping the driving it has already learned while training it to use what it picks up.\n\nGetting a ship around a circuit is only part of this. The pilot also has to deal with traffic, pickups, Turbo and the consequences of firing a weapon. Changes that help one part can upset another, so a lot of the work between training runs has been figuring out what actually changed and replaying it on other tracks.\n\nThe [YouTube playlist](https://www.youtube.com/playlist?list=PLaR-_mRLlb6E) has the driving and combat recordings. This is the longer account of how the training works and what the latest results mean.\n\n## Connecting the pilot to the game\n\nPython communicates with isolated native workers, each running NullG's C++ race simulation with its physics, pickups, weapons and rules. Gymnasium supplies the environment interface, with PyTorch and [Stable-Baselines3 PPO](https://stable-baselines3.readthedocs.io/en/master/modules/ppo.html) handling the main training implementation.\n\nThe policy is the network that turns observations into controls. During training, a second network, the critic, estimates the return expected from a state. PPO collects experience with the current policy and makes bounded updates to those networks.\n\nThe pilot gets structured information about motion, health, the connected road ahead, nearby racers, pickups, visible projectiles, inventory and previous controls. That makes the observation design quite concrete: how much road can it see, how is sideways motion represented, and what happens to those measurements at a section boundary?\n\nIt produces throttle, analog steering, left and right airbrakes, pitch and weapon commands at **30 Hz**. These pass through the normal player-input path. The session advances at **60 Hz**, with the selected source physics cadence; movement is held between decisions and weapon commands follow the engine's input timing.\n\nThe trained actor exports to a compact `.ngpolicy` package. A native C++ evaluator then produces the controls in the game, without Python in that runtime path. The package records which observations and configuration the model expects, along with its identity, so it can be checked against the training setup.\n\n![Title view from a NullG reinforcement-learning evaluation replay](/content/images/2026/10/b1b8220f350c-nullg-rl-full-title-preview.png)\n\n*An evaluation race rendered by the cinematic replay exporter. The driving policy and camera director are separate systems.*\n\n## Useful references\n\nThere is plenty of existing racing RL work to learn from. I've used different projects for different parts of this, adapting the ideas to an anti-gravity racer and NullG's own simulation.\n\n[Sony AI's account of training Gran Turismo Sophy](https://ai.sony/blog/how-to-train-your-race-car) is useful for thinking about the practice a driver needs: solo driving, traffic and specialized situations. That informs the mix of circuits, opponents and difficult starting states used here.\n\n[TMRL](https://github.com/trackmania-rl/tmrl) covers real-time control, observation histories, analog actions and off-policy learning in TrackMania. It helped motivate a controlled SAC steering comparison alongside PPO. [Linesight](https://linesight-rl.github.io/linesight/build/html/main_objects.html), another TrackMania project, uses IQN and explicit reference-line geometry. Its handling of progress and training coverage is useful even though NullG's integration is different. [Learn-to-Race](https://github.com/learn-to-race/l2r) provides an autonomous-racing environment and baselines for comparing conventional and learned controllers against a defined task.\n\n[CAPS](https://ai.bu.edu/caps/) informed the smooth-control studies, including looking at action variation and the frequency content of motion. [NeuralKart](https://github.com/rameshvarun/NeuralKart) provided a reference for searching for useful driving actions and learning from them. NullG does that search from native snapshots of states reached by its own learner. [DAgger](https://proceedings.mlr.press/v15/ross11a.html) explains why collecting corrections in those states matters: the learner's previous decisions determine where it ends up needing help.\n\n## The observations and steering\n\nOnce the basic loop worked, the next job was checking what information actually reached the network. That meant auditing scaling, feature connections, road preview and the steering adapter.\n\nA value can be present in an observation and still be nearly useless after normalization and clipping. The corrected observations use explicit physical scales and identity normalization, which leaves those scaled values unchanged. Checks also trace features all the way to the network inputs, so adding a field isn't mistaken for successfully teaching the model about it.\n\nThe road observations now include forward directions, heading error, longitudinal and lateral velocity, displacement across the road, curvature and clearance. Connected-road projection supplies a continuous position through section boundaries and qualified junctions. Progress rewards use distance along the road and only count new ground beyond the previous furthest point. Pickups and boost pads make limited contributions.\n\nThe XL input adapter now provides **513 signed positions** for analog steering. That made it possible to compare observation designs while keeping the control protocol and experience budget fixed. A separate PPO-versus-SAC study used the same restricted auxiliary controls. These smaller comparisons helped isolate decisions before asking the full driver to do more.\n\n## What happens between training runs\n\nBetween October 15 and 20, the work repeatedly went from recorded races and telemetry to a focused change, then back through evaluation. Simply turning up a steering-change penalty wasn't enough. It could make sampled actions quieter, so the checks also looked at the policy's mean output and the ship's actual heading and lateral movement. Longer road preview and different brake commitment were tested separately.\n\nExploration adds another complication. The actions sampled during training can behave differently from deterministic native inference, so both need checking. The exported driver is the one that eventually has to race.\n\nA steering update changes the trajectory from which Turbo is fired, even when the weapon weights stay fixed. States with naturally earned Turbo were added to the comparisons and examples used to preserve existing behaviour. Speed alone isn't enough to judge the result; health, recovery and what happens over the rest of the race matter too.\n\nFor corrective driving, the tools save a native state reached by the learner and try small steering interventions from that exact point. Accepted examples become labels for training, with unchanged actions retained as anchors. Work on Gare d'Europa and Odessa Keys added longer follow-through checks and preservation samples from the other circuits.\n\nTrying a correction involves quite a few steps: restore the state, check that the original run replays correctly, try alternative inputs, train a limited update, export it and run complete races. A turn can look better for a few seconds and still lead to trouble further down the road, so the follow-through checks were extended.\n\nThree completed combat jobs—the eight-hour run on October 18, the four-hour run on October 19 and the latest four-hour continuation—account for **more than 16 hours of training wall time**. They branch from selected earlier checkpoints, so that isn't sixteen continuous hours on one policy. Short comparisons, earlier driving work, evaluations and corrective searches are additional work. That total doesn't include reading, reviewing footage or changing the tools; there isn't a combined time counter for those.\n\n## Adding combat\n\nFor combat, I kept the actor's driving features and movement outputs frozen while training the weapon outputs and critic. Weight checks verify that the protected parameters stay unchanged. Full races are still necessary, because a new weapon decision can change the situation the driver has to deal with.\n\nFinishing position matters most in the reward, with limited extra credit for effective combat and useful pickups or pads. Firing more shots doesn't earn a bonus by itself.\n\nThe experiments covered longer credit assignment, combat reward weighting and critic calibration. Calibration fits the value estimate to returns collected from the retained policy without changing the actor. For extra practice, a race can start from a saved state where the driver had reached an opponent while carrying an offensive weapon. Those chances can be sparse in a normal race.\n\nThe latest configuration mixes ordinary races with **ten-second combat opportunity windows**. These start from verified native states and collect fresh actions from the current policy. The end of a practice window isn't counted as finishing or being eliminated from a race.\n\nThe opponents include original XL drivers, a retained learned challenger and mixed fields. The training rotation uses Talon's Reach, Sagarmatha, Gare d'Europa and Odessa Keys.\n\n## The latest four-hour run\n\nThis continuation started from the earlier combat run's hour-two checkpoint. Four CPU workers added **6,639,616 decisions**, taking that checkpoint lineage from **32,651,264 to 39,290,880**. Other experiments have their own experience counts.\n\nThe run retained the driving parameters and trained the weapon outputs and critic, using gamma **1.0**, GAE lambda **0.95**, combat reward scale **1.0** and **300-decision** opportunity windows. Hourly saves allowed intermediate checkpoints to be compared with the final one.\n\nThe review covered six policies, each with 48 standard evaluation races, plus cross-play and confirmation panels. [RLiable's evaluation guidance](https://research.google/blog/rliable-towards-reliable-evaluation-reporting-in-reinforcement-learning/) is useful here: paired conditions, several seeds and the distribution of outcomes give more information than a single good race.\n\nThe final checkpoint **won 18 of 48 races and finished 31** in the standard panel. In the additional 24-race confirmation panel, it **won 9 and finished 15**. The intermediate checkpoints had different strengths; the last save wasn't simply better at everything.\n\n![Gare d'Europa race from the final combat-training checkpoint](/content/images/2026/10/b1e12503f058-nullg-rl-full-preview.png)\n\n## The recordings\n\nThe learned pilot is in slot 0 in the latest full race. The automatic cameras move around the field, so every ship shown in a close-up isn't necessarily that pilot.\n\n### Gare d'Europa, complete race\n\nThis is the three-lap race from the final checkpoint of the four-hour continuation. It runs for about three minutes, with native race audio and cinematic cameras.\n\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/__kpDHYbsCg\" frameborder=\"0\" allowfullscreen\u003e\u003c/iframe\u003e\n\n[Watch the full Gare d'Europa race on YouTube](https://youtu.be/__kpDHYbsCg).\n\n### Racing and combat Short\n\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/lTbLUj1QKt4\" frameborder=\"0\" allowfullscreen\u003e\u003c/iframe\u003e\n\n[Watch the racing and combat Short](https://youtu.be/lTbLUj1QKt4).\n\n### Earlier training: update 3\n\nThe earlier upload, “NullG - 8 hour AI driving training 3 (reinforcement learning)”, shows another point in the driving work.\n\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/fkUTwW2p_8k\" frameborder=\"0\" allowfullscreen\u003e\u003c/iframe\u003e\n\n[Watch training update 3](https://youtu.be/fkUTwW2p_8k).\n\n### Earlier training: update 2\n\nThe preceding upload is “NullG - 8 hour AI driving training 2 (reinforcement learning)”.\n\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/9ddeaoyyIY4\" frameborder=\"0\" allowfullscreen\u003e\u003c/iframe\u003e\n\n[Watch training update 2](https://youtu.be/9ddeaoyyIY4).\n\nThe next experiments will build on these comparisons, working on weapon decisions while checking that the driving still holds up across the four circuits. The [reinforcement-learning page](https://projectable.me/projects/nullg/reinforcement-learning/) has more detail on the engine integration.\n","html":"\u003cp\u003eI\u0026#39;ve been working on getting a learned pilot to race in \u003ca href=\"https://projectable.me/projects/nullg/\" rel=\"nofollow\"\u003eNullG\u003c/a\u003e. It trains against the native simulation, exports into a model the game can run, and uses the same controls as a player. The latest work is on weapons: keeping the driving it has already learned while training it to use what it picks up.\u003c/p\u003e\n\u003cp\u003eGetting a ship around a circuit is only part of this. The pilot also has to deal with traffic, pickups, Turbo and the consequences of firing a weapon. Changes that help one part can upset another, so a lot of the work between training runs has been figuring out what actually changed and replaying it on other tracks.\u003c/p\u003e\n\u003cp\u003eThe \u003ca href=\"https://www.youtube.com/playlist?list=PLaR-_mRLlb6E\" rel=\"nofollow\"\u003eYouTube playlist\u003c/a\u003e has the driving and combat recordings. This is the longer account of how the training works and what the latest results mean.\u003c/p\u003e\n\u003ch2\u003eConnecting the pilot to the game\u003c/h2\u003e\n\u003cp\u003ePython communicates with isolated native workers, each running NullG\u0026#39;s C++ race simulation with its physics, pickups, weapons and rules. Gymnasium supplies the environment interface, with PyTorch and \u003ca href=\"https://stable-baselines3.readthedocs.io/en/master/modules/ppo.html\" rel=\"nofollow\"\u003eStable-Baselines3 PPO\u003c/a\u003e handling the main training implementation.\u003c/p\u003e\n\u003cp\u003eThe policy is the network that turns observations into controls. During training, a second network, the critic, estimates the return expected from a state. PPO collects experience with the current policy and makes bounded updates to those networks.\u003c/p\u003e\n\u003cp\u003eThe pilot gets structured information about motion, health, the connected road ahead, nearby racers, pickups, visible projectiles, inventory and previous controls. That makes the observation design quite concrete: how much road can it see, how is sideways motion represented, and what happens to those measurements at a section boundary?\u003c/p\u003e\n\u003cp\u003eIt produces throttle, analog steering, left and right airbrakes, pitch and weapon commands at \u003cstrong\u003e30 Hz\u003c/strong\u003e. These pass through the normal player-input path. The session advances at \u003cstrong\u003e60 Hz\u003c/strong\u003e, with the selected source physics cadence; movement is held between decisions and weapon commands follow the engine\u0026#39;s input timing.\u003c/p\u003e\n\u003cp\u003eThe trained actor exports to a compact \u003ccode\u003e.ngpolicy\u003c/code\u003e package. A native C++ evaluator then produces the controls in the game, without Python in that runtime path. The package records which observations and configuration the model expects, along with its identity, so it can be checked against the training setup.\u003c/p\u003e\n\u003cimg src=\"/content/images/2026/10/b1b8220f350c-nullg-rl-full-title-preview.png\" alt=\"Title view from a NullG reinforcement-learning evaluation replay\"\u003e\n\u003cp\u003e\u003cem\u003eAn evaluation race rendered by the cinematic replay exporter. The driving policy and camera director are separate systems.\u003c/em\u003e\u003c/p\u003e\n\u003ch2\u003eUseful references\u003c/h2\u003e\n\u003cp\u003eThere is plenty of existing racing RL work to learn from. I\u0026#39;ve used different projects for different parts of this, adapting the ideas to an anti-gravity racer and NullG\u0026#39;s own simulation.\u003c/p\u003e\n\u003cp\u003e\u003ca href=\"https://ai.sony/blog/how-to-train-your-race-car\" rel=\"nofollow\"\u003eSony AI\u0026#39;s account of training Gran Turismo Sophy\u003c/a\u003e is useful for thinking about the practice a driver needs: solo driving, traffic and specialized situations. That informs the mix of circuits, opponents and difficult starting states used here.\u003c/p\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/trackmania-rl/tmrl\" rel=\"nofollow\"\u003eTMRL\u003c/a\u003e covers real-time control, observation histories, analog actions and off-policy learning in TrackMania. It helped motivate a controlled SAC steering comparison alongside PPO. \u003ca href=\"https://linesight-rl.github.io/linesight/build/html/main_objects.html\" rel=\"nofollow\"\u003eLinesight\u003c/a\u003e, another TrackMania project, uses IQN and explicit reference-line geometry. Its handling of progress and training coverage is useful even though NullG\u0026#39;s integration is different. \u003ca href=\"https://github.com/learn-to-race/l2r\" rel=\"nofollow\"\u003eLearn-to-Race\u003c/a\u003e provides an autonomous-racing environment and baselines for comparing conventional and learned controllers against a defined task.\u003c/p\u003e\n\u003cp\u003e\u003ca href=\"https://ai.bu.edu/caps/\" rel=\"nofollow\"\u003eCAPS\u003c/a\u003e informed the smooth-control studies, including looking at action variation and the frequency content of motion. \u003ca href=\"https://github.com/rameshvarun/NeuralKart\" rel=\"nofollow\"\u003eNeuralKart\u003c/a\u003e provided a reference for searching for useful driving actions and learning from them. NullG does that search from native snapshots of states reached by its own learner. \u003ca href=\"https://proceedings.mlr.press/v15/ross11a.html\" rel=\"nofollow\"\u003eDAgger\u003c/a\u003e explains why collecting corrections in those states matters: the learner\u0026#39;s previous decisions determine where it ends up needing help.\u003c/p\u003e\n\u003ch2\u003eThe observations and steering\u003c/h2\u003e\n\u003cp\u003eOnce the basic loop worked, the next job was checking what information actually reached the network. That meant auditing scaling, feature connections, road preview and the steering adapter.\u003c/p\u003e\n\u003cp\u003eA value can be present in an observation and still be nearly useless after normalization and clipping. The corrected observations use explicit physical scales and identity normalization, which leaves those scaled values unchanged. Checks also trace features all the way to the network inputs, so adding a field isn\u0026#39;t mistaken for successfully teaching the model about it.\u003c/p\u003e\n\u003cp\u003eThe road observations now include forward directions, heading error, longitudinal and lateral velocity, displacement across the road, curvature and clearance. Connected-road projection supplies a continuous position through section boundaries and qualified junctions. Progress rewards use distance along the road and only count new ground beyond the previous furthest point. Pickups and boost pads make limited contributions.\u003c/p\u003e\n\u003cp\u003eThe XL input adapter now provides \u003cstrong\u003e513 signed positions\u003c/strong\u003e for analog steering. That made it possible to compare observation designs while keeping the control protocol and experience budget fixed. A separate PPO-versus-SAC study used the same restricted auxiliary controls. These smaller comparisons helped isolate decisions before asking the full driver to do more.\u003c/p\u003e\n\u003ch2\u003eWhat happens between training runs\u003c/h2\u003e\n\u003cp\u003eBetween October 15 and 20, the work repeatedly went from recorded races and telemetry to a focused change, then back through evaluation. Simply turning up a steering-change penalty wasn\u0026#39;t enough. It could make sampled actions quieter, so the checks also looked at the policy\u0026#39;s mean output and the ship\u0026#39;s actual heading and lateral movement. Longer road preview and different brake commitment were tested separately.\u003c/p\u003e\n\u003cp\u003eExploration adds another complication. The actions sampled during training can behave differently from deterministic native inference, so both need checking. The exported driver is the one that eventually has to race.\u003c/p\u003e\n\u003cp\u003eA steering update changes the trajectory from which Turbo is fired, even when the weapon weights stay fixed. States with naturally earned Turbo were added to the comparisons and examples used to preserve existing behaviour. Speed alone isn\u0026#39;t enough to judge the result; health, recovery and what happens over the rest of the race matter too.\u003c/p\u003e\n\u003cp\u003eFor corrective driving, the tools save a native state reached by the learner and try small steering interventions from that exact point. Accepted examples become labels for training, with unchanged actions retained as anchors. Work on Gare d\u0026#39;Europa and Odessa Keys added longer follow-through checks and preservation samples from the other circuits.\u003c/p\u003e\n\u003cp\u003eTrying a correction involves quite a few steps: restore the state, check that the original run replays correctly, try alternative inputs, train a limited update, export it and run complete races. A turn can look better for a few seconds and still lead to trouble further down the road, so the follow-through checks were extended.\u003c/p\u003e\n\u003cp\u003eThree completed combat jobs—the eight-hour run on October 18, the four-hour run on October 19 and the latest four-hour continuation—account for \u003cstrong\u003emore than 16 hours of training wall time\u003c/strong\u003e. They branch from selected earlier checkpoints, so that isn\u0026#39;t sixteen continuous hours on one policy. Short comparisons, earlier driving work, evaluations and corrective searches are additional work. That total doesn\u0026#39;t include reading, reviewing footage or changing the tools; there isn\u0026#39;t a combined time counter for those.\u003c/p\u003e\n\u003ch2\u003eAdding combat\u003c/h2\u003e\n\u003cp\u003eFor combat, I kept the actor\u0026#39;s driving features and movement outputs frozen while training the weapon outputs and critic. Weight checks verify that the protected parameters stay unchanged. Full races are still necessary, because a new weapon decision can change the situation the driver has to deal with.\u003c/p\u003e\n\u003cp\u003eFinishing position matters most in the reward, with limited extra credit for effective combat and useful pickups or pads. Firing more shots doesn\u0026#39;t earn a bonus by itself.\u003c/p\u003e\n\u003cp\u003eThe experiments covered longer credit assignment, combat reward weighting and critic calibration. Calibration fits the value estimate to returns collected from the retained policy without changing the actor. For extra practice, a race can start from a saved state where the driver had reached an opponent while carrying an offensive weapon. Those chances can be sparse in a normal race.\u003c/p\u003e\n\u003cp\u003eThe latest configuration mixes ordinary races with \u003cstrong\u003eten-second combat opportunity windows\u003c/strong\u003e. These start from verified native states and collect fresh actions from the current policy. The end of a practice window isn\u0026#39;t counted as finishing or being eliminated from a race.\u003c/p\u003e\n\u003cp\u003eThe opponents include original XL drivers, a retained learned challenger and mixed fields. The training rotation uses Talon\u0026#39;s Reach, Sagarmatha, Gare d\u0026#39;Europa and Odessa Keys.\u003c/p\u003e\n\u003ch2\u003eThe latest four-hour run\u003c/h2\u003e\n\u003cp\u003eThis continuation started from the earlier combat run\u0026#39;s hour-two checkpoint. Four CPU workers added \u003cstrong\u003e6,639,616 decisions\u003c/strong\u003e, taking that checkpoint lineage from \u003cstrong\u003e32,651,264 to 39,290,880\u003c/strong\u003e. Other experiments have their own experience counts.\u003c/p\u003e\n\u003cp\u003eThe run retained the driving parameters and trained the weapon outputs and critic, using gamma \u003cstrong\u003e1.0\u003c/strong\u003e, GAE lambda \u003cstrong\u003e0.95\u003c/strong\u003e, combat reward scale \u003cstrong\u003e1.0\u003c/strong\u003e and \u003cstrong\u003e300-decision\u003c/strong\u003e opportunity windows. Hourly saves allowed intermediate checkpoints to be compared with the final one.\u003c/p\u003e\n\u003cp\u003eThe review covered six policies, each with 48 standard evaluation races, plus cross-play and confirmation panels. \u003ca href=\"https://research.google/blog/rliable-towards-reliable-evaluation-reporting-in-reinforcement-learning/\" rel=\"nofollow\"\u003eRLiable\u0026#39;s evaluation guidance\u003c/a\u003e is useful here: paired conditions, several seeds and the distribution of outcomes give more information than a single good race.\u003c/p\u003e\n\u003cp\u003eThe final checkpoint \u003cstrong\u003ewon 18 of 48 races and finished 31\u003c/strong\u003e in the standard panel. In the additional 24-race confirmation panel, it \u003cstrong\u003ewon 9 and finished 15\u003c/strong\u003e. The intermediate checkpoints had different strengths; the last save wasn\u0026#39;t simply better at everything.\u003c/p\u003e\n\u003cimg src=\"/content/images/2026/10/b1e12503f058-nullg-rl-full-preview.png\" alt=\"Gare d\u0026#39;Europa race from the final combat-training checkpoint\"\u003e\n\u003ch2\u003eThe recordings\u003c/h2\u003e\n\u003cp\u003eThe learned pilot is in slot 0 in the latest full race. The automatic cameras move around the field, so every ship shown in a close-up isn\u0026#39;t necessarily that pilot.\u003c/p\u003e\n\u003ch3\u003eGare d\u0026#39;Europa, complete race\u003c/h3\u003e\n\u003cp\u003eThis is the three-lap race from the final checkpoint of the four-hour continuation. It runs for about three minutes, with native race audio and cinematic cameras.\u003c/p\u003e\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/__kpDHYbsCg\" frameborder=\"0\" allowfullscreen=\"\"\u003e\u003c/iframe\u003e\n\u003cp\u003e\u003ca href=\"https://youtu.be/__kpDHYbsCg\" rel=\"nofollow\"\u003eWatch the full Gare d\u0026#39;Europa race on YouTube\u003c/a\u003e.\u003c/p\u003e\n\u003ch3\u003eRacing and combat Short\u003c/h3\u003e\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/lTbLUj1QKt4\" frameborder=\"0\" allowfullscreen=\"\"\u003e\u003c/iframe\u003e\n\u003cp\u003e\u003ca href=\"https://youtu.be/lTbLUj1QKt4\" rel=\"nofollow\"\u003eWatch the racing and combat Short\u003c/a\u003e.\u003c/p\u003e\n\u003ch3\u003eEarlier training: update 3\u003c/h3\u003e\n\u003cp\u003eThe earlier upload, “NullG - 8 hour AI driving training 3 (reinforcement learning)”, shows another point in the driving work.\u003c/p\u003e\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/fkUTwW2p_8k\" frameborder=\"0\" allowfullscreen=\"\"\u003e\u003c/iframe\u003e\n\u003cp\u003e\u003ca href=\"https://youtu.be/fkUTwW2p_8k\" rel=\"nofollow\"\u003eWatch training update 3\u003c/a\u003e.\u003c/p\u003e\n\u003ch3\u003eEarlier training: update 2\u003c/h3\u003e\n\u003cp\u003eThe preceding upload is “NullG - 8 hour AI driving training 2 (reinforcement learning)”.\u003c/p\u003e\n\u003ciframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/9ddeaoyyIY4\" frameborder=\"0\" allowfullscreen=\"\"\u003e\u003c/iframe\u003e\n\u003cp\u003e\u003ca href=\"https://youtu.be/9ddeaoyyIY4\" rel=\"nofollow\"\u003eWatch training update 2\u003c/a\u003e.\u003c/p\u003e\n\u003cp\u003eThe next experiments will build on these comparisons, working on weapon decisions while checking that the driving still holds up across the four circuits. The \u003ca href=\"https://projectable.me/projects/nullg/reinforcement-learning/\" rel=\"nofollow\"\u003ereinforcement-learning page\u003c/a\u003e has more detail on the engine integration.\u003c/p\u003e\n","excerpt":"What changed in NullG's pilot training, from steering and Turbo to weapons, and what the latest four-hour combat run managed.","metaTitle":"Teaching NullG pilots to drive and use weapons","metaDescription":"What changed in NullG's pilot training, from steering and Turbo to weapons, and what the latest four-hour combat run managed.","status":"published","featured":false,"publishedAt":"2026-10-20T13:21:42Z","authorId":"190160de54f2192bd5b8734ec8808086","authorName":"Daniel Chote","authorSlug":"daniel","heroMediaId":"2de6b37e96d56188d9b74019123198cd","projectId":"ae4bdb66da6dc1729a263a475e1cdbe3","tags":[{"id":"10a974c0a281d4fcf3d44f5ef8dae079","name":"NullG","slug":"nullg"},{"id":"1c9e8a1339a8f8b797767e91a9673016","name":"Game","slug":"game"},{"id":"60103f47eb65b46c6138b9d6fc7b17d0","name":"Machine Learning","slug":"machine-learning"}],"createdAt":"2026-10-20T13:10:40Z","updatedAt":"2026-10-26T18:09:37Z"},{"id":"39b4d054f7e0bce8c6537f8d80a0a2e3","siteId":"projectable","slug":"msi-gtx-1050ti-4g-oc-the-poor-mans-1u-inference-gpu","title":"MSI GTX 1050Ti 4G OC - The poor man's 1u inference GPU","markdown":"At [BrandYourself](http://brandyourself.com), we are diving deep in to Machine Learning to enhance some of our upcoming features. The issue I have with the current ML battlefield is that cloud based GPU instances are still relatively new and very expensive. When you start adding high availability requirements, as well as potential concurrent load, it gets prohibitively expensive quickly.\n\n### The Solution\n\nI set out with the goal to build my own 1u low power server with an NVIDIA GPU. I found a good low power server candidate from [Supermicro](https://www.supermicro.com), the [5018D-FN4T](https://www.supermicro.com/products/system/1u/5018/sys-5018d-fn4t.cfm) 8 core D-1541 Intel® Xeon® processor. It has a 200W power supply and PCI-E 3.0 x16. It's not the cheapest platform option, but it has great performance: m.2 for storage and a full x16 PCI-E 3.0 slot.\n\nFor the GPU it took a lot of looking at photos to determine a card that could be easily modified to fit in a 1u chassis and still maintain an adequate cooling solution. With the launch of the GTX 1050Ti from NVIDIA, I was excited to see MSI launch a product that seemed like it may be my perfect fit. The [GEFORCE® GTX 1050 TI 4G OC](https://www.msi.com/Graphics-card/GeForce-GTX-1050-Ti-4G-OC.html) is a bus powered (75W) 4G 10 series GPU with a stock cooler that has the heatsink fins running end to end on the card. \n4G of ram means we can run 4 workers on the card (our inference model footprint for image processing right now is ~850Mb) and still have a small buffer incase something grows. \nThe 75W power consumption over the PCI bus means no extra power cables required and a lower expected running temperature. \nThe stock heatsink on the card already has end to end fins, so we have the ability to place a 1u high RPM fan at the end of the card and maintain a decent level of airflow over the heatsink.\n\n\n\n### Making it fit\n**WARNING: This WILL void your GPU's warranty...** \n\u003cp align=\"center\"\u003e\n![](/content/images/2017/04/IMG_0530.JPG)\n\u003c/p\u003e\nAs you can see, the card is already pretty small and the DVI port is in the first row. This means that we can easily cut the bracket and not make any physical modifications to the card itself (such as desoldering or brute force cutting of the DVI port).\n\nDisassembling is very straight forward.\n\n![](/content/images/2017/04/IMG_0505-1.JPG)\n \nThe 4 heatsink screws on the back of the card need to be removed.\n\n![](/content/images/2017/04/IMG_0506.JPG)\n\nCarefully disconnect the fan power connector and split apart. Avoid disturbing the heat transfer compound.\n\n![](/content/images/2017/04/IMG_0507.JPG)\n\nThere are 4 screws holding the fan shroud to the heatsink. Remove those.\n\n![](/content/images/2017/04/IMG_0508.JPG)\n\nPlace the heatsink back on (keeping the same orientation as before) and fasten in place with the original screws (the ones with the springs).\n\n![](/content/images/2017/04/IMG_0509.JPG)\n![](/content/images/2017/04/IMG_0510.JPG)\n\nNext, locate and remove the screws holding the mounting plate.\n\n![](/content/images/2017/04/IMG_0511.JPG)\n![](/content/images/2017/04/IMG_0512.JPG)\n\nAs you can see, this card is very slim.\n\n![](/content/images/2017/04/IMG_0513.JPG)\n![](/content/images/2017/04/IMG_0515.JPG)\n\nWith a pair of pliers, carefully straighten the board mounting support so that the plate will sit flush on a surface.\n\n![](/content/images/2017/04/IMG_0516.JPG)\n![](/content/images/2017/04/IMG_0517.JPG)\n\nApply some masking tape and draw your cutting line. Single slot plates are typically 18mm wide.\n\n![](/content/images/2017/04/IMG_0518.JPG)\n\nWhile using appropriate safety equipment, cut along your line using a bandsaw (a hacksaw would probably work, too).\n\n![](/content/images/2017/04/IMG_0519.JPG)\n\nFile off the sharp edges.\n\n![](/content/images/2017/04/IMG_0520.JPG)\n\nHooray, we have a single slot mounting plate!\n\n![](/content/images/2017/04/IMG_0521.JPG)\n\nBend back the board mounting support.\n\n![](/content/images/2017/04/IMG_0522.JPG)\n![](/content/images/2017/04/IMG_0523.JPG)\n\nNow we have a low profile GPU, without heavy modification or destruction of the PCB/components!\n\n![](/content/images/2017/04/IMG_0525.JPG)\n![](/content/images/2017/04/IMG_0526.JPG)\n![](/content/images/2017/04/IMG_0527.JPG)\n\nAs you can see, it mounts nicely inside this 1u chassis, and a 1u high RPM fan can be placed behind it. I mount the fan in place with some high strength double sided tape.\n\n![](/content/images/2017/04/Screen-Shot-2017-03-08-at-2-27-00-PM.png)\n\nThe GPU runs perfectly on ubuntu with the latest NVIDIA linux drivers and latest CUDA and CUDNN libraries. If you have any questions, drop me a comment below or tweet at me on twitter [@dchote](https://twitter.com/dchote).","html":"\u003cp\u003eAt \u003ca href=\"http://brandyourself.com\" rel=\"nofollow\"\u003eBrandYourself\u003c/a\u003e, we are diving deep in to Machine Learning to enhance some of our upcoming features. The issue I have with the current ML battlefield is that cloud based GPU instances are still relatively new and very expensive. When you start adding high availability requirements, as well as potential concurrent load, it gets prohibitively expensive quickly.\u003c/p\u003e\n\u003ch3\u003eThe Solution\u003c/h3\u003e\n\u003cp\u003eI set out with the goal to build my own 1u low power server with an NVIDIA GPU. I found a good low power server candidate from \u003ca href=\"https://www.supermicro.com\" rel=\"nofollow\"\u003eSupermicro\u003c/a\u003e, the \u003ca href=\"https://www.supermicro.com/products/system/1u/5018/sys-5018d-fn4t.cfm\" rel=\"nofollow\"\u003e5018D-FN4T\u003c/a\u003e 8 core D-1541 Intel® Xeon® processor. It has a 200W power supply and PCI-E 3.0 x16. It\u0026#39;s not the cheapest platform option, but it has great performance: m.2 for storage and a full x16 PCI-E 3.0 slot.\u003c/p\u003e\n\u003cp\u003eFor the GPU it took a lot of looking at photos to determine a card that could be easily modified to fit in a 1u chassis and still maintain an adequate cooling solution. With the launch of the GTX 1050Ti from NVIDIA, I was excited to see MSI launch a product that seemed like it may be my perfect fit. The \u003ca href=\"https://www.msi.com/Graphics-card/GeForce-GTX-1050-Ti-4G-OC.html\" rel=\"nofollow\"\u003eGEFORCE® GTX 1050 TI 4G OC\u003c/a\u003e is a bus powered (75W) 4G 10 series GPU with a stock cooler that has the heatsink fins running end to end on the card.\n4G of ram means we can run 4 workers on the card (our inference model footprint for image processing right now is ~850Mb) and still have a small buffer incase something grows.\nThe 75W power consumption over the PCI bus means no extra power cables required and a lower expected running temperature.\nThe stock heatsink on the card already has end to end fins, so we have the ability to place a 1u high RPM fan at the end of the card and maintain a decent level of airflow over the heatsink.\u003c/p\u003e\n\u003ch3\u003eMaking it fit\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eWARNING: This WILL void your GPU\u0026#39;s warranty...\u003c/strong\u003e\u003c/p\u003e\n\u003cp align=\"center\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0530.JPG\" alt=\"\"\u003e\n\u003c/p\u003e\nAs you can see, the card is already pretty small and the DVI port is in the first row. This means that we can easily cut the bracket and not make any physical modifications to the card itself (such as desoldering or brute force cutting of the DVI port).\n\u003cp\u003eDisassembling is very straight forward.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0505-1.JPG\" alt=\"\"\u003e\n\u003cp\u003eThe 4 heatsink screws on the back of the card need to be removed.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0506.JPG\" alt=\"\"\u003e\n\u003cp\u003eCarefully disconnect the fan power connector and split apart. Avoid disturbing the heat transfer compound.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0507.JPG\" alt=\"\"\u003e\n\u003cp\u003eThere are 4 screws holding the fan shroud to the heatsink. Remove those.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0508.JPG\" alt=\"\"\u003e\n\u003cp\u003ePlace the heatsink back on (keeping the same orientation as before) and fasten in place with the original screws (the ones with the springs).\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0509.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0510.JPG\" alt=\"\"\u003e\n\u003cp\u003eNext, locate and remove the screws holding the mounting plate.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0511.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0512.JPG\" alt=\"\"\u003e\n\u003cp\u003eAs you can see, this card is very slim.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0513.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0515.JPG\" alt=\"\"\u003e\n\u003cp\u003eWith a pair of pliers, carefully straighten the board mounting support so that the plate will sit flush on a surface.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0516.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0517.JPG\" alt=\"\"\u003e\n\u003cp\u003eApply some masking tape and draw your cutting line. Single slot plates are typically 18mm wide.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0518.JPG\" alt=\"\"\u003e\n\u003cp\u003eWhile using appropriate safety equipment, cut along your line using a bandsaw (a hacksaw would probably work, too).\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0519.JPG\" alt=\"\"\u003e\n\u003cp\u003eFile off the sharp edges.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0520.JPG\" alt=\"\"\u003e\n\u003cp\u003eHooray, we have a single slot mounting plate!\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0521.JPG\" alt=\"\"\u003e\n\u003cp\u003eBend back the board mounting support.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0522.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0523.JPG\" alt=\"\"\u003e\n\u003cp\u003eNow we have a low profile GPU, without heavy modification or destruction of the PCB/components!\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0525.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0526.JPG\" alt=\"\"\u003e\n\u003cimg src=\"/content/images/2017/04/IMG_0527.JPG\" alt=\"\"\u003e\n\u003cp\u003eAs you can see, it mounts nicely inside this 1u chassis, and a 1u high RPM fan can be placed behind it. I mount the fan in place with some high strength double sided tape.\u003c/p\u003e\n\u003cimg src=\"/content/images/2017/04/Screen-Shot-2017-03-08-at-2-27-00-PM.png\" alt=\"\"\u003e\n\u003cp\u003eThe GPU runs perfectly on ubuntu with the latest NVIDIA linux drivers and latest CUDA and CUDNN libraries. If you have any questions, drop me a comment below or tweet at me on twitter \u003ca href=\"https://twitter.com/dchote\" rel=\"nofollow\"\u003e@dchote\u003c/a\u003e.\u003c/p\u003e\n","excerpt":"","metaTitle":"","metaDescription":"","status":"published","featured":false,"publishedAt":"2017-04-06T14:51:54Z","authorId":"190160de54f2192bd5b8734ec8808086","authorName":"Daniel Chote","authorSlug":"daniel","heroMediaId":"90d32feb088156ee2cd4008f16ec17a7","projectId":"","tags":[{"id":"60103f47eb65b46c6138b9d6fc7b17d0","name":"Machine Learning","slug":"machine-learning"},{"id":"ea8ba22edf4d47c7325d88a051118500","name":"GPU","slug":"gpu"},{"id":"9b3625f33e46c5576a2ffb4c6ff5304c","name":"Supermicro","slug":"supermicro"},{"id":"614e3fef27b2bf48aec6759cb2c26bfb","name":"Low Power Server","slug":"low-power-server"},{"id":"39509342d1486390e4356778e40bda83","name":"BrandYourself","slug":"brandyourself"}],"createdAt":"2026-10-18T13:16:27Z","updatedAt":"2026-10-18T13:16:27Z"}],"site":{"id":"projectable","theme":"projectable","title":"Daniel Chote's Project Blog","description":"Software Development, Game Development, Hardware Hacking, Sim Racing, Multi-rotors and Robots!","timezone":"America/New_York","postsPerPage":10,"logoMediaId":"21557257b3f2ecc4b3fa875b808097a7","coverMediaId":"6d7e99f0997d13aac98da38677ca7064","copyright":"© Daniel Chote","twitterUrl":"https://twitter.com/dchote","linkedinUrl":"https://www.linkedin.com/in/dchote","googleTagId":"G-07GJBQVM38","hosts":[{"host":"new.projectable.me","canonical":false},{"host":"projectable.me","canonical":true},{"host":"www.projectable.me","canonical":false}]},"tag":{"description":"","name":"Machine Learning","slug":"machine-learning"},"tagId":"60103f47eb65b46c6138b9d6fc7b17d0","total":3}

Machine Learning

Software Development, Game Development, Hardware Hacking, Sim Racing, Multi-rotors and Robots!

NullG updates: the XL fleet, shields and a first Pulse race
I've been working on the rebuilt XL ships, shields, split-screen and learning pilots, plus the first playable Pulse preview and more PSP imports.
Teaching NullG pilots to drive and use weapons
What changed in NullG's pilot training, from steering and Turbo to weapons, and what the latest four-hour combat run managed.
MSI GTX 1050Ti 4G OC - The poor man's 1u inference GPU
At BrandYourself , we are diving deep in to Machine Learning to enhance some of our upcoming features. The issue I have with the current ML battlefield is that