Add Subway Surfers-style lane runner model comparison - #11
Open
Alex-Wengg wants to merge 9 commits into
Open
Alex-Wengg wants to merge 9 commits into
Alex-Wengg wants to merge 9 commits into
Conversation
Retry of the runner idea after the Flappy Bird trial ruled out per-frame control: one decision per row, five moves (left/right/jump/slide/stay), and every option labeled with its consequence over the three visible rows (the Tetris harness pattern). Option order rotates each row. - DecisionPolicy: the 13 model backends move out of FlappyBirdPolicy into a shared target (choose(state:instruction:options:)). FlappyBirdPolicy is now a thin wrapper; Flappy results are unchanged (GLiNER multilingual seeds 1-4 still 1 pipe / 10.98 s). - maximumOptions per model: Kai/Lex Choice packages have 3 candidate slots (marker_positions [1, 3]), NanoJev 4, so they cannot take the runner's 4-5 moves and now fail up front instead of in the worker. - LaneRunner engine (solvable by construction), LaneRunnerCheck headless runner (model, heuristic, random), LaneRunnerDemo app that shows the exact model input and counts late rows. Seeds 1-4, 200-row cap: heuristic 800/800; GLiNER base 184, GLiNER multilingual 160, Kev 0.6B 132 (66% first-position picks), GLiClass 70, GLiNER small 56, Jeff 30 = random 30, Laya 22, Verdict 16, LFM 12, Kev 0.5B 10. Every run ended on an option labeled "crash". Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Model loading (asset SHA-256 checks over hundreds of MB, then Core ML compilation) ran on the main actor, so the window stopped responding while "Loading..." showed. Load in a detached task instead. Launched from a terminal the unbundled app also needs a regular activation policy to take focus. LANE_RUNNER_MODEL / LANE_RUNNER_SEED / LANE_RUNNER_ROW_MS preselect a run that loads and starts without clicks, for demos. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Replaces the flat Canvas board with a third-person SceneKit view: subway trains and rail track from Kenney's Train Kit; character, gold coins, low fences, overhead beam, and trees from the Platformer Kit (both CC0; license files bundled under Resources/Kenney, ~500 KB of OBJ/MTL/colormap). Rows slide toward the camera over each row interval; the runner eases between lanes, jumps, slides, and falls on a crash. Game rules, model inputs, and results are unchanged. Also delays the heuristic autostart so the controller picker's reset cannot cancel it, and widens the window for the 3D view. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Live TIME metric (pauses excluded) beside distance and coins, and the final time on the run-over card. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Kenney's poles model is a solid log palisade, so the high bar looked like a wall rather than something to slide under. Draw two thin posts and a red/white beam at chest height with open space below, and drop the unused poles/platform models. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Lane movement, jump height, and slide squash ran on one node, so each row's lane move snapped y back to 0 and the scale reset cut jumps and slides short. Split the runner into lane / hop / stride / body nodes so the motions no longer fight. Also retime the track: the passed row now slides from half a row ahead to half a row behind, crossing the runner mid-row where a sine jump peaks and a smoothstep slide is lowest (previously the obstacle arrived after the jump had landed). Lane changes ease with a small lean, the runner bobs while running, and a crash stops at the obstacle before falling. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When the runner crosses a coin in its lane, at the moment the coin reaches it: the coin pops up, spins, shrinks and fades; a short burst of soft gold sparks plays; a billboarded "+1" rises and fades; and the runner does a quick squash-and-settle bounce. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Like Subway Surfers and Temple Run, runs now ramp up: - Track density (engine): a 30-row easy opening with the original rules, then over 240 rows open breather rows disappear (up to 65% of the time) and blocked lanes rise (trains 40 -> 55%, open 20 -> 5%). Back-to-back obstacle rows are only generated when every lane passable in the first can legally enter the second (same lane with the right move, or an adjacent open lane), so tracks stay survivable: the heuristic clears 600 rows on seeds 1-12. - Speed (app): each row is 0.5% faster, floored at 40% of the starting interval; the panel shows the current interval and multiplier. Tracks changed, so results are re-run (seeds 1-4, 300-row cap): GLiNER base 129, GLiNER multilingual 62, GLiNER small 48, Laya 28, GLiClass 20, LFM 20, random 14, Kev 0.6B 12, Verdict/Jeff/Kev 0.5B 10. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #10. Retries the runner idea with per-row decisions instead of per-frame control. Each of the 5 moves is labeled with its consequence, following the Tetris harness pattern.
DecisionPolicy: the 13 model backends are now a shared target;FlappyBirdPolicywraps it (Flappy results unchanged).LaneRunnerengine (survivable by construction),LaneRunnerCheck(model / heuristic / random),LaneRunnerDemoapp.LANE_RUNNER_MODELandLANE_RUNNER_SEEDstart a run with no clicks.Seeds 1–4, 300-row cap (1,200 max):
Every run ended when the model picked an option labeled
crash. Only GLiNER base regularly got past the easy opening. Details inSources/LaneRunnerDemo/README.md.🤖 Generated with Claude Code