No training and no H100s yet — agents iterate on code logic alone, optimizing heuristic policies for performance in simulation.
The reference task runs this improvement loop on the MuJoCo - Ant environment - (Gymnasium).
+This stage ships nine tasks — each runs the same improvement loop in its own simulated environment.
+Improve a programmatic Ant locomotion policy under a fixed autonomous research budget.
+ simulation_heuristics_ant_v1 ↗ +Mirrors the simulation-heuristics Ant v1 reference task; snippets are illustrative.
+Code snippets in the nodes mirror the Ant v1 reference task; all nine task packages live on GitHub.
A dome torso on four two-segment legs — 8 actuated joints in the MuJoCo physics engine. Episodes return a score the agent can measure.
gymnasium.farama.org — Ant ↗"},
+ {id:"simulation_heuristics_halfcheetah_v1",env:"MuJoCo HalfCheetah",lbl:"HalfCheetah",img:"assets/env-halfcheetah.png",
+ d:"Improve an interpretable HalfCheetah controller under a fixed autonomous research budget.",
+ start:"an asymmetric CPG/PD baseline — weaker than the reference",
+ pop:"A planar cheetah with six actuated joints in the MuJoCo physics engine — the goal is fast, stable forward running.
gymnasium.farama.org — HalfCheetah ↗"},
+ {id:"simulation_heuristics_pong_ram_v1",env:"Atari Pong (RAM)",lbl:"Pong · RAM",img:"assets/env-pong.png",
+ d:"Improve a programmatic Atari Pong controller that acts from the current 128-byte RAM state.",
+ start:"a weak, late-reacting RAM controller",
+ pop:"Classic Atari Pong via EnvPool — the controller never sees pixels; it acts from the live 128-byte RAM state.
envpool.readthedocs.io ↗"},
+ {id:"simulation_heuristics_breakout_ram_v1",env:"Atari Breakout (RAM)",lbl:"Breakout · RAM",img:"assets/env-breakout-ram.png",
+ d:"Improve a RAM-only programmatic Atari Breakout policy toward the 864-point article result.",
+ start:"a RAM-state starter policy — normalized to 0",
+ pop:"Atari Breakout via EnvPool, observed as the 128-byte RAM state — no pixels.
envpool.readthedocs.io ↗"},
+ {id:"simulation_heuristics_breakout_rgb_v1",env:"Atari Breakout (RGB)",lbl:"Breakout · RGB",img:"assets/env-breakout-rgb.png",
+ d:"Improve a pixel-only Atari Breakout policy toward the article's 864-point RAM-to-RGB transfer result.",
+ start:"a pixel-only starter policy — normalized to 0",
+ pop:"Atari Breakout via EnvPool, observed as rendered RGB frames only — the RAM-to-RGB transfer setting.
envpool.readthedocs.io ↗"},
+ {id:"simulation_heuristics_montezuma_v1",env:"Montezuma's Revenge",lbl:"Montezuma",img:"assets/env-montezuma.png",
+ d:"Improve a native-image Montezuma policy beyond a brittle open-loop replay.",
+ start:"a working but weak native-RGB controller",
+ pop:"EnvPool MontezumaRevenge-v5 from native RGB observations — the classic hard-exploration Atari game.
envpool.readthedocs.io ↗"},
+ {id:"simulation_heuristics_vizdoom_d1_v1",env:"ViZDoom D1 medikit",lbl:"D1 · medikit",img:"assets/env-vizdoom-d1.png",
+ d:"Improve a rendered-pixel ViZDoom D1 medikit policy under a fixed autonomous research budget.",
+ start:"a working but weak programmatic controller",
+ pop:"EnvPool D1Basic-v1 — a ViZDoom medikit scenario played from rendered screen pixels.
vizdoom.farama.org ↗"},
+ {id:"simulation_heuristics_vizdoom_d3_v1",env:"ViZDoom D3 battle",lbl:"D3 · battle",img:"assets/env-vizdoom-d3.png",
+ d:"Improve a screen-CV ViZDoom D3 battle policy under a fixed autonomous research budget.",
+ start:"a working but weak programmatic controller",
+ pop:"EnvPool D3Battle-v1 — a ViZDoom battle scenario driven by screen-CV features.
vizdoom.farama.org ↗"},
+ {id:"simulation_heuristics_atari57_v1",env:"Atari-57 suite",lbl:"Atari-57",img:"assets/env-atari57.png",
+ d:"Improve one aggregate 342-policy Atari-57 submission produced by 342 independently accounted heuristic-search trajectories.",
+ start:"an aggregate starter artifact spanning all 57 games",
+ pop:"The full Atari-57 suite via EnvPool — one programmatic policy slot for every game, observation mode, and independent repeat.
envpool.readthedocs.io ↗"}
+ ];
+ var ti=0,
+ tpCount=document.getElementById('tpCount'),tpEnv=document.getElementById('tpEnv'),
+ tpDesc=document.getElementById('tpDesc'),tpId=document.getElementById('tpId'),
+ loopStart=document.getElementById('loopStart'),envImg=document.getElementById('loopEnvImg'),
+ envLbl=document.getElementById('loopEnvLbl'),
+ pager=document.querySelector('.task-pager');
+ function renderTask(n){
+ ti=(n+TASKS.length)%TASKS.length;
+ var t=TASKS[ti];
+ tpCount.textContent="0"+(ti+1)+" / 09";
+ tpEnv.textContent=t.env;
+ tpDesc.textContent=t.d;
+ tpId.textContent=t.id+" ↗";
+ tpId.href="https://github.com/benchflow-ai/GenesisBench/tree/main/tasks/"+t.id;
+ loopStart.textContent=t.start;
+ CONTENT.env={t:"Environment — "+t.env,h:t.pop};
+ if(envImg.getAttribute("src")!==t.img){envImg.src=t.img;envImg.alt=t.env+", pencil illustration";}
+ envLbl.textContent=t.lbl;
+ if(current&¤t.dataset.k==="env"&&pop.classList.contains("show"))show(current);
+ }
+ if(pager){
+ document.getElementById('tpPrev').addEventListener('click',function(){renderTask(ti-1);});
+ document.getElementById('tpNext').addEventListener('click',function(){renderTask(ti+1);});
+ pager.addEventListener('keydown',function(e){
+ if(e.key==='ArrowLeft'){renderTask(ti-1);e.preventDefault();}
+ if(e.key==='ArrowRight'){renderTask(ti+1);e.preventDefault();}
+ });
+ renderTask(0);
+ setTimeout(function(){TASKS.forEach(function(t){var im=new Image();im.src=t.img;});},1500);
+ }
})();