The Brain
The same remote that plays ARC-AGI-3 and Mario, pointed at physics. Every clip on this page is a real rollout recorded on our machine: no training, no gradients, no network. It freezes the world, probes a fan of candidate commands, rewinds, and locks whichever one moved the needle.
The brain, firing, driving a robot
Three of its best runs, filmed frame for frame. On the left is the actual robot the brain drove. Everything on the right is the machinery behind that motion at the same instant — the codes it fanned out, the one it locked, the motor command it sent, and the analog oscillator that manufactured them all. No policy was trained. Nothing is scripted.
Gold comet: the logistic×golden oscillator walking its chaotic attractor. Dots: the codes it fired this beat, blue→bright by score. The ringed one is what it locked.
A four-legged robot taught to walk by nothing but feedback. Every step is a code the brain fired, watched, and locked — no gait was ever programmed.
Recorded 2026-08-11 · best of 6 scouted seeds, filmed on seed 42 (42:397 · 7:332 · 13:103 · 101:243 · 2024:249 · 777:305) · 2,100 filmed frames · 23,930 probes across the three runs.
We stopped guessing and wrote the controller.
Everywhere else on this page the brain is blind — it gropes forward one frame at a time and scored 51.3 on HalfCheetah doing it. A body is not a puzzle. So here we did the honest thing and wrote an actual program: one heartbeat, one word per joint, and let the search tune only the words. Same body, same physics, 4359.4.
A flat-out sprint. Six motors, one shared phase, and a reflex term that keeps the spine from tumbling at speed. The random policy face-plants and slides backwards; this thing pins the throttle for the full thirty seconds.
theta += 2*pi * f * dt // one heartbeat for the whole body
a[j] = b[j] // where the joint rests
+ A1[j] * sin(theta + p1[j]) // the swing
+ A2[j] * sin(2*theta + p2[j]) // the snap
+ W[j] . tanh((obs - mu)/sd) // the reflexFour legs, eight actuators. Random tips over; the hunt keeps it upright and moving for the whole clip.
The robot wall
11 bodies, one loop, zero retraining. Click any tile to load it into the control room.
Classic control
Textbook problems on the same remote. Nothing in the loop changes, only how far ahead it looks.
Freeze and fan out
Snapshot the full physics state. Draw candidate commands from the chaosalator: logistic chaos at r = 3.99 mixed with a golden-ratio walk. Every candidate is tested from the identical frozen state, so the comparison is fair.
Watch the needle
Run each candidate forward through a lookahead window and score the return. One bit of device knowledge is supplied per surface: whether an episode ending means death or arrival.
Rewind and lock
Restore the snapshot, commit the winner for a few ticks, repeat. Receding horizon, no weights, no memory between episodes. The search is the entire mechanism.
Scoreboard
| Control surface | Brain | Random | Delta | Look / commit | Cands | Rewind |
|---|---|---|---|---|---|---|
| Half CheetahHalfCheetah-v5 | 16.6 | -174.9 | +191.6 | 15/3 | 20 | yes |
| AntAnt-v5 | 164.8 | -8.3 | +173.1 | 15/3 | 20 | yes |
| HopperHopper-v5 | 207.5 | 10.5 | +197.1 | 15/3 | 15 | yes |
| Walker 2DWalker2d-v5 | -25.0 | 0.3 | -25.3 | 15/3 | 15 | yes |
| HumanoidHumanoid-v5 | 117.0 | 109.0 | +8.0 | 40/3 | 20 | yes |
| Humanoid StandupHumanoidStandup-v5 | 39659 | 19341 | +20318 | 15/3 | 20 | yes |
| SwimmerSwimmer-v5 | 35.4 | -6.6 | +42.0 | 15/3 | 12 | yes |
| Reacher ArmReacher-v5 | -10.5 | -43.5 | +33.0 | 10/2 | 15 | yes |
| Pusher ArmPusher-v5 | -87.3 | -154.4 | +67.2 | 10/2 | 15 | yes |
| Inverted Double PendulumInvertedDoublePendulum-v5 | 141.7 | 44.4 | +97.3 | 15/3 | 12 | yes |
| Inverted PendulumInvertedPendulum-v5 | 166.3 | 5.7 | +160.7 | 15/3 | 12 | yes |
| PendulumPendulum-v1 | -203.9 | -1143.5 | +939.5 | 20/3 | 12 | yes |
| AcrobotAcrobot-v1 | -404.7 | -500.0 | +95.3 | 25/3 | 10 | yes |
| Mountain CarMountainCar-v0 | -200.0 | -200.0 | 0.0 | 25/3 | 10 | yes |
| Cart PoleCartPole-v1 | 500.0 | 22.0 | +478.0 | 20/3 | 10 | yes |
| Lunar LanderLunarLander-v3 | -74.0 | -135.9 | +61.8 | 10/3 | 12 | yes |
Mean return over 3 episodes per surface, random baseline run from the identical seeds. These means are unselected: no seed picking, every episode counted. The reels above are single hand-picked episodes, so their on-screen scores differ from these means.
What this footage does and does not prove
- The baseline is a random policy, not a trained agent. A tuned SAC or PPO policy beats these numbers on every locomotion task.
- The hunt gets to rewind. It burns 24,078 simulated probe rollouts across these 16 clips; random gets none. Where a simulator exposes no snapshot the loop degenerates to blind streaming and collapses, so a surface only appears here once we can actually freeze and restore it.
- The reels are cherry-picked and we show the cherries. Each surface was scouted across 8 seeds unrendered, and only the best-scoring take was filmed. Every rejected take is printed on its clip's take board, so you can see exactly how much of the reel is selection.
- Walker 2D loses the benchmark and Mountain Car ties at the floor. Walker's unselected 3-episode mean is -24.96 against random's 0.35 — its winning reel is one good seed, not a solved surface. Both stay on this page unedited rather than dropped.
- Probe steps do not consume the episode clock; committed steps do. Both sides are scored on committed steps only.
- Nothing is learned between episodes. Every clip starts as ignorant as the last one.