Φ
Universal remote · robot footage

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.

recorded 2026-08-11T23:53:00Z · 8 seeds scouted per surface, the winner filmed · 4,731 filmed frames
Live control room · real recorded runs

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.

Ant · the brain
0
reward
step 0/702
Chaosalator board

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.

Probe fan
Motor command
Cumulative reward — whole run0.0
Motors
8
Sensors
105
Codes/beat
20
Probes
4680
Brain
397
Random
-18

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.

The Gait Floor

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.

Half Cheetah0
6 motors, live
score
4359.4
seed 0
random flail
-298.9
same body, same seed — 1000 steps
top speed
6.21 m/s
mean 4.69
body
HalfCheetah-v5
stock physics, untouched
words tuned
133
6 motors, 17 sensors
steps held
1000
no fall

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.

Every seed we scouted, unedited: seed 0: 4359.4 · seed 1: 4317.9 · seed 2: 4317.0 · seed 3: 4262.5 · seed 4: 4232.8 · seed 5: 4271.4
the whole program — five lines, running on every body above
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 reflex
Search never touches the structure. It only tunes the words: rest, swing, snap, reflex, tempo.
16
Control surfaces filmed
14/16
Beat the random baseline
24,078
Probe rollouts rewound
348d
Largest state space
CONTROL ROOMAntAnt-v5beats random
BRAIN · 360.4
RANDOM · -21.2
cumulative reward, this clip
take boardevery seed scored, the winner filmed
s42 324.6s7 262.6s13 49.6s101 173.4s2024 174.0s777 236.3s31337 344.7s5 360.4

Four legs, eight actuators. Random tips over; the hunt keeps it upright and moving for the whole clip.

probe fan
20 candidates scored from one frozen state. Gold is the locked one.
4D rotor
Tesseract spun in two planes, vertices driven by the live command.
motor command8 actuators
frame1/600
reward now
cumulative
lock1/200
candidates / lock20
lookahead15 ticks
commit3 ticks
actuators8d
state dims105d
probes this clip4,000
filmed seed5
Every number on this panel comes out of the recorded run in /robots/ant.json, the same file the video was rendered from. Probes are counterfactual: simulated from a frozen state and rewound, so they never appear in the footage.

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 surfaceBrainRandomDeltaLook / commitCandsRewind
Half CheetahHalfCheetah-v516.6-174.9+191.615/320yes
AntAnt-v5164.8-8.3+173.115/320yes
HopperHopper-v5207.510.5+197.115/315yes
Walker 2DWalker2d-v5-25.00.3-25.315/315yes
HumanoidHumanoid-v5117.0109.0+8.040/320yes
Humanoid StandupHumanoidStandup-v53965919341+2031815/320yes
SwimmerSwimmer-v535.4-6.6+42.015/312yes
Reacher ArmReacher-v5-10.5-43.5+33.010/215yes
Pusher ArmPusher-v5-87.3-154.4+67.210/215yes
Inverted Double PendulumInvertedDoublePendulum-v5141.744.4+97.315/312yes
Inverted PendulumInvertedPendulum-v5166.35.7+160.715/312yes
PendulumPendulum-v1-203.9-1143.5+939.520/312yes
AcrobotAcrobot-v1-404.7-500.0+95.325/310yes
Mountain CarMountainCar-v0-200.0-200.00.025/310yes
Cart PoleCartPole-v1500.022.0+478.020/310yes
Lunar LanderLunarLander-v3-74.0-135.9+61.810/312yes

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.