Survival of the quickest.
Run, step or jump over 25 cm square blocks, pre-placed every 10 ±2 metres. Watch your champion race alongside its challengers.
One race at this duration. Changes pause with the same champion; Resume continues.
Playback changes viewing time.
One frozen champion and parallel practice bodies sharing a learning actor.
Loading shared champion…
Recover an older saved champion
These are archived runners from the previous version. Choosing one makes it your single current champion for all settings.
Blocks are 25 × 25 cm, about half standing knee height. The entire course is laid out before the start, with random gaps of 8–12 metres, including the first block. Every runner gets the same fixed positions as native (course 420). Importing a runner keeps this course. The two engines can still produce different movement; use the same race duration when comparing results. Blocks never appear, move or disappear during a run: reach and pass each one to get to the next. Terrain sensors, contacts and forces include them. The champion physically reruns every race; during practice it runs a separate demonstration.
Adaptive search
KEEP THE BEST · VARY THE CHILDRENHybrid runs eight mutation races, then one PPO cycle. Three independent branches breed for reliable performance. One branch per round gets matched disturbance checks; only the clean race decides the title.
Changing this slider applies its level to a new batch immediately and pauses. The champion stays the same. Mutation size controls how far selected weights and biases move, in either direction. A value may remain exactly unchanged; no minimum nudge is forced. After a win, up to half the child slots also try extending that successful change by 1%, 2%, 5%, 10%, 20%, 50% and 100%.
Independent choice for every unlocked weight and bias in each random child. The default 10% selects about one in ten, leaving about 90% untouched. At 0%, all stay untouched. At 100%, all are selected, but a zero or unrepresentably small adjustment can still leave a value unchanged. Applies to random mutations; winning-change extensions and PPO learning are separate. Changes pause with the same champion and prepare a new batch.
Champion children narrow to 0.000000000001%, widening after 60 failed races there. Each independent branch has its own range, with a 0.001% floor and periodic widening. Lifetime title defences keep counting. Three quarters of random children use sampled-command scaling; one quarter explores the full range.
Periodic actor–critic training
PRACTICEControls gradient step size. PPO clips its objective and stops actor updates when sampled policy change becomes large. Bigger steps can lose useful behaviour; the champion stays protected.
Maximum Gaussian action noise, before tanh and smooth motors. Training starts at a quarter of this level and adapts after validation. Clean title races have no exploration noise.
Practice mixes blocks starts, short episodes from recorded points in the champion’s stride, and occasional fallen starts. Bodies continue across batches. The first two updates train the critic alone. Matched lessons begin at up to four seconds and lengthen as balance improves; damaging updates roll back to its best checkpoint while the critic keeps learning. Title races always use the same blocks start and full duration.
Explicitly reseeds the learner from your current champion and starts a fresh critic. Your champion is unchanged.
Refresh restores your automatically saved champion. Start fresh deliberately replaces that saved champion with champion #2688 and resets its version counter. Export a runner file to keep it. Download champion #2688 JSON to keep it separately. This is Jeff’s exported generation-11 brain, preserved at full precision. Reliable recovery and obstacle skills are still training goals, not established abilities.
Learning focus
CHOOSE WHAT LEARNS0% freezes a group’s weights and biases; 100% uses the full mutation range and actor optimizer step. Shared brain covers both hidden layers and their memory inputs. Output groups cover the 128 connections and bias feeding that command.
Frozen weights do not freeze a limb: changed movement changes its sensor inputs. Enable the shared brain to learn new coordination. Changes pause training and discard unfinished experience; they keep the champion. Cadence controls the rhythm clock, not a uniform increase in running speed.
Runner network
| From | Input | Weight | Input × weight |
|---|
102 inputs (70 sensors + 32 feedback traces) feed 128 → 128 hidden neurons and 13 outputs: 31,104 connections and 269 biases, or 31,373 actor parameters. Twelve outputs drive the waist, neck and both hips, knees, ankles, shoulders and elbows; the last controls a rhythm clock. The first 20 memory traces use 0.1-second smoothing; the other 12 use 0.5 seconds. Memory resets at each start. Current body and joint angles, velocities, foot loads, traction, heel/forefoot contacts and head/torso/pelvis contacts are sensed. Forward ground-height probes and obstacle distance, height and width are also available; they sense the square blocks used in races and training. Selected lines show signed weights and exact thickness scaling; the faint background is grouped for display speed. All calculations and exports use native 64-bit numbers; displayed decimals do not round the network.
Runner collection · share & swap 0 / 50 SAVED
Send a runner file or code to a friend, then bring theirs into any slot. Opening this panel pauses the race. Replacing a slot restarts the current runners at the start; arrange as many slots as you like, then race.
01 / Choose a runner
Saved runners stay in this browser after reload. Export a file to keep a backup or share. Your best replay keeps the exact starting controller that earned the record in this session. Track exports also use the race’s starting checkpoint. Practice-slot exports contain the current actor, without its critic or exploration noise.
02 / Bring a friend
Classic runners join your saved collection first. Use a version 18 feet-and-ankles runner, or import an aelife5 portable brain to open its training gym. The new rigid body has different joints and mechanics; previous versions are incompatible. Refresh restores the last automatically saved champion. This collection is separate: its runners remain available for manual import into any slot.
03 / Pick their place
Replacing a slot restarts the current runners.
The selected target decides the title at the end of this race.
Upright streak & distance by heat
WAITING FOR A RESULTRunner #1
GAIT INSPECTORRecent champion evaluations
NO COMPLETED HEATSHow mutation, learning and evaluation work
01 / Race and learn.
Hybrid runs eight mutation races, then one PPO cycle. Children rotate between the title holder and three independently breeding parents. Each branch selects for half mean plus half worst performance across clean, forward-nudged and backward-nudged runs. All start in the same blocks pose. Only one branch is checked per round to bound the extra work. Branch parents can improve without beating the record; their children can still win the ordinary title. Each range narrows independently. Winning title changes also produce directional probes. Focus locks apply to every child and PPO.
02 / Protect the title.
Only noise-free evaluations can replace the champion. Every entrant, including the reigning champion, physically runs from the same blocks pose on the same pre-placed course. The champion also repeats a noise-free demonstration during practice and reliability checks, without contributing training samples or changing the title. Balanced scores positive net forward travel while upright multiplied by the fraction of the race spent upright, counting every upright spell. Total distance and longest upright streak remain visible separately. Exact ties retain the champion. Changing duration, target or controls preserves the current champion and pauses the session. Practice bodies can explore worse behaviour without taking the title.
03 / Learn to recover.
PPO practice includes blocks starts, short sections of the champion’s recorded stride and occasional face-up or face-down lessons lasting at least 16 seconds. Upright rewards recognise recovery, and flight bonuses remain small compared with forward progress. The learner has its own short-stride lesson score and best actor/optimizer checkpoint. Each validation uses one blocks start and two fixed recorded stride states, with gentle nudges in the latter two. Lessons start at up to four seconds and grow by two seconds when all three sustain balance, up to the selected race time. Damaging updates restore that checkpoint, reduce noise and step size, and retain the critic’s experience. A new title holder does not automatically erase the learner. Recovery and obstacle skills remain training goals.
04 / Feet, ankles and muscles.
Thirteen rigid segments total 75 kg. Physics uses two 240 Hz substeps per 120 Hz control tick. Twelve joints have physical angle stops, direction-dependent strength, speed-dependent torque and approximate power limits. Knees fold towards the back of the thigh through about 149°; opposite bending is blocked, within the contact solver’s small tolerance. Heel and forefoot contact provide leverage for ankle push-off. Ground collision, gravity and motor torque produce movement; there is no upright force or automatic get-up animation. This remains a simplified 2D model, not a validated human musculoskeletal simulation.
05 / Keep your champion.
The current champion is saved automatically in this browser, including its exact brain, version and course. Refresh restores it paused with new challengers; PPO practice starts again from that brain. Start fresh deliberately replaces the saved champion with champion #2688. Explicit collection entries are separate, and exports give you a portable backup. One champion serves every target and duration.
06 / Browser limits.
The worker runs at every speed, including in inactive tabs while the browser permits it. Turbo uses available computing time and yields for controls. Device sleep or browser suspension can stop training. More racers and a larger brain reduce throughput. The actor has two 128-neuron layers plus feedback memory; the critic uses two 64-neuron layers. Neither size nor long training guarantees improvement. Planck.js licence.