Operant lab · Subject R-01 · Robinhood Chain
We gave a virtual rat a brain made of two trained neural networks and taught it to press a lever. Its head is the mouse and the lever is the click. Next, it launches our coin on pons by itself, doing every one of the 11 clicks.Then it launched our coin on pons by itself, doing every one of the 11 clicks.
not a real rat brain a simulated body driven by artificial networks
01The subject
R-01 is a virtual rat. There is no animal and no camera: it is a physics simulation of a rat's body, and a computer brain has to control it.
Its body is DeepMind's open-source virtual rodent, a detailed model of a real rat's skeleton built for neuroscience research (the rat from Aldarondo et al., Nature 2024). It has 67 joints moved by 38 motors and weighs 0.34 kg.
It lives inside MuJoCo, a physics engine that works out gravity, friction and collisions every 2 milliseconds. Nothing is animated by hand: whenever the brain is in control, it decides 50 times a second where each motor should move its joint, and the physics does the rest.
That makes it a real balancing act. Push the wrong motors and it wobbles, slips or falls over, the way any body would.
Its home is an operant box: the classic lab chamber where rats learn that pressing a lever pays off. Ours has four walls, a floor and a spring-loaded lever hinged on the front wall.
02The brain
We never programmed the lever press or the head-steering. We built two artificial neural networks and let them learn by trial and error, the way a lab rat learns a lever: try something, get rewarded, repeat.
Numbers go in: what the rat can sense, like its joint angles and where the lever is. They flow through layers of simple units, and every connection has a dial that turns the signal up or down. Numbers come out: where each motor should move its joint.
R-01's two networks: 2,312 units, 643,072 dials.
At first the dials are random, and so are the movements. We score every attempt: points for a clean lever press and for standing well, a penalty for falling over. After each round of practice the dials are nudged towards whatever scored more.
This is reinforcement learning. The method we used is called PPO.
Each practice step is 20 milliseconds of simulated time. A network that starts from random dials needs a huge number of tries, so training runs many copies of the rat in parallel.
Stage 1The lever-press network
It feels 200 numbers about its body and the lever and drives all 38 motors. Its job: press the lever with a paw and stay standing. A press only counts if it is clean: a paw pushes the lever past 11.5°, the head and belly stay off it, and the rat is upright.
Then we tested the lever-press network on its own: 800 fresh attempts in four test suites. Every single time it made a clean press and was still standing 1.2 s later:
Stage 2The steering network
It sees 21 numbers: where the lit target is compared with the cursor and how big it is, how its head is turned, and whether it is standing upright. It drives the 4 neck motors that turn the head, plus one PRESS signal.
Its job: point the head so the cursor lands on the lit target, and press only when it's on it.
How the two work together
From the recorded training logs of the two networks in R-01's brain (not live): every row of the steering log, 160 of the lever log's 1,043.
Final training run, from 5 M to 35 M steps, under the strict clean-press rules. We kept the 32.5 M-step snapshot.
Practice had a difficulty dial that rose as it improved. The dip near 0.9 M is where training resumed from a saved checkpoint and the dial restarted from easy. By the end it had climbed back only to 0.32 of full difficulty, so the last value is on easier practice targets than the 2.96 it scored at full difficulty before the resume.
03The launch
R-01 launches our coin, Labrat ($LABRAT), on pons, a coin launchpad on Robinhood Chain, and it does every click itself. Here is how a rat uses a website.
The rig lights up the next thing to click, like the cue light in a lab box, and scrolls it into view.
R-01 turns its head and steers the cursor onto the lit target.
It presses the lever. A click inside the lit target reaches the page. A click anywhere else is ignored and never sent.
If it clicked into a text box, the rig types that box's text. The rat does every click, not the typing.
Between steps. The rig puts R-01 back in its standing start pose, and the brain rests while the rig types or scrolls. That is the jump you see in the replay.
Who does what. We choose the name, ticker, description, image and settings before the run, and a person presses Start. The rig types our text and, only if every check passes, signs the launch transaction once. The rat's brain makes all 11 clicks, and nothing reaches the page without one.
each one a lever press by the rat
LabratLABRATA step only moves on when the rat presses the lever with the cursor inside the lit target. A press anywhere else is ignored and never reaches the page, so the rat has to hit every one of the 11 targets itself.
pons builds the launch transaction. Before anything is signed, the rig decodes it and checks it field by field. Every check has to pass:
signed onceOnly then does the rig sign it, once. Anything that doesn't match stops the launch unsigned. In a dry rehearsal the rig runs every check and then refuses to sign, so nothing is created.
The result
Launching soon
Not launched yet. R-01 will launch it itself, live, on pons. Its contract, the launch transaction and the block appear here once it has.
01
ProblemAn early lever network found a loophole: flop the whole body onto the lever. The lever went down, so it scored, but that isn't a press.
FixWe closed it. Now a press only counts with a paw on the lever, the head and belly off it, and the rat upright.
02
ProblemSweeping the head down hard to reach low targets tipped the rat over.
FixWe limited how far it could pitch its head up and down. Turning sideways stayed free, for the small sideways moves small targets need.
03
ProblemThe lever network was trained to recover its balance for a moment after each press. Taking the body back too early made 6 of 10 test runs through the pons steps fall.
FixNow it keeps the body for 0.8 s after the click to finish recovering before steering resumes.
04
ProblemWe added a rest between steps, so the rat stands still while the rig types or scrolls. With the head limit, it could then no longer reach the last target: the small Confirm button low on the screen.
FixWith the rest in place the limit was no longer needed, so we removed it: 11 of 11, no falls, on 6 of 6 test runs.
04Proof
Everything the rat does can be checked: the brain it uses is fingerprinted on-chain, and the session re-runs bit for bit.
A
Before the run we compute a sha256 fingerprint over the whole brain: both networks, the physics scene and the code that runs them. The rig types it into the coin's description, so it goes on-chain inside the launch transaction. Change one dial and the fingerprint changes.
The coin's description
launched by a virtual rat: its trained brain steered the cursor and clicked every button. brain sha256 <the brain's 64-character fingerprint>
coversscene.xmlpolicy.ptsteer.ptenv.pycursor_env.pysteer_env.pysession.pyptload.py
B
Every session saves its commands, the brain's actions and every physics frame. The replay re-runs the same brain from the same start with the same commands, and prints MATCH only if every frame, every click and the session fingerprint come out identical.
$ python replay_session.py runs/<the launch session>
brain commit <the brain fingerprint> OK
frames identical
clicks 11 recorded, identical
recorded proof <the session fingerprint>
replayed proof <the same fingerprint>
MATCHExact on the recording machine. A different CPU or numpy build can round the networks' float maths differently, so every session records the Python, numpy and MuJoCo versions it used.
C
The code, both trained networks, the training logs behind the charts above and recorded sessions are in the public repo. Read the rules the rat trained under, check the fingerprint yourself, or run a replay.
LabratDevRH/labratgithub.com/LabratDevRH/labratruns/final/ the two networkstrain.py · env.py · steer_env.py the training and the rulesruns/lever_v3/log.jsonl · runs/steer_v1/log.jsonl the training logssession.py · replay_session.py the launch session and its bit-exact replay05The live panel
The badge in the corner of the panel at the top tells you which of two things you're seeing.
Badge legend: what each badge means when the panel shows it
Training runs many simulated rats in parallel, with random exploration switched on, and every so often saves a checkpoint: a snapshot of all the dials.
A small program next to the training takes the newest saved checkpoint and plays it in its own separate simulation, in real time at 25 frames a second, making its best move each time. Whenever training saves again, the rat you see gets the new dials, so you can watch it improve. The charts are the training log, row by row as it is written.
“LIVE” appears only while a training run is actually streaming. The rest of the time the panel says REPLAY, or STANDBY while no recording is loaded.
When nothing is training, you're watching a replay of a recorded launch session, named next to the badge: a dry rehearsal of the Labrat launch, where the rat does all 11 steps on the real pons page and the rig checks the transaction and then refuses to sign, so nothing is created. It is re-run from its saved log with the same brain, the same start and the same commands, so the lit targets and clicks are the real ones from that session.
Between targets the rig puts the rat back in its standing start pose, so it jumps, and the long waits while the rig types play fast-forward. It is a recording, not live. The readout beside it shows the recorded training logs of the two networks it uses.
Never a camera, never a real rat. The 3D rat is a drawing of the simulated body's pose, frame by frame, as the physics engine computed it. The screen on the chamber wall is a drawing for you too: it shows the cursor and the lit target. The rat never sees a screen or the pons page: its steering network is told where the target is as numbers.
Reading the charts. An attempt (an “episode” in the training log) is one practice try: about 5 s at the lever for the lever-press network, up to 15 s and 4 targets for the steering network. “Falls” is the share of attempts that ended with the rat falling over.