DAISYCHAIN · LOCAL SIM · RECORDED RUNS

robot learning arcade

Play a training mission. Record what happened. Help a roboticist and their robot learn from every run.

connecting locally
mood · curious 29 DOF
REFERENCE ROBOT · UNITREE G1

G1 humanoid in MuJoCo

The compact, nimble G1 keeps robot meaning attached to every run.

SIM + DATA CONTRACT head: pat · ↔ drag · feet: boost · body: dance · 2×: tickle · r: surprise · j: joke
physics
MuJoCo 3.10.0
model
29-DoF MJCF
joints
12 legs · 3 waist · 14 arms
G1 evidence
25 Hz · 100D observation · 29D action
Arcade boundary

The live games use a simpler 4D planar MuJoCo controller at 20 Hz. The G1 pilot is visual reference—not native whole-body G1 actuation.

Policy Lab checks 29D position targets and 29D torque commands as different control modes.
01training arcadelego target · episode dock
ROOM 1 · PICK A TASK

Choose your mission

you are inside mujoco—the same physics engine professionals use to train robots. each mission creates a digital simulation episode: brain + body food that helps embodied ai learn what to do and how to move.

03easy-bake policy labflag dock 04launch padevidence dock
co-op learning analytics loop

help your roboticist friend’s robot learn from every run.

play training games. sort the multi-dataset run inbox. catch policy red flags. return launch evidence so the robot team can stage the next policy faster.

  1. 01training gameplay + record
  2. 02run inboxsort many datasets
  3. 03red flagscheck policy output
  4. 04robot teamreturn launch evidence
methods + limits

preserve provenance; map schema, missingness, units, timing, frames, and action meaning; compare matched state/action traces; cluster correlated anomalies; keep uncertainty and abstention visible. this is hypothesis-generating decision support—not retraining, hardware approval, or a stability promise. the experienced owner judges any staged physical test. faster means fewer avoidable loops—not skipping hardware gates.

02live run mailroomlego target · label dock
V2 · LABEL THE RUN

Say what the controls mean.

Save the run with its units, frame, and control mode. The next room can compare it without guessing.

send to the harmonizer →
run labelCHECK ONCE
source family
arcade planar task
action units
normalized planar intent
reference frame
seeded world frame
command mode
teleoperation intent
episode_000000279 framesneeds a run label
episode_0000011,759 framesneeds a run label

Rule: the owner checks one label. The next matching run gets the same label as a suggestion.

Choose a playable mission to begin locally.

Live robot game arena

World—
Seed—
Time0.0s
Score1000
G1 PILOTshot-1 avatar · input-linked
MUJOCO · LIVE FORCES
SEEDED PROCEDURAL TASK
JIMINY · LIVEgoal + observation → local response
I’ll watch the goal and help you choose the next small move.

Goal critic: waiting for grounded simulator evidence.

OBSERVE
Waiting for the local simulation… Caption evidence will appear here.
Teleop View
Deterministic local instruction router—not a trained policy.
W ASD

Drag the humanoid, click a destination, or steer with WASD. Every game records the same bounded [x, y, rotate, gripper] action. Use Q/E to rotate and Space to grasp or release. A gamepad also works. Press R to replay the seed, N for the next one, P to pause, and Esc for a safe stop.