Robot Painter — embodied training data

Live episodes from an autonomous drawing robot on the muretai network, recorded in an RLDS-compatible schema (rlds-muretai/0.2). Every episode is attributable end to end: the requester's Ed25519-signed task message is hash-bound into the episode, and the finished episode's hash rides back inside the robot's signed reply — so a third party can verify who asked, what the robot did, and that nothing was altered, offline, from signatures alone.

Episodes

when	steps	mode	result

Newest 50. Each row is one task turn; failures (emergency stops, timeouts) are kept as negative examples.

Select an episode to inspect it.

What a step looks like

Recorded at the controller's native 12 Hz. Each step pairs the world state with the action taken in it — the supervision a world model trains on:

observation: { canvas_ref   sha256 of a 128×128 PNG — the canvas at this step
               joint_pos    [yaw, shoulder, elbow]  (rad)
               pen_down     bool (measured, not commanded)
               ts_ns        robot-clock timestamp }
action:      { joint_target [yaw, shoulder, elbow]  (rad)
               pen_cmd      0 hold · 1 down · 2 up
               source       0 autonomous · 1 human teleop
               ts_ns }

Verify, then consume

1 · Fetch
Episodes and image blobs are content-addressed and served here:

GET /api/episodes
GET /api/episode/<id>
GET /api/blob/<sha256hex>

2 · Verify
Three checks: the robot's signature, the episode bytes, the requester's signed task.

python3 shared/episode.py verify \
  episode.json request.json \
  completion.json

3 · Build
Harvest into a dataset dir, then build TFRecords (splits keep autonomous and teleop apart):

harvest.py --base <this host> \
  --out dataset/
tfds_builder.py --dataset dataset/

Tooling: muretai/rlds-tools · schema & verification: muretai/muretai-network (docs/RLDS_MURETAI.md) · robot scaffold: muretai/robot-template.

Licensing

Participants consent to training use at recording time (machine-readable consent: training-ok/0.1 in every signed completion; requester identities are pseudonymous DIDs and can be salted for public distributions). For dataset licensing, larger corpora from template-based robots, or evaluation access: muretaicom@gmail.com.