so101-tic-tac-toe-norm — the same 210 episodes, in LeRobot NORMALISED units

A units-converted copy of cagataydev/so101-tic-tac-toe (210 teleop episodes, 57,092 frames @ 30 Hz, 20 task strings, cameras observation.images.top + observation.images.wrist). Same episode indices, same tasks, same videos (byte-identical). Only action and observation.state differ:

	source	this copy
arm joints	degrees (lerobot use_degrees=True)	[-100, 100] (lerobot RANGE_M100_100)
gripper	[0, 100]	[0, 100] (unchanged)
wire semantics on agi.diy	so101-degrees-gripper-percent-v1	so101-normalized-v1 (what agi-wfm/so101, agi-act/so101, agi-smolvla/so101 speak)

Conversion uses the recording follower's own calibration (blue_follower.json) and lerobot's formulas from lerobot/motors/motors_bus.py (_normalize, drive_mode 0): norm = deg · (4095/360) · 200/(range_max − range_min) per joint — a pure scale (the mid-range offsets cancel), so per-episode stats and meta/stats.json scale with the same factors. Scales: shoulder_pan 1.0927 · shoulder_lift 0.9685 · elbow_flex 1.0336 · wrist_flex 0.9913 · wrist_roll 0.5556 · gripper 1.

Proof (agi_wfm/data/so101_units.py in cagataycali/agi-wfm, branch ttt/wfm): round trip max abs error 7.6e-6° over all 57,092 frames; agreement with lerobot's own _normalize 7.6e-6 over 2,000 random tick vectors across the calibrated range; 0 samples outside [-100, 100] (a live arm driven in normalised mode never saturates on these actions); converted range min [-44.9, -99.4, -80.8, 6.2, -48.4, 0.1] max [93.2, 64.8, 99.9, 99.9, 66.0, 44.9].

Live arm: read and drive the follower with lerobot use_degrees=False (the default) and the same calibration id so the policy's outputs are in the units it was trained on. Cell names are in the ROBOT's frame — read the source card.

Trained on this copy: cagataydev/agi-wfm-so101-ttt (the flagship of api.agi.diy — 190 train episodes of this set, 1 : 1 with so101-tic-tac-toe-sim); it answers in these units. The source dataset and the model form the agi collection.