maximilianofir/so101-left-rack-balanced-1000-dr-speed05-1-20261005

Balanced left-rack SO-101 demonstrations

1000 accepted demonstrations; 250 each with 0, 1, 2, or 3 preloaded rack vials. Four total vials, rack on the left, both cameras at 30 fps. Source commit: f6c017c6b7b2882d39b0aaee0f1a828ac3f9580e. The shoulder/elbow joint-target speed cap is sampled uniformly from 0.5–1.0 times its baseline (1–2 rad/s) at every reset. This is a target-rate limit, not a measured physical-speed guarantee. Path timing and wrist/gripper response retain their defaults.

Uses the branch's complete DR task: randomized object placement, mat rotation, external-camera tilt/focal length, HDRI/yaw, sky exposure/temperature, and light-box exposure. Robot colour remains the configured orange; the wrist camera retains its calibration. A single environment ensures shared HDRI and focal length are resampled on every reset.

Only successful episodes are included; this introduces success-selection bias in sampled conditions. This is a training dataset, not an independently held-out validation set. See collection_manifest.json for batch seeds, configuration and zero-based episode ranges.