SO-101 SmolVLA Multitask and Long-Horizon Evaluation v0.1.0

SO-101 SmolVLA Multitask and Long-Horizon Evaluation v0.1.0

Included

• Parameterized SO-101 follower/leader setup, calibration, teleoperation, and
  three-task recording scripts.
• Dataset schema validation and safe merge workflow.
• Verified SmolVLA v1 training configuration and 030000 checkpoint metadata.
• Sync, RTC, episodic rollout, W&B labeling/upload, camera alignment, and
  hierarchical long-horizon task scripts.
• Both arm calibrations, camera reference/audit images, all dataset and rollout
  metadata, W&B terminal training summary, environment exports, documentation,
  upstream provenance, and 413-row project inventory.

Storage

• Git LFS objects in this release: none.
• Raw dataset Parquet/video, rollout data/video, six full weights, and optimizer
  states: MANIFEST_ONLY with absolute original paths, sizes, and SHA-256.
• Hugging Face caches, raw W&B cache, upstream LeRobot checkout, secrets, and
  unrelated LIBERO assets are excluded.

Known limitations

Success rates are not yet formally measured. RTC can execute stale actions
after a failed grasp, camera/lighting changes produce distribution shift, and
the red-cube stage in watermelon-then-cube evaluation sees a box state absent
from typical red-cube demonstrations. The task queue is lexical and supports a
bounded object vocabulary.

Reproduction

Use LeRobot commit b895ed0fe4d78016783e708405c5879875908f42,
restore manifest-only assets at documented paths, run checksum verification,
activate the exported lerobot environment, and follow the safety and
evaluation protocols before physical motion.