THE WORKFLOW
Teleoperation, recording, policy and evaluation are separate milestones.
A robot can be assembled and teleoperated without having a high-quality dataset. A dataset can be recorded without having a trained model. A trained policy can run inference without being safe or successful on a physical task. Each transition deserves its own evidence.
REAL OPENRGD BASELINE
Our first confirmed milestone is the native SO-101 data-acquisition path.
Previous OpenRGD work sessions reported working SO-101 leader/follower teleoperation with persisted calibration identities, a single local 15-second no-video smoke recording (448 samples, episode 0) and replay. These are operator-confirmed results, not independently archived performance measurements.
No ACT/SmolVLA task success rate, real-world VLA inference latency or synchronized camera dataset is claimed from that smoke session. See the public lab ledger for the exact boundary between reports and pending artifacts.
FOUR LAYERS
The complete system is more than a policy checkpoint.
SO-101 embodiment
Joint order, motor mapping, calibration, mechanical limits and the actual follower state.
LeRobot acquisition
Demonstrations, camera observations, episode timing, task metadata and raw evidence.
ACT / SmolVLA
Learned action predictions and chunks conditioned on the dataset’s feature schema.
OpenRGD runtime boundary
Semantic intent, capability grounding, traceability and separately governed physical execution.
REPRODUCIBILITY
A public protocol is useful before there is a public benchmark.
The SO-101 lab publishes a machine-readable evidence ledger, experiment protocol and empty CSV trial template. The protocol names task definitions, failure categories, version requirements, timing fields and a conservative matched-trial floor. Publishing the method first prevents success criteria from being improvised after results arrive.