ohong/intel-robotics-dataset

Intel Robotics — Second Look dataset

Recorded data, detector annotations, and QC evidence from Second Look, a Physical AI challenge project pairing an SO-101 robotic arm with Anomalib defect detection and a VLA policy. Source code and docs live in the paired GitHub repository (not included here — see below).

This is a working project dataset, not a finished benchmark. Several pieces are explicitly flagged below as unverified; read those notes before training on this data.

Contents

Path	What it is
artifacts/recording/	LeRobot-format episode recordings, capture manifests, QC reports, ACT loader smoke tests
artifacts/cv/	Anomalib detector captures, annotations, and coverage reports
artifacts/discovery/	Environment/hardware discovery evidence
artifacts/verification/	Intel deployment verification checks, benchmarks
artifacts/models/h100-patchcore-v3/	Frozen LEGO PatchCore detector (H100 fit; Torch + OpenVINO IR) used by the Intel pilot
artifacts/cv/local-cpu-patchcore-*/	Earlier local CPU PatchCore fits and their evaluation images (superseded by the H100 model)
artifacts/placement/	Receipts for the 27-episode placement dataset: audit, split, encoder manifest, sampled visual QC, H100 transfer receipt
pc-context/local/observations/	Local observation logs from the Intel PC

The five-episode recorded pilot

artifacts/recording/pilot-five-20260916T004831Z/ — the main recorded dataset. Real teleop demonstrations, SO-101 follower arm, two cameras.

• 5 episodes, 3,249 frames, 30 fps, two task captions ("pick up a good piece" / "pick up a defective piece").
• State/action: 6-dim, joint order shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos, normalized units.
• Cameras: low (1920x1080) and high (1280x720), AV1-encoded, gap-free timestamps across all 3,249 frames.
• Structural, decode, and native ACT-loader checks pass. Not verified: defect/bin correctness, inter-camera/camera-to-joint synchronization, specimen or session independence. See qc-readonly/ reports alongside the snapshot for details.
• Detector annotations over all 3,249 frames are in artifacts/cv/: 2,614 rows are invalid/no-inference, and the 635 scored rows are all flagged anomalous with target_association=UNKNOWN. Detector-conditioned sorting training is blocked on this data — the annotations don't yet establish which detection corresponds to which manipulated object.

Placement dataset (receipts only)

artifacts/placement/ holds receipts for a later 27-episode inspection-to-bin recording (15 BLUE / GOOD, 12 PINK / BAD, 16,981 frames, frozen 17/5/5 split, manifest SHA-256 943b9964…). The episode videos themselves are still only on the Intel PC and are not in this dataset yet. No policy has been trained on them.

License

Not yet assigned — ask the maintainer before reuse.

Related

• Source code, docs, and handoff state: github.com/ohong/intel-robotics (start with README.md and AGENT_STATE.md).
• Project status: paused September 23, 2026. Inspection pilot only; no autonomous physical trials.