cons909/V1-LeRobot-SO101-Eval-Videos
SO-101 SmolVLA Evaluation Videos — V1 Full evaluation footage and per-trial logs for V1-LeRobot-SO101-Learning-by-Demonstration-SmolVLA — a screwdriver-to-bin pick-and-place task taught to a low-cost SO-101 robotic arm via imitation learning, using LeRobot and the SmolVLA policy. This dataset contains all 150 evaluation trial videos (50 trials x 3 camera configurations) plus their trial-by-trial logs. It's the full record behind the results reported in the GitHub repo's README —… See the full description on the dataset page: https://huggingface.co/datasets/cons909/V1-LeRobot-SO101-Eval-Videos.
SO-101 SmolVLA Evaluation Videos — V1
Full evaluation footage and per-trial logs for **V1-LeRobot-SO101-Learning-by-Demonstration-SmolVLA** — a screwdriver-to-bin pick-and-place task taught to a low-cost SO-101 robotic arm via imitation learning, using LeRobot and the SmolVLA policy.
This dataset contains all 150 evaluation trial videos (50 trials x 3 camera configurations) plus their trial-by-trial logs. It's the full record behind the results reported in the GitHub repo's README — the repo itself only ships a handful of sample clips to stay lightweight; everything else lives here.
The chart above ("video" / 47.2 — 181) is Hugging Face's own auto-generated preview of clip durations across all 150 videos, in seconds. The spike near 181s is trials that ran out the 180s timeout; everything to the left ended earlier because the arm finished or the operator judged the attempt over. It's a duration summary, not a quality or success indicator.
Watch a video right now
Click any link below to stream it directly in your browser (no download required):
- 1 camera (left side) — a successful trial · a failed trial
- 2 cameras (side left + side right) — a successful trial · a failed trial
- Side + claw — a successful trial · a failed trial
To browse all 150 videos: click the Files tab at the top of this page → open 1cam/, 2cam/, or side_claw/ → videos/ → click any trial_XXX_side.mp4 filename. Each file page has a built-in player (press play) and a Download button. Or use the full per-trial tables below — every row links straight to its video.
Configurations evaluated
Full methodology (pre-registered protocol, grading rubric, controls) is documented in the GitHub repo's `docs/EVALUATION_PROTOCOL.md`.
Full per-trial results
Every trial, every outcome, every video — click "watch" to stream it directly. This is the same data as the trial_log.csv files, just browsable. For scripts/analysis, use the CSVs directly (1cam/trial_log.csv, 2cam/trial_log.csv, side_claw/trial_log.csv) — same columns, no need to scrape this table.
<details> <summary><b>1 camera (left side) — 50 trials</b></summary>
</details>
<details> <summary><b>2 cameras (side left + side right) — 50 trials</b></summary>
</details>
<details> <summary><b>Side + claw — 50 trials</b></summary>
</details>
outcome is one of 8 pre-registered categories — see `docs/EVALUATION_PROTOCOL.md` for exact definitions. trial_log.csv also logs avg_loop_latency_ms, stop_reason, notes, and checkpoint_step, not shown in the table above.
Privacy note: for both multi-camera configs (2cam, side_claw), only one camera's footage is included per trial — the other camera's feed was dropped before publication to avoid a family member appearing in frame. This does not affect any logged result, only which raw video ships.
Source code
The evaluation harness that produced this data, the Raspberry Pi robot-control servers, and the full methodology/build-log writeup all live in the companion GitHub repo: [cons909/V1-LeRobot-SO101-Learning-by-Demonstration-SmolVLA](https://github.com/cons909/V1-LeRobot-SO101-Learning-by-Demonstration-SmolVLA).
License
MIT, matching the GitHub repo.