josefchen/RoboJev-evals

RoboJev evaluation records

In the data we have 29 model and protocol records. We have 927 human-labelled episodes consisting of 268 ArmnetBench SO-101 episodes, 459 Robometer unseen-lab episodes and 200 community clips. There are 65 community clips that are ambiguously tagged as scored: false, which means there are 862 episodes for the purpose of these numbers.

Interactive explorer · Writeup and charts · Model · Live demo

What is included

• Each line in benchmark.jsonl contains the task, human label, dataset, episode, evaluation group and recorded P(success) per model for each episode. Entries where a prediction is not available will be set to null, and model and protocol names are kept as keys in the JSON.
• The benchmark.csv file contains the same data as a single, flat table.
• The summary.json contains aggregated statistics along with the input protocols, the release revision and paired bootstrap comparisons.
• The file catalogue.jsonl contains 68,555 sampled episodes, approximately 501 hours of video, from 7,641 public LeRobot datasets. Note that this is a sample, which means not all the episodes available on Hugging Face are listed here.

There are currently no success labels in the catalogue from the new model. Predictions of the old four-frame release are left out, and the new model has not yet regraded the entire catalogue.

Methods

For the released RoboJev and ten-frame JEV baseline, we use 10 main camera frames and final frames from other cameras if they exist (dense-v1). For clip-v1, we only use 4 front-camera frames. video means whole video input. The frontier LLMs answer zero-shot; RoboJev was trained using data from other ArmnetBench sessions. Because these inputs don't all contain the same number of frames, we shouldn't interpret the difference in protocols as a purely training effect.

The SO-101 balanced accuracy is based on a threshold of 0.5. Ranking on the unseen labs is based on AUROC. The community accuracy is based on the 135 clear outcomes that were labelled by one person. Non-reported results are not assumed to be zeroes. The earlier variants will be preserved for inspection along with their protocol.

This is the released model e089a9b69a365744b80765057d86583647af95a9 (robojev-27b-da). The model card provides information about training, calibration, how the splits were built and limitations. Cost and latency are historical measurements of recorded runs, and are not current quotes from the vendor, and the cost extrapolations don't include download, decode and idle time.

Sources and rights

Episodes are from ArmnetBench, Robometer's unseen labs, and the public LeRobot datasets identified in each community record. Generated evaluation records and the explorer code are Apache-2.0. Original datasets and videos have separate licenses and this dataset repo does not redistribute any video.

Community labels can be confusing, benchmark labels can be wrong, and a model can confidently answer incorrectly. The list of controversial unseen-lab episodes is in the benchmark/ folder of the model's repository.