CrimsonRobot/crimson-ma2-evaluation

Crimson MolmoAct2: preliminary real-robot evaluation

Right-arm bottle pick-and-place rollouts recorded on 2026-09-24 (Asia/Bangkok). This release contains three checkpoints, 31 logged attempts, videos from three cameras, selected recorded telemetry, and joint trajectories. It is an evaluation evidence release, not a training dataset or a completed six-checkpoint benchmark.

Model source: Kavin60606/crimson-ma2-ckpts. The policy task text was pick and place the bottle. Each checkpoint was assigned five bottle positions and two repetitions per position. The labels are operator annotations, not independently certified task outcomes. place=1 is the source's success label.

Results as labeled — all logged attempts retained

Checkpoint	Attempts	Success labels	SR	Wilson 95% interval	Reach	Grasp	Carry	Place
s2000/001000	10	1	10.0%	1.8–40.4%	4	1	1	1
s1000/001500	11	1	9.1%	1.6–37.7%	7	1	1	1
s1000/000500	10	0	0.0%	0–27.8%	1	0	0	0

Stage counts use the same denominator as the attempts column, not a conditional denominator. The dominant stage failure inferred from these labels is: not reached (6/10) for s2000/001000; reached but not grasped (6/11 all attempts) for s1000/001500; and not reached (9/10) for s1000/000500. These stage failures are not causal diagnoses of model versus system faults. For s1000/001500, selecting the latest attempt per position/repetition gives 1/10 success labels and 6/10 reach; this is sensitivity analysis only, not an approved exclusion rule. summary.csv provides both views. rollouts.csv retains all 31 rows: do not feed it into a script assuming exactly ten rows per checkpoint without resolving the duplicate policy.

Important qualifications

• Results are published as recorded, with no fixed-time-limit reclassification, as requested by the recording owner. Actual engagement durations varied; some episodes contain release/re-engage cycles. See engagement_timing.csv and each episode.json for intervals. Recording length includes setup and is NOT policy execution time. This is not a standardized fixed-duration benchmark. Success labels have not been independently certified; video includes post-disengagement activity, which must not automatically be attributed to the policy.
• In sampled review of the success-labelled s1000/001500 attempt, the bottle was visibly separated on the table after policy ownership ended. This limits attribution of the release to the policy; the operator's success label is retained unchanged.
• s1000/001500 position 4, repetition 1 has two attempts. The earlier attempt stopped after approximately 14.6 seconds because of consecutive discontinuous chunks. The source notes mark a later attempt as superseding it. The later attempt also stopped for consecutive discontinuous chunks (approximately 39.4 seconds), despite a less specific source label. All attempts remain in this release. A model-caused safety stop is not automatically an invalid trial eligible for replacement.
• Some stop reasons were inconsistently described as “system” versus “model” in source notes. Those notes remain annotations, not proven causal diagnoses; recorded status reasons are provided separately. Two failures in the first checkpoint have blank failure descriptions.
• The release does not assign an official six-checkpoint Spearman correlation, top-1/top-2 agreement, or statistically significant winner. Only three checkpoints were tested and the protocols/duplicate rules require qualification. The proposed “difference below two successes means tie” rule is a descriptive benchmark convention, not a statistical significance test.
• The recorded pipeline applies command limits and stop gates. These are closed-loop policy-plus-executor results, not unrestricted model-only performance measurements.

Files and how to inspect a rollout

1. Open attempts.csv. attempt_id is unique; (ckpt,pos,rep) is not necessarily unique. VIDEOS.md provides a direct camera link for every attempt.
2. Follow episode_path to the corresponding folder under episodes/.
3. Play cam_high.mp4, cam_wrist.mp4, or cam_wrist2.mp4.
4. Use the matching *.frames.jsonl.gz for original per-frame receipt/capture timestamps.
5. Read trajectory_state.csv.gz for recorded measured/commanded joint positions; trajectory_approved.csv.gz for supervisor-approved positions and velocities.
6. streams/*.jsonl.gz supplies selected original telemetry fields without smoothing or resampling. Consult DATA_DICTIONARY.md.

The MP4s are lossy H.264 viewing derivatives, 15 frames/s with one original JPEG per encoded frame. Their player time is a nominal frame timeline; original timestamp indices are authoritative for measurements and cross-camera synchronization. They contain no audio. They include setup and post-policy activity: not all visible movement was policy-controlled. Original JPEG image files are not included.

rollouts.csv preserves source outcome labels and notes. attempts.csv adds stable IDs, duplicate flags, checkpoint/config digests and recording links. outcome_basis explicitly identifies the source operator labels without time-limit reclassification. release_audit.json lists exported stream counts, omitted-field counts and video checks. manifest.json provides SHA-256 and byte size for published files (excluding itself). The standalone code/package_release.py and its tests document the conversion and statistics; they require local raw recordings and cannot reconstruct original JPEGs from the MP4s.

Scope, privacy and limitations

Only the 31 CSV-referenced recordings from these three evaluation sets are included. The unscored episode numbered 000007 in the s1000/001500 set is excluded. Earlier exploratory trial runs and training episodes are not included. No model weights are redistributed.

This is a sanitized telemetry export, not a byte-identical raw-log archive. Explicit stream field allowlists omit network/link/session/host information and arbitrary nested config payloads. Local filesystem paths and private-address strings are redacted. Camera indices, scientific numeric arrays, outcome annotations and selected status reasons are preserved. The schema and omission counts make the export scope visible. Sampled visual privacy checks cover every camera and attempt; sampling cannot guarantee absence of incidental content in every frame. Public release was authorized by the recording owner.

Telemetry records the executed/approved trajectory, not necessarily every raw predicted model action chunk. Complete 10-step model predictions are not supplied in this release. No external tool-position metrology is provided; joint data are encoder/model based.

No additional reuse license is granted in this release. Public accessibility alone is not a license to reuse recordings; contact the repository owner for reuse permission. Do not replay these trajectories on hardware without an independent setup and safety assessment.