IROS 2026: Collaboration with RobotiqRead recap

Annotation & Evaluation infra for robots.

Annotate robot runs and demonstrations. Evaluate task performance and investigate failures with video and sensor evidence.

Build with RoboLens ↗
RoboLens / shared evidenceAnnotation + evaluation
From pretraining to deployment

Recordings

Teleoperation recording

Egocentric, teleoperation and UMI data.

Video, robot state and tactile signals
RoboLensShared evidence

Annotation, evaluation
or both.

Prepare data for training

Annotation

Reviewed labels

Task steps, outcomes and interventions.

Trained VLMs, QC and human review
Understand the behavior

Evaluation

Failure evidence

Task outcomes, comparisons and failure evidence.

“Which steps failed?”

From a question to the relevant evidence

Recording source

Annotation

Training-ready data.
At scale.

Egocentric, teleoperation and UMI recordings, annotated for pretraining and deployment.

Trained VLMs
Our in-house models are optimized for annotation accuracy and cost.
Review included
Quality control and human reviewers check labels, timing and outcomes against your criteria.
Training-ready exports
Labels, timestamps and source references in structured formats.
See annotation examples
Robot data annotationRecording → labels
One run. A useful record.Actions, outcomes and assistance on one timeline
Task steps
Outcomes
Human assistance
Keep uncertainty visible

Ambiguous boundaries, missing context and uncertain outcomes are flagged for review.

Reviewed labels + source referencesFor policy training and task evaluation
Illustrative alignment

Evaluation

Catch regressions.
Debug faster.

See whether a change improves robot performance. Find the failing steps and curate the data for your next iteration.

See evaluation workflow

Illustrative scenario. Claude Code prompt: Did the perception update regress? Compare 120 test runs and suggest data to curate. RoboLens MCP compares the test batch, failure annotations, robot states and motion. Response: Misses rose 2/60 → 8/60, mostly under partial occlusion. The tote stays visible as its track drops, then the robot pauses. Curate occluded frames + matched successes for box and visibility review. Example counts and signals are simulated.

Evidence reviewIllustrative
Candidate · partial occlusion10.0 s
Failure annotationMissed tote detection8.0–10.0 s
Tote track
TrackedLost
Robot state
ApproachHold
TCP speed0–0.2 m/s
6 s8 s10 s
Suggested curation

Occluded tote frames + matched successes. Review boxes and visibility labels before retraining.

Illustrative workflow · example counts and signalsSwipe to see the evidence →
Catch regressions
Compare policy or software versions on matched tasks, using recorded runs or repeatable simulation evaluations.
Debug faster
Ask Codex or Claude Code through RoboLens MCP to locate failures and return the video, labels and signals.
Train on the right examples
Curate recurring failures and matched successes. Retest the same cases after training.

Work with RoboLens

Start with one task.

Start with a scoped annotation or evaluation pilot. We agree on your data, success criteria and deliverables, then configure RoboLens with your team.

Build with RoboLens ↗