Navigation data across every lighting and crowd condition.
AMRs succeed or fail on perception under conditions nobody planned for: glare at 4 p.m., a crowd leaving a shift, a pallet left in a corridor. Agentuor makes those conditions visible in the data and in the evaluation.
The challenge
Mobile robot perception datasets tend to be biased toward the conditions in which they were convenient to record. Dynamic agents — people, carts, other robots — are under-labeled because they are tedious to track. Semantic maps go stale. When a robot fails in production, the team cannot say whether the condition was in the training data, let alone how the model performed on it.
Built for this data
LiDAR obstacle labeling
Point-level classes and cuboids for static and dynamic obstacles across sweeps, fused with camera views.
Dynamic agent tracks
People, vehicles, and robots tracked over time with trajectory consistency checks.
Semantic map annotation
Floor, wall, doorway, and zone labels that update as facilities change.
Condition coverage
Agents chart the corpus by lighting, crowd density, and clutter and recommend what to record next.
Condition-sliced evaluation
Detection and navigation metrics by lighting, density, and environment type.
Failure tracing
Jump from a near-collision in evaluation to the frames and labels that should have covered it.
Data and labels
- Multi-LiDAR sweeps with wheel odometry and IMU
- Camera video with synchronized poses
- Static obstacle and dynamic agent labels with tracks
- Semantic map layers and zone annotations
- Lighting, crowd, and environment metadata
Typical workflow
Ingest routes
Recorded routes are segmented and tagged by facility and conditions.
Label with assistance
Pre-labels for common classes; agents flag track breaks and identity swaps.
Review dynamics
Ambiguous agent interactions routed to expert review.
Evaluate by condition
Rollouts gated on performance under the hardest slices.
Outcomes teams work toward
Common questions
Can we import existing HD maps?
Yes, as reference layers for annotation and evaluation.
How are identity swaps in tracks handled?
Agents detect implausible trajectory jumps and propose corrections for confirmation.
Is outdoor data supported?
Yes — including weather and time-of-day metadata for slicing.
Show us an episode. We'll show you the gaps.
Start with a walkthrough on your own data.