mobile manipulation

Long-horizon episodes, labeled consistently.

Mobile manipulation produces the messiest data in robotics: minutes-long episodes, many sub-tasks, partial successes, and objects that move, deform, or disappear. Agentuor keeps every boundary and outcome consistent across thousands of them.

The challenge

A single fetch-and-place episode can contain navigation, approach, grasp attempts, regrasps, transport, and placement — each with its own success criterion. When annotators segment these by hand, boundaries drift between people and across weeks, and outcome labels get applied to the wrong sub-task. Models trained on that data learn the inconsistency. Teams discover the problem only when a policy behaves erratically on a task the dataset supposedly covered.

how agentuor helps

Built for this data

Event segmentation assistance

Agents detect merged sub-tasks and propose boundaries at the frame where the robot's state actually changes.

Outcome enforcement

Every grasp attempt and every episode must carry a success, failure, or partial label before the batch can close.

Object tracks across views

Wrist, head, and external camera tracks reconciled into a single object identity with depth-consistent boxes.

Guideline drift detection

When two annotators place the same boundary differently, the disagreement is clustered and resolved once.

Per-sub-task evaluation

Score policies on grasp, transport, and placement separately so a failing stage is visible.

Demonstration filtering

Rank teleop demonstrations by cleanliness and route noisy ones out of imitation-learning sets.

Data and labels

  • Synchronized RGB-D from wrist and head cameras with joint positions, gripper state, and force-torque
  • Sub-task boundaries and per-attempt outcomes on a shared timeline
  • 2D boxes, instance masks, and 3D cuboids for target and distractor objects
  • Scene metadata: surface type, clutter level, lighting, object novelty
  • Operator identity and intervention markers for teleoperated sessions

Typical workflow

1

Ingest sessions

Episodes are segmented from continuous recordings and tagged with task and scene metadata.

2

Pre-segment

Agents propose sub-task boundaries and grasp attempts for annotators to confirm.

3

Annotate and review

Object tracks and outcomes are labeled; ambiguous boundaries go to expert review.

4

Evaluate by stage

Policies are scored per sub-task and per condition, with failures linked to source episodes.

what changes

Outcomes teams work toward

Fewermerged events reaching the training set
Per-stagevisibility into where a policy fails
Oneboundary convention across every annotator

Common questions

How long can episodes be?

There is no practical limit; the timeline tooling is built for multi-minute recordings with thousands of frames.

Can we define our own sub-task ontology?

Yes. Sub-tasks, outcomes, and attempt types are all configurable and versioned.

Do you support deformable objects?

Yes — mask and keypoint labels handle cloth, cables, and food items; agents flag identity ambiguity when shape changes.

Show us an episode. We'll show you the gaps.

Start with a walkthrough on your own data.