Custom Robot Manipulation Datasets
Built to Your Task Spec

What "Custom" Means in Practice

A custom robot manipulation dataset is demonstration data collected against a written specification: your task, your objects, your scene constraints, your action space, your success criteria. Nothing in it existed before you asked for it. That distinction matters because the alternative, adapting whatever a public corpus happens to contain, works right up until your task leaves the corpus's coverage. Most commercially interesting tasks do.

The commercial model behind this is capture as a service. Labs that train manipulation policies increasingly buy aligned state-action demonstration data the way they buy compute: specified up front, delivered on a date, quality checked against agreed criteria. The buyer keeps researchers on modeling. The vendor runs operators, rigs, and QA as a production discipline.

What a Delivered Dataset Contains

A manipulation dataset is a set of episodes. Each episode is one attempt at the task, recorded end to end. A well-built episode carries:

• Observation streams. Scene cameras and wrist cameras at agreed resolutions and frame rates, plus proprioception: joint positions, end-effector pose, gripper aperture, and force or torque readings where the rig provides them.
• Action stream. The commanded actions in the action space you train in, whether joint targets, end-effector deltas, or absolute poses, recorded on the same clock as the observations.
• Alignment guarantees. One clock across every stream, verified per episode. Misaligned demonstrations are worse than fewer demonstrations, because the policy learns a systematic lag.
• Labels. Success and failure against written criteria, plus any stage or subtask segmentation the spec calls for.
• Calibration and metadata. Camera intrinsics and extrinsics, rig configuration, and a data card describing collection conditions, operator protocol, and known limitations.

Delivery format is part of the spec, not an afterthought. We deliver LeRobot, RLDS, or GR00T-compatible exports with a loader script; the trade-offs between them are covered in our format guide for buyers.

Four Signals You Need Custom Data

Not every team should buy a dataset. These are the situations where custom collection earns its cost:

• Your task is missing from open corpora. Search DROID and the Open X-Embodiment mix for your task first. If the closest match is a cousin of what you need, fine-tuning on cousins gets you a policy that does the cousin task. Our custom versus open comparison treats this honestly, including the cases where open data is enough.
• Your embodiment or action space is unusual. Cross-embodiment transfer keeps improving, and it still pays a tax. Data collected in your action space with your camera placement removes that tax.
• Your objects deform, reflect, or occlude. Cloth, cables, and packaging are thinly covered in public data. This is our specialty; see the deformable manipulation page.
• Your demonstration collection is eating research time. If PhD students are running teleop shifts, the lab is paying senior rates for operator work and losing the research hours on top.

How Big Should It Be?

The honest answer is that episode counts are estimated, then measured. Task horizon, policy class, and whether you co-train on open corpora all move the number, and the only trustworthy signal is a pilot batch trained into your actual pipeline. That is why our engagements front-load the pilot: fifty-odd real episodes in your format inside two weeks, so the full-run size is set by your validation curves instead of by a sales conversation. Buyers co-training on an open prior routinely need fewer custom episodes than a from-scratch estimate suggests, which is a budget argument for doing the open-data homework first.

Diversity matters more than raw count once past the minimum. A thousand episodes of the same object on the same table teach less than four hundred episodes across specced variations in objects, initial states, and lighting. The diversity axes belong in the spec, with sampling rules, so coverage is auditable at delivery rather than assumed.

How a Collection Runs

Every engagement follows the same shape. We turn your task into a collection spec and you sign it. A pilot batch of real episodes lands in your pipeline within two weeks, and you check cameras, alignment, labels, and format while changes are still cheap. Full collection then runs in our purpose-built capture studios in Asia Pacific with trained operators, and every episode passes synchronization, calibration, and completeness checks or is re-collected. Sixty days after the signed spec, the dataset is yours, exclusively and permanently.

If you have never written a dataset spec, start with our spec checklist. A precise spec is most of the difference between a dataset that trains and a dataset that disappoints.