Teleoperation Data Collection Service: Methodology, QA, Formats and Pricing

We collect teleoperated demonstration data for robot learning: a trained operator drives the robot through your task, and every episode is recorded, checked and delivered in the format your training pipeline reads. Projects run from 20-episode pilots to 10,000+ episode production campaigns.

This page explains how a project works, what we record, how we check it, what you receive and how it is priced. For the service overview and the quote form, see Robot Data Collection Services.

Last updated October 18, 2026.

How a campaign works

A campaign is run as an engineering project, in six phases:

1. Kickoff and task design. We define the task, success criteria, scene variation, robot embodiment, sensor stack and target policy. The output is a collection brief with measurable acceptance criteria.
2. Hardware configuration and timing verification. We set up the robot, operator interface, cameras, lighting and safety limits. Every sensor joins a shared clock before production starts.
3. Operator training and qualification. Operators rehearse the exact task and pass a proficiency test on success rate, motion quality, consistency and reset discipline.
4. Collection sprints. Qualified operators collect in controlled sprints. Supervisors watch throughput, failure modes, scene coverage and per-operator quality.
5. Quality control and validation. Episodes are checked for task completion, timestamp integrity, frame loss, schema consistency, motion anomalies and metadata completeness.
6. Delivery and handoff. Approved episodes ship with manifests, calibration files, quality reports and an engineering handoff.

Operator interfaces

We pick the interface to fit the task's precision, throughput, workspace and robot, rather than using one tool for everything.

Method	Best for	Precision	Typical demos per day	Setup effort
Leader-follower arms	Precise bimanual and contact-rich tasks (ALOHA-style)	Highest	30–80	Medium
VR (Meta Quest 3)	Large workspaces and mobile manipulation	High	40–100	Medium
SpaceMouse / keyboard	Simple Cartesian tasks and pilots	Medium	60–150	Low
Haptic gloves	Dexterous hands and finger control	High	20–60	High
Kinesthetic teaching	Backdrivable arms and precise paths	Highest	20–50	Low
Scripted demos	Repeatable baselines and calibration	Deterministic	100+	Medium

The hardware behind these: leader-follower setups on ALOHA-style WidowX/ViperX arms and OpenArm, Meta Quest 3 headsets, 6-DOF SpaceMouse devices and SenseGlove Nova 2 haptic gloves. For bimanual tasks we run dual leader-follower or dual VR configurations. Glove-driven collection for dexterous hands, including egocentric and UMI data, is covered in more depth on the dexterous hand data page.

Supported robots

We collect on OpenArm 1, DK1 bimanual, Franka FR3, UR3e and UR5e, Unitree G1, xArm 6/7, Kinova Gen3 and most ROS 2-compatible arms. You can ship us your robot, or we can procure one. Integrating a new robot typically takes 3–5 business days and is scoped before production.

No robot yet? The same hardware is available to buy or lease.

What we record

An episode can include any combination of these streams, aligned to one task timeline:

Signal	Contents	Typical rate	Delivered as
Joint states	Position, velocity, effort	20–100 Hz	Float32 arrays with timestamps
RGB cameras	Head, wrist, side, overhead	30–60 FPS	MP4/JPEG or embedded arrays
Depth cameras	Aligned depth with intrinsics	15–30 FPS	16-bit depth or point cloud
End-effector pose	XYZ plus quaternion or 6D pose	20–100 Hz	Base- and task-frame transforms
Gripper state	Position, force, open/close	20–100 Hz	Continuous or discrete
Force/torque	6-axis wrench and contact events	100–1,000 Hz	Calibrated SI units
Annotations	Success, phases, language, keyframes	Per episode or frame	JSON, CSV or embedded

All streams carry monotonic timestamps and documented clock provenance. Exact rates and tolerances are fixed in the collection brief.

Quality assurance

Timing

Cameras, robot state and actions share a verified clock, using hardware-triggered cameras and shared clock sources. Each delivery includes a timing report against these targets:

Metric	Target
Camera-action sync	< 5 ms
Cross-camera skew	< 1 ms
Frame drop rate	< 0.1%
Jitter (σ)	< 2 ms
Clock drift	< 0.5 ms/hr

The metrics are calculated from the delivered episodes, not from device specifications, and any exceptions are listed in the handoff report. The Robot Time Fabric article explains why timing matters for training.

Operators

• Qualification testing: operators pass task-specific trials before their episodes count toward delivery.
• Real-time monitoring: supervisors track success rate, cycle time, frame loss and scene coverage during sprints.
• Per-operator metrics: episode quality is attributed by operator, so drift and retraining needs show up early.

The 10-point episode checklist

1. Synchronized timestamps: cameras, joints and actions aligned within 5 ms using shared clock sources.
2. Consistent episode structure: identical observation/action schema, dimensions, data types and key names.
3. Operator qualification: operators pass a proficiency test on the specific production task.
4. Task success verification: every episode is reviewed; failures are flagged and separated on request.
5. Scene reset consistency: object positions, lighting and workspace state follow documented reset ranges.
6. Frame drop monitoring: episodes above the allowed loss rate are recollected.
7. Gripper state consistency: gripper signals are validated against visible contact and commanded actions.
8. Joint limit compliance: trajectories are checked for limit violations, clipping and unsafe motion.
9. Metadata completeness: robot, operator, task, calibration, scene and timing metadata ship with every episode.
10. Annotation standards: labels follow a documented ontology, with spot checks for consistency.

Delivery formats

Datasets are delivered in the format you specify in the brief. We handle conversion and validate against the schema before handoff.

Format	Works with	What you get
HDF5	ACT, ALOHA, robomimic	Per-episode files with synchronized camera, state and action streams
RLDS / TFRecord	Open X-Embodiment, Octo	TensorFlow-native trajectories with standardized steps and metadata
LeRobot (Parquet)	Hugging Face LeRobot	Hub-ready datasets with Parquet state/action tables and encoded video
Custom	Your training stack	Your schema, field names, chunking, compression and validation rules

For the trade-offs between them, see HDF5 vs RLDS vs LeRobot.

Available annotations include timestamped task phases, subtask boundaries, language instructions for VLA training, keyframes and success/failure labels. Custom annotation schemas are available for enterprise campaigns.

How many demonstrations you need

The answer depends more on task diversity and the policy you are training than on a universal number. As a starting point:

Task class	Example	Typical demos	Goal
Simple single-arm	Pick-and-place, one object	20–50	Pilot, behavior cloning
Moderate single-arm	Variable objects and poses	100–300	Robust task policy
Bimanual	Coordinated assembly or handoff	300–1,000	ACT, diffusion policy
High diversity	Many objects, scenes, operators	1,000–5,000	Generalization
VLA / generalist	Language-conditioned multi-task	5,000+	Foundation-model adaptation

Pricing

We recommend starting with a pilot to validate the task and the instrumentation, and scaling only after it passes.

Package	Scope	Price	Includes
Pilot	20 demonstrations	$2,500	Task design and collection brief, hardware setup and timing validation, expert operator collection, QA report and one delivery format, 1–2 week turnaround, 2 weeks of post-delivery support
Campaign	100 demonstrations	$8,000	Everything in Pilot, plus multi-station parallel collection, 2–4 dedicated operators, weekly batch deliveries, scene diversity management and up to 2 delivery formats
Enterprise	Custom scale or ongoing	Custom	Dedicated collection infrastructure, on-site or co-located deployment, SLA, custom robot integration, all formats and data platform access

What moves the price is task complexity, the number of camera views, the teleoperation method and the QA requirements. Simple tabletop tasks cost less per episode; contact-rich bimanual tasks with dexterous hands cost more. Volume discounts apply to campaigns over 500 episodes.

A standard 100-demonstration campaign takes 2–6 weeks, depending on task complexity and how much scene diversity you need. Enterprise projects are scoped individually.

Data ownership and confidentiality

We sign mutual NDAs before discussing proprietary tasks, robot configurations or research goals. Data collected under contract is owned by the client. We do not use client data for other purposes or include it in public datasets.

Frequently asked questions

What does a teleoperation data collection pilot cost?

A 20-demonstration pilot is $2,500, including task design, hardware setup and timing validation, collection, a QA report and one delivery format. A 100-demonstration campaign is $8,000.

Which formats do you deliver?

HDF5 (ACT/ALOHA), RLDS/TFRecord (Open X-Embodiment, Octo), LeRobot Parquet (Hugging Face), or a custom schema.

Can you collect data on our robot?

Yes, for the robots listed above and most ROS 2-compatible arms. New integrations typically take 3–5 business days.

How do you make sure the data is usable?

Every delivery includes a timing report measured on the delivered episodes, and episodes go through the 10-point checklist above. Episodes that fail task-success or frame-loss checks are flagged or recollected.

Who owns the data?

You do. Data collected under contract belongs to the client and is not reused or published.

Start a project

Tell us the robot, the task, the sensor stack and the policy you plan to train, and we will turn it into a pilot scope.

Request a quote · Robot data collection guide · Data collection cost breakdown · Sample datasets