TELUS Digital Alternatives for Physical AI Data

Quick facts

Topic

Telus Digital

Audience

Procurement leads, ML ops, robotics engineers

Deliverable

Buyer-facing reference + procurement guidance

What TELUS Digital Is Built For

TELUS Digital is an annotation-first AI data services provider: multimodal labeling, multilingual collection, and post-training work for generative models^[1]. It reports 1M+ contributors across 500+ languages, with RLHF, red-teaming, and expert-in-the-loop review layered on top of automated labeling. The model assumes the data already exists. You bring the raw sensor streams (RGB-D video, LiDAR scans, IMU logs, joint encoders) and TELUS Digital adds the human judgment layer.

Physical AI inverts that dependency. A manipulation policy needs episodes that were never recorded: real-world pick-place, navigation, and teleoperation captured with specific hardware in the target environment. Scale AI's physical AI push and DROID's 76,000-trajectory dataset exist because embodied data has to be collected against a hardware spec; no volume of crowdwork retrofits it. The 500-language claim is beside the point when the bottleneck is finding collectors with UR5e arms and RealSense D435 cameras who can record failure modes (grasp slips, collision recovery, re-grasping) into RLDS episode structure with sub-millisecond timestamp alignment.

TELUS Digital vs Truelabel: Side-by-Side

The two products barely overlap. One applies human judgment to data you own; the other produces episodes that do not exist yet. Line them up against what a robotics buyer actually weighs and the split is clean.

Why Annotation Scale Does Not Solve Capture

TELUS Digital's 1M+ contributors are real leverage where the task is stable and quality gates are statistical: bounding boxes, semantic segmentation, RLHF preference labels^[2]. Headcount scales those. It does not scale physical capture, because robotics data quality tracks collector expertise far more than crowd size. Open X-Embodiment's 22 datasets came from 21 research labs, and DROID's 86 building locations required people who could deploy a robot in a real home rather than annotate images from a laptop.

The gap is what a crowdworker cannot produce. Multi-sensor fusion (RGB-D for visual grounding, LiDAR for 3D geometry, IMU for ego-motion, joint encoders for proprioception, force-torque for contact dynamics) has to be recorded together, at capture time. RT-1's 130,000-episode dataset used 13 robots over 17 months; both it and Open X-Embodiment's 527 skills across 160,266 tasks needed purpose-built rigs^[3]. Synchronization is the hard constraint: a 10ms skew between RGB frames and joint positions corrupts inverse-kinematics training, so truelabel enforces alignment via MCAP monotonic clocks and RLDS episode boundaries. TELUS Digital's QA checks label accuracy; capture QA checks calibration matrices, frame-rate consistency, and trajectory smoothness. The enrichment a labeling workforce cannot add is the differentiator: teleoperation metadata (demonstrator ID, intervention timestamps, re-grasp counts), failure-mode tags (collision, grasp slip, timeout), and domain-randomization parameters.

Robotics-Ready Delivery Formats

TELUS Digital exports JSON, CSV, or COCO bounding boxes, which is fine for 2D vision and unusable as embodied training data. Robotics models consume HDF5 hierarchical arrays for trajectories, MCAP containers for ROS2 streams, and Parquet for metadata. LeRobot expects episodes as HDF5 groups keyed `/observations/images/cam_high`, `/actions/joint_positions`, and `/episode_metadata`^[4].

Truelabel delivers that structure by default. A kitchen-manipulation bounty returns HDF5 files with `/rgb_wrist` (480×640×3 uint8), `/depth_wrist` (480×640 float32), `/joint_positions` (7-DOF float64), `/gripper_state` (binary), and `/task_success` (boolean), plus a hardware manifest (camera serial, firmware, calibration date), capture-environment metadata, and per-trajectory provenance. Format fidelity is where integration time hides. OpenVLA's 970,000-trajectory run spent real engineering converting many source datasets into one schema; truelabel arrives pre-converted to your target (RLDS, LeRobot, robomimic, CALVIN) with validation and loader code, so a delivery is `h5py.File('data.hdf5')` ready instead of a custom-parser project.

How Truelabel Delivers Physical AI Data

Procurement runs as a bounty, not an RFP. You specify robot platform (Franka FR3, UR10e, Stretch RE1), task domain, sensor suite, and episode count; vetted collectors bid with portfolio samples, hardware proof, and per-episode pricing; you pick on evidence. Every delivery carries the provenance a scraped folder cannot: contributor consent artifacts, location releases where applicable, and per-trajectory metadata that satisfies EU AI Act Article 10 training-data documentation.

1. 01

   Scope

   Post a bounty with task, environment, sensor suite (for example RealSense D435i on a Franka FR3), and volume. Intake validates feasibility and pricing before capture.

2. 02

   Bid

   Collectors submit portfolio samples, hardware proof (serial, calibration report, sample trajectory), and a per-episode rate. You compare on quality and price.

3. 03

   Capture

   The collector records in the target environment with approved hardware, time-syncing RGB-D, proprioception, and force-torque via ROS2 or MCAP timestamps.

4. 04

   Enrich

   Post-capture adds failure-mode tags, teleoperation metadata, and domain-randomization parameters that make episodes usable for policy learning, not just viewable.

5. 05

   Deliver

   Datasets ship in RLDS, LeRobot, MCAP, HDF5, or Parquet to S3, GCS, or Azure, with a sample packet and QA evidence before you commit to scale.

When to Choose Which

Match the vendor to the missing half. Choose TELUS Digital when the constraint is human judgment on data you already hold: 500,000 images needing boxes, 100,000 model outputs needing preference ranks, multilingual transcripts needing sentiment. Its contributor scale, domain verticals (medical, legal, financial)^[5], and post-training tooling (RLHF, red-teaming, constitutional feedback) are built for that. Multilingual grounding for VLA models like RT-2 is a genuine need, but downstream of the trajectories.

Choose truelabel when acquisition is the constraint: RGB-D manipulation episodes, LiDAR navigation runs, teleoperation demos, sensor-fusion datasets. Training diffusion policies or vision-language-action models needs diverse real-world episodes simulation cannot supply. BridgeData V2's 60,096 trajectories span just 24 environments, all captured in-house on one robot setup; a marketplace parallelizes that kind of capture across many collectors and far more settings at once. Around 10,000 vetted collectors across 100 countries deliver training-ready HDF5, MCAP, and Parquet with per-trajectory provenance, priced per episode (kitchen manipulation $40 to $80, warehouse navigation $60 to $120, bimanual assembly $100 to $200) with no annual minimum. Post a bounty, compare bids, and start training without a six-month procurement cycle.

Other Physical AI Data Alternatives

Scale AI runs an end-to-end physical AI engine (capture, annotation, QA), and its Universal Robots partnership signals enterprise traction, but access starts at large annual minimums that price out most labs and startups. Appen and Sama collect crowdsourced image and video, yet skew to 2D vision and autonomous-vehicle work without sensor fusion, trajectory formats, or teleoperation metadata; neither ships HDF5 or RLDS.

Labelbox, Encord, V7, and Roboflow are annotation platforms rather than capture marketplaces: they assume you already possess the sensor streams. For robotics teams the missing step is acquiring those streams, which is the gap truelabel fills. It does not compete with annotation tooling; it solves the upstream problem that tooling assumes is already done.

Related pages

Use these to move from category-level context into specific task, dataset, format, and comparison detail.

External references and source context

1. Appen AI Data

   TELUS Digital's multimodal and multilingual AI data service positioning

   appen.com ↩
2. appen.com data annotation

   Global contributor scale claims for annotation platforms

   appen.com ↩
3. Project site

   Open X-Embodiment aggregation methodology and dataset diversity

   robotics-transformer-x.github.io ↩
4. LeRobot dataset documentation

   LeRobot HDF5 episode structure with observation and action groups

   Hugging Face ↩
5. imerit.net resources

   Domain-specific annotation verticals (medical, legal, financial)

   imerit.net ↩
6. labelbox

   Labelbox annotation platform capabilities

   labelbox.com
7. encord

   Encord annotation and data management platform

   encord.com
8. Introduction to HDF5

   HDF5 hierarchical data format for trajectory storage

   The HDF Group
9. MCAP guides

   MCAP container format for ROS2 sensor streams

   MCAP
10. Apache Arrow Parquet files

    Apache Parquet columnar file format

    Apache Arrow
11. FR3 Duo

    Franka FR3 Duo bimanual robot platform

    franka.de
12. sama

    Sama AI data annotation and post-training services

    sama.com
13. Appen AI Data

    Appen AI data collection and annotation platform

    appen.com
14. cloudfactory.com accelerated annotation

    CloudFactory annotation services and quality workflows

    cloudfactory.com
15. appen.com data collection

    Multimodal and multilingual data collection breadth claims

    appen.com
16. ROS: an open-source Robot Operating System

    ROS (Robot Operating System) message types and ecosystem

    ICRA Workshop on Open Source Software
17. appen.com data annotation

    Fraud detection and quality assurance in distributed annotation

    appen.com
18. sama.com resources

    Enterprise service agreement pricing models for AI data services

    sama.com
19. iMerit model evaluation and training data

    Use case fit for annotation platforms vs data marketplaces

    imerit.net
20. LeRobot GitHub repository

    LeRobot dataset schema and training integration

    GitHub
21. truelabel physical AI data marketplace bounty intake

    Truelabel delivery timeline and bounty workflow

    truelabel.ai

FAQ