Real-world manipulation data for autonomy.

Explore related data pages

Why It's Different

Built for model teams
that need signal, not demos.

01

Real production context

Collected inside active warehouse and industrial workflows, with natural variance, interruptions, and recovery behaviour that matter for training and evaluation. Not a staged lab environment.

02

Engineer-ready structure

Episodes include synchronised observations, actions, rewards, dones, per-step metadata, and external JSON. Teams can evaluate format compatibility and ingest data without custom preprocessing.

03

Rich multi-modal observability

Multi-view RGB and depth, robot pose and kinematics, joint angles, gripper state, wrist force/torque (Fx/Fy/Fz in Newtons per frame), and prompt-conditioned metadata where relevant.

04

Flexible delivery format

Start with standard datasets for evaluation and benchmarking. Expand into custom collection programmes with defined skills, task structures, specific embodiments, and milestone-based delivery.

What You Get

Designed for pre-training,
post-training, and evaluation.

Telepath datasets are structured for teams working across the full model development cycle.

Episode containers

.rlog format with observations, actions, rewards, dones, infos, and episode-level metadata. Companion JSON for prompts, tags, and post-processed corrections.

Multi-view visual streams

RGB and depth from multiple viewpoints per episode, covering pick, place, transfer, and recovery sequences under real operating lighting conditions.

Robot state & telemetry

Robot pose (6-DoF ENU frame), joint angles, gripper state (Robotiq 2F-140, Schmalz SBPG, Robotiq ePick), and wrist F/T sensor data per frame.

Prompt-conditioned tasks

Multiskill episodes with conditional inputs and prompt-driven manipulation. Suitable for instruction-following and generalisation benchmarks.

Full schema documentation

Loading workflows, robot setup documentation, coordinate system reference, and data dictionary. No reverse-engineering required on ingestion.

Episode format reference

Standard episode
container structure

episode_id str
observations dict[str, array]
actions array[float]
rewards array[float]
dones array[bool]
infos array[dict]
force_torque Fx/Fy/Fz · N/frame
pose 6-DoF · ENU frame
metadata companion .json

Format is consistent across collection programmes. Evaluation samples include full schema and loading scripts on request.

Embodiment & Task Coverage

More than one collection
format. More than one use.

Arm configuration

Single-arm & dual-arm

Pick, place, and transfer episodes across single and dual-arm setups. Robotiq and Schmalz gripper variants. SKU-conditioned and prompt-conditioned collection.

Extended embodiment

Humanoid-style setups

Unitree G1 configurations included in the dataset. Suitable for teams working on cross-embodiment transfer or humanoid-specific policy training.

Task structure

Narrow to long-horizon

From single-skill picks to multiview pick-and-place and multi-step prompt-driven manipulation. Task diversity supports both narrow optimisation and generalisation research.

Data collected in production environments gives model teams access to natural behavioural variance, interruptions, and recovery episodes that matter for training — not only for narrow task optimisation, but for broader generalisation across operating conditions and embodiments.

Get Started

Need a dataset for a specific embodiment, benchmark, or manipulation capability?

We can start with a sample set, align on technical requirements, and define a collection plan around your training or evaluation goals.

Discuss a collection plan