An operator wearing an AR headset and haptic gloves teleoperating a dual-arm robot, multiple camera-feed monitors showing the follower robot
Home / Data collection

Capture the world, three ways.

Teleoperation for fidelity. UMI for real-world scale. Simulation for infinite, perfectly-labeled variation. We run all three — and blend them per task.

TELEOPERATION RIG
LEADER ARMFOLLOWER ARM
latency 12 msjoint mirror 1:1
01 / TELEOPERATION

Expert demonstrations on real hardware.

Trained operators drive your robot — or ours — through leader–follower rigs, VR, and exoskeletons, recording synchronized multi-view RGB-D, joint states, end-effector pose, gripper force, and tactile streams on one clock.

This is the gold standard for manipulation data: real dynamics, real contact, real sensor noise — exactly the distribution your policy will see at deployment.

  • leader–follower rigs
  • VR / exoskeleton
  • bimanual + mobile bases
  • force & tactile capture
  • scripted scene variation
Why our capture is different

Every clip arrives training-ready.

Raw recordings aren't enough. Whatever the method, the output is the same: every angle on one clock, sliced into clean, labeled episodes you can train on directly.

01

Synced to one clock

Every camera and sensor is hardware-synchronized and frame-stamped, so vision, depth, force, and joint states all line up to the millisecond.

02

Split into clean episodes

Continuous runs are segmented into individual task episodes with clear start and end points — the unit your policy actually trains on.

03

Labeled with full provenance

Each episode carries task metadata, operator ID, and scene context, so you always know exactly what you're training on and where it came from.

UMI · IN-THE-WILD
kitchenwarehouseretailhome
handheld · no robot requiredepisodes/hr 120+
02 / UMI & DexUMI

Human-speed capture, anywhere.

UMI is a handheld gripper with a wrist fisheye and onboard pose tracking — collectors demonstrate tasks anywhere, no robot on site. We recover 6-DoF trajectories with SLAM and retarget to your embodiment.

For multi-finger hands, DexUMI goes further: the collector wears a hand exoskeleton tuned to your robot hand, capturing dexterous motion with real tactile feedback, while a vision pipeline swaps in your robot's hand frame-by-frame. Published results report ~3× faster collection than teleoperation.

  • handheld grippers
  • DexUMI exoskeleton
  • dexterous, multi-finger
  • tactile + 6-DoF pose
  • robot-hand inpainting
  • embodiment retargeting
SIMULATED EPISODES
SIM · physics onseed 04821
domain randomizationvariants
03 / SIMULATION

Infinite variation, perfect labels.

Built on Isaac Sim and MuJoCo with assets from our sim-to-real library, we generate photoreal, physics-validated episodes with domain randomization over lighting, texture, pose, and dynamics.

Ground truth comes free — masks, depth, 6-DoF poses, contact events — and unsafe or impossible edge cases are just another seed.

  • Isaac Sim · MuJoCo
  • domain randomization
  • free ground-truth labels
  • rare-event synthesis
  • sim-to-real validated
Start collecting

Scope a capture pilot.

Tell us the task and the embodiment. We'll recommend a teleop / UMI / sim mix and ship a benchmark batch you can train on.

Scope a pilot → or email contact@intellisant-ai.com