Lightwheel

Access 100,000+ Hours of Teleoperation Data

Curate from rare, high-quality training datasets
Train large-scale physical AI models with comprehensive teleoperation data that establishes new performance benchmarks

Simulation Data Collection

Data Collection

Fixed Solution

Robot

Hand

Teleops Solution

Unitree G1

Default 3 Finger

1. AR/VR - Oculus, Apple Vision Pro, Pico Controller or + Glove

2. Exoskeleton - JoyLo

3. Motion Capture Suit - IO (Suit) + IO (Glove), IO (Suit) + Pico Controller

4. Any other combinations

Inspire

Glaxea R1 Pro

Arx X7s

Frank Panda

Customised Solution

Only difference is pricing

Simulation Data Collection

Data Augmentation

Need-Factors in Teleoperation (Real World + Simulation)

When collecting teleoperation data, your face fundamental limitations regardless of whether you operate in simulation or the real world

Need-Factor

Why It Exists

Impact on Policy Robustness

Environmental Diversity

Teleop captures only specific lighting,backgrounds,weather conditions present during collection

Policy fails when deployed under different lighting (night shift vs day shift) or environments (cluttered vs clean workspace)

Object Variation

Demonstrations show limited object poses, orientations, conditions(pristinevs damaged)

Policy cannot handle rotated parts,defects, wear-and-tear, or natural variation in manufacturing

Physical Perturbations

Real/sim teleop happens under nominal conditions-no external disturbances, sensor noise, calibration drift

Policy brittle to real-world messiness:vibrations, sensor glitches, slight misalignments

Failure Mode Coverage

Humans demonstrate success cases;rarely show recovery from failures, edge cases, or anomalies

Policy has no training data for: dropped objects, collisions, occlusions, unexpected obstacles

Scale Constraints

Collecting 10,000+ diverse real demos is prohibitively expensive; sim teleop lacks visual realism at scale

Insufficient data diversity leads to overfitting and poor generalization

Need-Factor

Environmental Diversity

Why It Exists

Teleop captures only specific lighting,backgrounds,weather conditions present during collection

Impact on Policy Robustness

Policy fails when deployed under different lighting (night shift vs day shift) or environments (cluttered vs clean workspace)

Need-Factor

Object Variation

Why It Exists

Demonstrations show limited object poses, orientations, conditions(pristinevs damaged)

Impact on Policy Robustness

Policy cannot handle rotated parts,defects, wear-and-tear, or natural variation in manufacturing

Need-Factor

Physical Perturbations

Why It Exists

Real/sim teleop happens under nominal conditions-no external disturbances, sensor noise, calibration drift

Impact on Policy Robustness

Policy brittle to real-world messiness:vibrations, sensor glitches, slight misalignments

Need-Factor

Failure Mode Coverage

Why It Exists

Humans demonstrate success cases;rarely show recovery from failures, edge cases, or anomalies

Impact on Policy Robustness

Policy has no training data for: dropped objects, collisions, occlusions, unexpected obstacles

Need-Factor

Scale Constraints

Why It Exists

Collecting 10,000+ diverse real demos is prohibitively expensive; sim teleop lacks visual realism at scale

Impact on Policy Robustness

Insufficient data diversity leads to overfitting and poor generalization

Mimicgen Vs Cosmos

Dual-Engine Solution: MimicGen + Cosmos Predict

Engine

MimicGen

Cosmos Predict

Technology

Object-centric trajectory adaptation

Generative world model

Input

10-200 human demonstrations

Text/image/video prompts

Process

Transform demos to new object poses

Generate novel visual scenarios

Strength

Geometric accuracy, task structure preservation

Visual diversity,edge cases

Multiplier

200x from 10 demos

50x per scenario

Use Case

Primary spatial variation engine

Appearance/condition augmentation

Engine

Technology

MimicGen

Object-centric trajectory adaptation

Cosmos Predict

Generative world model

Engine

Input

MimicGen

10-200 human demonstrations

Cosmos Predict

Text/image/video prompts

Engine

Process

MimicGen

Transform demos to new object poses

Cosmos Predict

Generate novel visual scenarios

Engine

Strength

MimicGen

Geometric accuracy, task structure preservation

Cosmos Predict

Visual diversity,edge cases

Engine

Multiplier

MimicGen

200x from 10 demos

Cosmos Predict

50x per scenario

Engine

Use Case

MimicGen

Primary spatial variation engine

Cosmos Predict

Appearance/condition augmentation

Access 100,000+ Hours of Teleoperation Data

Curate from rare, high-quality training datasets
Train large-scale physical AI models with comprehensive teleoperation data that establishes new performance benchmarks

Simulation Data Collection

Data Collection

Fixed Solution

Robot

Hand

Teleops Solution

Unitree G1

Default 3 Finger

1. AR/VR - Oculus, Apple Vision Pro, Pico Controller or + Glove

2. Exoskeleton - JoyLo

3. Motion Capture Suit - IO (Suit) + IO (Glove), IO (Suit) + Pico Controller

4. Any other combinations

Inspire

Glaxea R1 Pro

Arx X7s

Frank Panda

Customised Solution

Only difference is pricing

Simulation Data Collection

Data Augmentation

Need-Factors in Teleoperation (Real World + Simulation)

When collecting teleoperation data, your face fundamental limitations regardless of whether you operate in simulation or the real world

Need-Factor

Why It Exists

Impact on Policy Robustness

Environmental Diversity

Teleop captures only specific lighting,backgrounds,weather conditions present during collection

Policy fails when deployed under different lighting (night shift vs day shift) or environments (cluttered vs clean workspace)

Object Variation

Demonstrations show limited object poses, orientations, conditions(pristinevs damaged)

Policy cannot handle rotated parts,defects, wear-and-tear, or natural variation in manufacturing

Physical Perturbations

Real/sim teleop happens under nominal conditions-no external disturbances, sensor noise, calibration drift

Policy brittle to real-world messiness:vibrations, sensor glitches, slight misalignments

Failure Mode Coverage

Humans demonstrate success cases;rarely show recovery from failures, edge cases, or anomalies

Policy has no training data for: dropped objects, collisions, occlusions, unexpected obstacles

Scale Constraints

Collecting 10,000+ diverse real demos is prohibitively expensive; sim teleop lacks visual realism at scale

Insufficient data diversity leads to overfitting and poor generalization

Need-Factor

Environmental Diversity

Why It Exists

Teleop captures only specific lighting,backgrounds,weather conditions present during collection

Impact on Policy Robustness

Policy fails when deployed under different lighting (night shift vs day shift) or environments (cluttered vs clean workspace)

Need-Factor

Object Variation

Why It Exists

Demonstrations show limited object poses, orientations, conditions(pristinevs damaged)

Impact on Policy Robustness

Policy cannot handle rotated parts,defects, wear-and-tear, or natural variation in manufacturing

Need-Factor

Physical Perturbations

Why It Exists

Real/sim teleop happens under nominal conditions-no external disturbances, sensor noise, calibration drift

Impact on Policy Robustness

Policy brittle to real-world messiness:vibrations, sensor glitches, slight misalignments

Need-Factor

Failure Mode Coverage

Why It Exists

Humans demonstrate success cases;rarely show recovery from failures, edge cases, or anomalies

Impact on Policy Robustness

Policy has no training data for: dropped objects, collisions, occlusions, unexpected obstacles

Need-Factor

Scale Constraints

Why It Exists

Collecting 10,000+ diverse real demos is prohibitively expensive; sim teleop lacks visual realism at scale

Impact on Policy Robustness

Insufficient data diversity leads to overfitting and poor generalization

Mimicgen Vs Cosmos

Dual-Engine Solution: MimicGen + Cosmos Predict

Engine

MimicGen

Cosmos Predict

Technology

Object-centric trajectory adaptation

Generative world model

Input

10-200 human demonstrations

Text/image/video prompts

Process

Transform demos to new object poses

Generate novel visual scenarios

Strength

Geometric accuracy, task structure preservation

Visual diversity,edge cases

Multiplier

200x from 10 demos

50x per scenario

Use Case

Primary spatial variation engine

Appearance/condition augmentation

Engine

Technology

MimicGen

Object-centric trajectory adaptation

Cosmos Predict

Generative world model

Engine

Input

MimicGen

10-200 human demonstrations

Cosmos Predict

Text/image/video prompts

Engine

Process

MimicGen

Transform demos to new object poses

Cosmos Predict

Generate novel visual scenarios

Engine

Strength

MimicGen

Geometric accuracy, task structure preservation

Cosmos Predict

Visual diversity,edge cases

Engine

Multiplier

MimicGen

200x from 10 demos

Cosmos Predict

50x per scenario

Engine

Use Case

MimicGen

Primary spatial variation engine

Cosmos Predict

Appearance/condition augmentation