Training data—not a proxy.
UMI-L produces clean, precisely synchronized demonstrations designed for direct use in real-robot training and evaluation workflows.
PHYSICAL AI DATA INTELLIGENCE
DATA
INTELLIGENCE
FOR PHYSICAL AI.
DATA GOVERNANCE & QUALITY GATINGMULTIMODAL ENRICHMENTDATA SEMANTIC PROFILINGMODEL READINESS
ROBOT VISION // MANIPULATION LEARNING LIVE PIPELINE
RGB VIDEO DEPTH MULTI-VIEW TELEOPERATION KINEMATICS JOINT STATES FORCE / TORQUE ACTIONS RGB VIDEO DEPTH TELEOPERATION KINEMATICS
HOKFANG PRESENTS // A FIRST LOOK PRE-RELEASE SYSTEM
INTRODUCINGUMI - L.
Robot-grade data.
Captured in the real world.
Captured in the real world.
UMI-L is a complete data collection system for five-finger dexterous manipulation. It supports dexterous hands with up to 22 actively controlled degrees of freedom, with adaptable hardware and SDK support for additional dexterous hands and grippers. No robot platform is required.
REQUEST EARLY ACCESS ↗UP TO 22 DoFNO ROBOT PLATFORMREAL CONTACT FEEDBACKLOW-COST PROCESSINGROBOT-READY OUTPUT
UMI-L // MULTIMODAL SIGNAL CAPTURE PRECISELY SYNCHRONIZED
Direct control. Real contact.
The operator controls the dexterous hand directly and feels real contact feedback, enabling natural in-the-wild collection without a robot body.
The complete demonstration.
First-person video, bimanual spatial poses, dexterous-hand operation data, contact, and force-control signals are recorded on one timeline.
Scale it. Deploy it.
Low-cost pre- and post-processing supports multi-scene, coordinated five-finger datasets at scale. Deploying on the same hand platform reduces data adaptation and policy-transfer work.
UMI-L / HIGH-FIDELITY EMBODIED DATA COLLECTIONNOW IN DEVELOPMENT // EARLY ACCESS INQUIRIES OPEN
FROM RAW COLLECTIONS TO
TRAINING-READY EPISODES.
RAW COLLECTIONINGEST
RGB / DEPTHSTATE / ACTIONFORCE / TELEOP
QUALITY GATEGOVERN
VALID EPISODESALIGNED STREAMSTRACEABLE LINEAGE
ENRICHED SIGNALSPROFILE
KINEMATICSTRAJECTORIESSEMANTICS
TRAINING-READY DATADELIVER
VERSIONEDQUALITY-SCOREDMODEL-READY
01MULTIMODAL SIGNAL EXTRACTION
WRIST TRAJECTORY // TEMPORAL LIVE EXTRACTION
FRAME-ALIGNED // QC PASS
Wrist Motion Trajectory
PARALLEL SIGNALS→SYNCED
HAND JOINTS // KINEMATICS LIVE EXTRACTION
FRAME-ALIGNED // QC PASS
Joint & Grip State
02HUMAN VIDEO → ROBOT REPRESENTATION
SOURCE // EVERYDAY VIDEO LIVE EXTRACTION
FRAME-ALIGNED // QC PASS
Unstructured Human Demonstration
ACTION RETARGETING→SYNCED
OUTPUT // ROBOT-LEARNABLE LIVE EXTRACTION
FRAME-ALIGNED // QC PASS
Robot-Compatible Action Representation
03SINGLE VIEW → WORLD-MODEL COMPLETION
SOURCE // OBSERVED VIEW LIVE EXTRACTION
FRAME-ALIGNED // QC PASS
Raw Single-View Demonstration
GENERATIVE VIEW COMPLETION→SYNCED
OUTPUT // WORLD MODEL LIVE EXTRACTION
FRAME-ALIGNED // QC PASS
World-Model Multi-View Completion
KNOW WHAT YOUR DATA
CAN ACTUALLY SUPPORT.
CORPUS EVIDENCE // ILLUSTRATIVE PROFILE
Uncurated Collection vs Representative Mix
83PROFILE INDEX
41UNCURATED
HOKFANG DELIVERED BATCHUNCURATED COLLECTIONILLUSTRATIVE AXES
DIVERSITY28 → 82
UNCURATED28
HOKFANG82
DEMONSTRATION COMPLETENESS58 → 90
UNCURATED58
HOKFANG90
TEMPORAL SYNCHRONIZATION74 → 88
UNCURATED74
HOKFANG88
EXECUTION SMOOTHNESS46 → 84
UNCURATED46
HOKFANG84
RECOVERY BEHAVIOR18 → 78
UNCURATED18
HOKFANG78
BALANCED COVERAGE22 → 76
UNCURATED22
HOKFANG76
VALIDATE PERFORMANCE.
FRAME BY FRAME.
MEASURED EXAMPLE
REPORT COMPLETEEvaluation results, not internal recipes.
NARROW DATA MIX8%TASK SUCCESS
→FULL DATA MIX78%TASK SUCCESS
Same task evaluation context
MEASURED OUTCOME
DELIVERABLESDATASET SCORECARD GAP MAP READINESS REPORT CUSTOM TRAINING-READY DATASET
EVALUATION ROLLOUT
BIMANUAL CUBE TRANSFER
01 APPROACH02 GRASP03 TRANSFER04 PLACE
PICK & PLACE
01 LOCALIZE02 GRASP03 FOLD04 VERIFY
GARMENT FOLDING