Ingestion & Standardization
Unify ego, exo and multi-view video with teleoperation commands, kinematics, joint states, force-torque, telemetry, actions and task metadata in consistent, versioned schemas.
PHYSICAL AI DATA INTELLIGENCE
THE MISSING INFRASTRUCTURE
DATA
INTELLIGENCE
FOR PHYSICAL AI.
HokFANG transforms fragmented multimodal robotics data — egocentric and exocentric video, synchronized multi-view capture, teleoperation demonstrations, kinematics and joint states, force-torque signals, sensor telemetry and action logs — into structured, traceable and quality-scored datasets for training, evaluation and deployment.
MULTI-VIEW VIDEOTELEOPERATIONKINEMATICSFORCE / TORQUE
ROBOT VISION // MANIPULATION LEARNING LIVE PIPELINE
01 TASK ANALYSIS02 DATA CAPTURE03 QUALITY EVALUATION04 MODEL VALIDATION05 TARGETED RECOLLECTION
RGB VIDEO DEPTH MULTI-VIEW TELEOPERATION KINEMATICS JOINT STATES FORCE / TORQUE ACTIONS RGB VIDEO DEPTH TELEOPERATION KINEMATICS
[ 001 // THE MARKET GAP ]
PHYSICAL AI HAS A
DATA-READINESS PROBLEM.
Robotics teams generate RGB and depth video, kinematics, joint states, force-torque signals, telemetry, actions, annotations and task context across real-world environments. Yet much of it remains fragmented, inconsistently structured and difficult to evaluate.
Egocentric and exocentric views show what happened; teleoperation commands and robot-state data explain how it happened. Reliable training requires every modality to be synchronized, traceable, quality-scored and packaged as one coherent episode.
[ 002 // WHAT HOKFANG BUILDS ]
A DATA INTELLIGENCE PLATFORM
FOR MODEL-READY ROBOTICS DATA.
Data Quality Intelligence
Measure multimodal completeness, temporal synchronization, diversity, task coverage, environmental conditions and failure cases.
Model Readiness & Validation
Evaluate whether datasets are suitable for training, fine-tuning, benchmarking or real-world task validation.
Provenance & Governance
Track data origin, permissions, transformations, versions and delivery requirements throughout the lifecycle.
[ 003 // CORE COMPETENCY ]
REAL SCENE → DATA DESIGN →
MODEL EFFECT → COMPLETE WORKFLOW.
HokFANG does not simply deliver data. We connect real tasks, data specifications, usage rights and model validation into a reusable workflow.
01
REAL TASKFAILURE OBSERVEDCAM_03
REAL-WORLD SCENE
Industrial tasksFailures · constraints · edge conditions
→02
COLLECTION SPEC
SCENARIOS12
CAMERAS03
TARGET EPISODES2.4K
RIGHTS ✓DATA ORCHESTRATION
Controlled provenanceWhat to collect · how much
→03
MULTI-VIEW
KINEMATICS
ACTIONS
FORCE / TORQUE
SYNC + QA98.4% PASSINTELLIGENT PROCESSING
Parse · QA · alignGeneralize · score · standardize
→04
FIXED MODEL / TASK
BASELINE58%
HOKFANG88%
+30 PTS VALIDATEDMODEL VALIDATION
Training-readyPerformance comparison · fit evaluation
→05
FAILURE ATTRIBUTION
OCCLUSION27%
DARK FABRIC34%
SLEEVE FLIP19%
→ NEXT 2,400 EPISODESFEEDBACK CAPTURE
Failure attributionData gaps · next collection
MODEL FEEDBACKTHE NEXT COLLECTION STARTS WITH THE LAST FAILURE.
[ 004 // MODEL EVALUATION ]
PROVE THE DATA ON THE MODEL
SAME MODEL.
SAME TASK.
MEASURED LIFT.
HokFANG evaluates the delivered dataset on the customer's robotics model. A controlled A/B protocol holds the model, task suite and test conditions constant, isolating the measurable contribution of the data.
FIXED MODEL FIXED TASK SUITE CONTROLLED TEST SET
EVAL_RUN // GARMENT_MANIPULATION_V07 REPORT COMPLETE
MODEL π₀.₅ CHECKPOINT 42KTEST SET 1,200 EPISODESSEEDS 5CONDITIONS CONTROLLED
CONTROL / BASELINE DATA58.2%TASK SUCCESS
95% CI 55.4–61.0
VALIDATED LIFT+29.7POINTSp < 0.01
TREATMENT / HOKFANG DATA87.9%TASK SUCCESS
95% CI 85.9–89.9
GENERALIZATION+21.4 pts
OOD FAILURE RATE−22.1 pts
VALID EPISODES+29.0 pts
RESIDUAL FAILURE CLUSTERSDARK FABRIC 34%PARTIAL OCCLUSION 27%SLEEVE INVERSION 19%
NEXT COLLECTION SPEC2,400 TARGETED EPISODES
3 environments · 4 lighting conditions · 2 garment geometries
[ ROBOT TASKS // COMPLETE ACTION SEQUENCES ]
REAL ACTION.
FRAME BY FRAME.
Each sequence shows the complete physical task—not a symbolic animation. The object moves only after contact, manipulation stays mechanically plausible, and the final state is visibly verified.
01 APPROACH02 GRASP03 TRANSFER04 PLACE
PICK & PLACE
Detect → approach → close gripper → lift → transfer → release inside the destination.
01 LOCALIZE02 GRASP03 FOLD04 VERIFY
GARMENT FOLDING
Localize sleeve keypoints → bimanual grasp → sequential folds → verify the final geometry.
01 DETECT02 PICK03 CLASSIFY04 SORT
VISION SORTING
Detect and classify the object → grasp → transfer → release into the correct bin.
[ 005 // CLOSED-LOOP MODEL VALIDATION ]
WE DO NOT STOP
AT DATA DELIVERY.
HokFANG runs the delivered data on the robotics model and tests the model inside the target task. We prove whether the data improves behavior, explain where the model still fails, and turn that evidence into the next collection plan.
HokFANG Data
Quality-scored, task-specific episodes enter the evaluation.
Customer Model
The same model receives baseline data and HokFANG data.
Task Simulation
Fixed tasks and test conditions isolate the effect of the data.
Measured Lift
Compare success, generalization, robustness and failure rate.
Next Data Plan
Failures become precise scenarios for targeted recollection.
BASELINE MODEL58%TASK SUCCESS
+30 PTSWITH HOKFANG DATA88%VALIDATED IMPROVEMENT
CONCLUSION // Improvement is strongest in standard folds. Remaining failures cluster around dark fabrics, partial occlusion and sleeve inversion. NEXT COLLECTION: 2,400 targeted episodes across those three failure conditions.
MODEL FAILURE EVIDENCEDRIVES THE NEXT DATA COLLECTIONTHE LOOP GETS SMARTER
[ 006 // BUILT FOR ]
WHO HOKFANG
IS BUILDING FOR.
Consistent, traceable datasets without distracting the core engineering team.
Well-documented real-world data for reproducible research and benchmarking.
Dataset quality, task coverage, provenance and model suitability before training.
Operational data prepared for applied AI, automation and deployment validation.