Skip to content
RFL_GLOBAL
中文
  • WAVE 5 // RESEARCH
  • SYNTHETIC DATA
  • SIM-TO-REAL

CORNUCOPIA

LARGE-SCALE SYNTHETIC VLA DATA

Reproduces InternData-A1: 630K+ manipulation trajectories, 7,433 hours of data, 4 embodiments, 70 tasks, 227 scenes. Proves synthetic data replaces real demonstrations at pretraining scale. Zero-shot sim-to-real transfer eliminates the data bottleneck.

MODULE STATUS: RESEARCH

PYGMALION

630K+

DIVISION
ANIMA
WAVE
W5
DOMAIN
SIMULATION
WAVE 5 // ANIMA SUITE
SIMULATION — LARGE-SCALE SYNTHETIC VLA DATA
CORNUCOPIA // W5 // 014/079
01THE CHALLENGE

REAL DATA DOESN'T SCALE

Training VLA models requires enormous datasets of robot manipulation demonstrations. Collecting real-world data is slow, expensive, and dangerous — each trajectory needs a physical robot, a human operator, and a controlled environment. At the scale needed for foundation models, real data collection becomes the bottleneck.

Current approaches require thousands of hours of teleoperation across diverse environments and embodiments. The cost grows linearly with dataset size, there's no way to parallelize physical robots, and rare failure modes are nearly impossible to capture. CORNUCOPIA proves that synthetic generation can replace real demonstrations at pretraining scale.

02THE SOLUTION

WHAT CORNUCOPIA DELIVERS

CORNUCOPIA generates large-scale synthetic manipulation datasets that match or exceed the quality of real demonstrations. The InternData-A1 reproduction pipeline creates 630K+ trajectories across 4 embodiments, 70 tasks, and 227 procedurally generated scenes — with zero-shot transfer to real robots.

PIPELINE

  1. 01Scene generation — procedural creation of 227 diverse manipulation environments with physics-accurate object properties and placements
  2. 02Trajectory synthesis — automated generation of 630K+ manipulation trajectories across 70 task types with motion planning and physics simulation
  3. 03Multi-embodiment rendering — trajectories replayed across 4 robot embodiments with embodiment-specific kinematics and viewpoints
  4. 04Sim-to-real validation — zero-shot transfer testing on physical robots proves synthetic data quality matches real demonstrations

CAPABILITIES

  • MASSIVE SCALE630K+ trajectories, 7,433 hours — dataset sizes impossible to collect with real robots→ >80% cost reduction compared to real-world data collection
  • MULTI-EMBODIMENT4 robot embodiments with accurate kinematics — train once, deploy anywhere→ Single pipeline generates data for multiple robot platforms simultaneously
  • TASK DIVERSITY70 manipulation tasks across 227 procedurally generated scenes→ Covers manipulation primitives from grasping to multi-step assembly
  • ZERO-SHOT TRANSFERSynthetic-trained policies transfer directly to real robots without fine-tuning→ Domain randomization and physics accuracy close the sim-to-real gap
03ENGINEERING

WHY THIS IS HARD

Generating synthetic data that actually transfers to real robots requires solving multiple compounding challenges:

  1. 01Physics fidelity: simulated contact dynamics must match real-world friction, deformation, and mass — even small errors compound across long trajectories
  2. 02Visual domain gap: rendered images must cover the visual diversity of real environments — lighting, textures, camera noise, occlusion patterns
  3. 03Trajectory quality: automated motion planning must produce naturalistic, human-like manipulation — not just kinematically valid but dynamically smooth
  4. 04Scale engineering: generating 630K trajectories requires massively parallel simulation with consistent quality across millions of physics steps
  5. 05Embodiment transfer: kinematics and dynamics differ between robot platforms — trajectories must be retargeted without losing task semantics

CORNUCOPIA solves this through aggressive domain randomization, physics-accurate contact simulation via PYGMALION, and a quality validation pipeline that filters trajectories before they enter the training set.

04BENCHMARKS

SYSTEM PERFORMANCE

Measured across the InternData-A1 reproduction pipeline:

SYSTEM PERFORMANCE
METRICVALUE
Total Trajectories630K+ manipulation demonstrations
Data Hours7,433 hours of simulation data
Cost Reduction>80% vs real-world data collection
Sim-to-Real TransferZero-shot — no real fine-tuning required
05BUILD STATUS

WHAT'S BUILT TODAY

1/6 COMPONENTS COMPLETE
WHAT'S BUILT TODAY
COMPONENTSTATUSNOTES
Dataset ArchitectureCOMPLETEInternData-A1 schema and storage pipeline production-ready
Trajectory GeneratorIN PROGRESSMotion planning + physics simulation pipeline scaling to 630K
Multi-EmbodimentPLANNED4 embodiment retargeting system designed, awaiting integration
Sim-to-Real TransferPLANNEDDomain randomization and validation protocol designed
Core modelsIN PROGRESSPYGMALION physics backbone being adapted for trajectory synthesis
API layerPLANNEDDataset access API designed, pending generation pipeline completion
06APPLICATIONS

WHERE CORNUCOPIA DEPLOYS

  • APP_01

    VLA PRETRAINING

    Foundation model pretraining at scale — 630K+ trajectories provide the data volume needed for VLA models to learn general manipulation without expensive real-world collection.

  • APP_02

    MULTI-ROBOT TRAINING

    Train once across 4 embodiments — a single synthetic dataset covers multiple robot platforms, eliminating per-platform data collection.

  • APP_03

    RARE SCENARIO COVERAGE

    Generate trajectories for edge cases and failure modes that are dangerous or impossible to collect in the real world — dropped objects, near-collisions, recovery behaviors.

07TECHNOLOGY

UNDER THE HOOD

FOUNDATION: INTERNDATA-A1 REPRODUCTION

  • 630K+ manipulation trajectories across 70 task categories
  • 227 procedurally generated scenes with physics-accurate properties
  • 4 robot embodiments with full kinematic and dynamic modeling
  • 7,433 hours of rendered multi-viewpoint simulation data

CORNUCOPIA IMPLEMENTATION

  • Procedural scene generation with physically-based materials and object placement
  • Automated trajectory planning with collision-aware motion primitives
  • Domain randomization across lighting, textures, camera parameters, and physics noise
  • Quality validation pipeline filtering trajectories by task completion and physics consistency

ANIMA MODULE INTEGRATION

  • PYGMALION provides the physics simulation backbone for trajectory generation
  • Generated datasets feed directly into VLA pretraining pipelines
  • Quality metrics align with downstream evaluation benchmarks

PIPELINE SPECS

  • 630K+ trajectories generated
  • 70 task categories
  • 227 procedural scenes
  • 4 robot embodiments
  • 7,433 simulation hours
  • Zero-shot transfer validated
08PAPERS

RESEARCH BASIS

  1. [01]InternData-A1 reproduction — proving synthetic manipulation data replaces real demonstrations at pretraining scale with zero-shot sim-to-real transfer