Galbot

Chinese robotics unicorn achieving large-scale commercial deployment with Synthetic Data First strategy

Overview

Galbot (Galaxy General Robotics) is a China-based full-stack robotics unicorn that achieved the first large-scale commercial deployment in the VLA field with its “Synthetic First, Real Data as a Complement” philosophy. They build production-ready systems with 99% synthetic + <1% real data.

ItemDetails
FoundedMay 2023
HeadquartersBeijing, China (Haidian District)
R&D CentersBeijing, Shenzhen, Suzhou, Hong Kong
Co-foundersHe Wang (CTO), TengZhou Yao
AffiliationsBeijing Academy of AI (BAAI), PKU EPIC Lab
Total Funding$800M+ (as of December 2025)
Valuation$3B (as of December 2025)
Mission”Make robots for every industry and every home”

He Wang: Born 1992, graduated from Tsinghua University, earned PhD from Stanford University in 2021 (advisor: Leonidas J. Guibas). Currently an Assistant Professor at Peking University CFCS and founder of PKU EPIC Lab.


Key Products

Galbot G1

Semi-humanoid mobile manipulator:

SpecificationValue
Height173cm
Weight85kg
Arm Span190cm
Max Reach Height240cm
Payload5kg (single arm)
Battery Life10 hours continuous
Item Handling5,000+ SKU types

Galbot S1

Industrial heavy-duty robot (launched 2025):

SpecificationValue
Payload50kg (continuous dual-arm)
ApplicationManufacturing, heavy industry

Key Achievements

Commercial Deployment

MetricValue
Galbot Store30+ Chinese cities
Smart Pharmacy/Warehouses30+ fully unmanned
Workers per warehouse0
MTBF1 month+
Continuous operation10 hours/charge

Key Partnerships

  • Manufacturing: CATL, Bosch, Toyota, Hyundai
  • Healthcare: Xuanwu Hospital (patient rooms, pharmacies, guidance systems)

GraspVLA Performance (LIBERO Zero-shot)

ModelLongGoalObjectCondition
OpenVLA33.7%56.6%65.4%fine-tuned
π062.7%79.4%93.8%fine-tuned
GraspVLA82.0%91.2%94.1%zero-shot

Outperforms fine-tuned models without any fine-tuning


Technical Architecture

Cerebrum-Cerebellum Structure

Dual system inspired by human brain architecture:

ComponentRoleImplementation
CerebrumHigh-level policy - what to doVLA (Imitation + Web Grounding)
CerebellumLow-level motor - how to doRL-based 100Hz control

GraspVLA

ComponentSpecification
Vision EncoderDINO-v2 + SigLIP
LLM BackboneInternLM2 1.8B
Action ExpertFlow Matching
Training Data1B synthetic + 100M+ web grounding
Training Cost~$5,000 (160×RTX 4090, 10 days)

Data Strategy

Synthetic Data Pipeline

Scene Synthesis

Trajectory Generation
  ├─ Physics-based Energy Optimization (DexGraspNet)
  ├─ Human Videos → Synthetic (GenHOI)
  └─ Large-Scale RL (UniDexGrasp++)

Validation & Rendering
  ├─ MuJoCo physics validation
  └─ Isaac Sim ray-tracing

Sim2Real Transfer (1B frames convergence)

Scaling Law Discovery

  • 1B frames: sim/real performance curves converge
  • Sim2real gap decreases with data scale
  • Scale impossible to achieve via teleoperation

Data Scale

Data TypeScale
Synthetic trajectoriesBillion-scale
DexGraspNet 2.0 grasps426M
Web grounding (GRIT)100M+ bboxes
Real data ratio<1%

Research Portfolio

Grasping

ResearchVenueKey Contribution
DexGraspNetICRA 2023 FinalistMillion-scale dexterous grasp
UniDexGrasp++ICCV 2023 FinalistLarge-scale RL, policy distillation
DexGraspNet 2.0CoRL 20247 embodiments, 426M grasps
DexonomyRSS 2025100+ grasp taxonomy
GraspVLA2025Billion-scale synthetic VLA

Sim2Real Solution

ResearchKey Contribution
DexNDMWorld model for sim2real gap correction

DexNDM Approach:

  1. Train generalist policy in simulator
  2. Train neural dynamics model with small real data
  3. Fix sim2real gap via back-propagation (differentiable)
ResearchKey Contribution
NavFoMCross-embodiment navigation foundation model
TrackVLA30+ minute human tracking

Capability Assessment (6 Axes)

AxisRatingEvidence
Long-horizonCloth folding deformable manipulation sim2real
PrecisionDriver, hammer manipulation via DexNDM
Deployment RobustnessMTBF 1 month+, 30+ unmanned warehouses
Multi-taskGraspVLA zero-shot 82% (LIBERO)
Cross-embodiment7 hand embodiments supported
Zero-shotOutperforms π0 fine-tuned

Limitations

Acknowledged Limitations (from presentation)

LimitationDescription
Task ScopeGrasping-specialized, not yet generalist
Dexterous Sim2RealInitially failed → solved with DexNDM
Specialist vs GeneralistMotion control still specialist policy

Analytical Limitations

LimitationDescription
Not Pure Sim2RealIncludes web grounding (100M+ real images)
Lack of Theoretical ExplanationNo theory on why 1B frames converge
Embodiment GeneralizationFranka Panda-centric, limited cross-robot transfer

vs Competitors

ItemGalbotPhysical IntelligenceFigure
Deployment Scale30+ cities + 30+ warehousesResearch demoResearch demo
Data Strategy99% syntheticCross-embodiment realVLA (Helix)
StrengthOnly commercial deploymentPursuing generalityIndustrial optimization
WeaknessGrasping-specializedWeak zero-shotNo consumer support

References


See Also