Perception under partial observability
Developing perception systems that integrate RGB-D, audio, motion, tactile, and proprioceptive signals to enhance reliability in dynamic, partially observable environments.
ResearchPhysical AI research
Physical AI research
driven by real-world data
Robgence researches multimodal perception, embodied learning, sim-to-real transfer, and vision-language-action (VLA) models — with publications at ICRA, CoRL, RSS, and NeurIPS. Every research initiative is grounded in data collected from real-world operational environments through our global network of 20,000+ trained operators across 50+ cities.
4Research Focus Areas
ICRA · CoRL · RSS · NeurIPSPublication Venues
100%Real-World Training Data
500K+Egocentric Frames Studied
Learning from human demonstrations
Exploring imitation learning, teleoperation, and human demonstration datasets for training embodied AI systems to perform complex tasks in the real world.
Closing the reality gap
Building pipelines which utilize synthetic data, digital twins, and high-fidelity real-world datasets to reduce the reality gap.
Foundation models for embodied intelligence
Investigating large-scale egocentric and multimodal datasets that power VLA models, and next-generation robotics foundation models.
Our VisionWe’re not just building
We’re not just building
data infrastructure. We’re building toward embodied intelligence.
Robgence began by solving one of the most fundamental challenges in Physical AI: access to high-quality real-world data. However, data collection is only the beginning.
Every environment we capture, every interaction we record, and every dataset we deliver contributes to a larger mission. Our research extends beyond data infrastructure into the core challenges that will define the next generation of embodied intelligence — humanoid robotics, world models, manipulation learning, and autonomous decision-making.
By combining large-scale data operations with applied robotics research, we’re helping build the foundations for systems that can understand, adapt to, and operate within the physical world with greater capability and autonomy.
The Frontiers We’re Building Toward
4Research Domains
∞Real-World Scale
Selected Publications
Collaborate with us Recent Work
Robgence research has been presented at ICRA, CoRL, RSS, and NeurIPS — covering tactile-conditioned manipulation, data-efficient imitation learning, and real-world demonstration collection at scale.
/ 01ICRA · 2026
Tactile-conditioned policies for unstructured manipulation
Integrating tactile feedback with visual observation improves grasp success rates on previously unseen objects compared to vision-only baselines.
Read Paper
/ 02CoRL · 2025
Distribution-aware data scaling for Physical AI
Curriculum-based distribution sampling during dataset construction reduces out-of-distribution failure modes and improves policy generalization across environment types.
Read Paper
/ 03RSS · 2025
Operator-in-the-loop imitation learning from active manufacturing facilities
Real-world teleoperation data collected in active production environments outperforms controlled lab demonstrations for contact-rich assembly tasks.
Read Paper
/ 04NeurIPS Workshop · 2024
Synthetic + real curriculum for multi-finger grasping
A staged synthetic-to-real curriculum halves the volume of real-world demonstrations needed to achieve robust multi-finger grasping on novel objects.
Read Paper
FAQ
Research questions
Common questions about Robgence's research focus, publications, and collaboration opportunities.
Get in touchWhat does Robgence research?
Robgence research focuses on four core areas of Physical AI: multimodal perception under partial observability, embodied learning from human demonstrations and teleoperation, sim-to-real transfer for closing the reality gap, and vision-language-action (VLA) models for robotics foundation models. All research is grounded in real-world data collected through our global operator network.
Where has Robgence published research?
Robgence has published at leading robotics and machine learning venues including ICRA (International Conference on Robotics and Automation), CoRL (Conference on Robot Learning), RSS (Robotics: Science and Systems), and NeurIPS workshops.
How does Robgence's real-world data approach differ from lab-based robotics research?
Robgence collects all training data in the environments where robots are actually deployed — manufacturing floors, warehouses, homes, and healthcare facilities — rather than controlled lab settings. This real-world grounding reduces the domain gap between training conditions and deployment conditions, improving policy robustness on contact-rich and unstructured tasks.
Can research teams collaborate with Robgence?
Yes. Robgence actively collaborates with university research groups and enterprise AI teams. Collaboration can take the form of joint data collection, dataset licensing, annotation partnerships, or co-authorship on published work. Contact us via the Collaborate page to discuss a partnership.